>
volge

The AI Rulebooks

AI GOVERNANCE · RULES OF THE GAME
How AI and agents are governed sovereignly by nations, and what each regime makes you do. A read on the United States, India and the six Gulf states, set against the European instrument they are all drafted beside, and on the one gap every single regime shares.
No 01 · 2027

Rules of the Game is a series of books, and a read on the reg­u­lat­ory land­scape: what binds, who it binds, and what it makes an op­er­ator ac­tu­ally do. Each edi­tion takes one sub­ject and sep­ar­ates the in­stru­ment in force from the rule that is pro­posed and the in­ten­tion that has only been an­nounced. The three are not the same thing, and they are routinely re­por­ted as though they were.

How AI and agents are gov­erned sov­er­eignly by na­tions, and what each re­gime makes you do. A read on the United States, India and the six Gulf states, set against the European in­stru­ment they are all draf­ted be­side, and on the one gap every single re­gime shares. A read on the na­tional rule­books for ar­ti­fi­cial in­tel­li­gence across the United States, India and the six Gulf states, and on the gap every one of them shares.

Every rule car­ries its in­stru­ment, its is­su­ing body and its date. A date that is sched­uled rather than in force is said to be sched­uled, and a re­gime an­nounced without an in­stru­ment is re­cor­ded as an an­nounce­ment and kept off the dead­line board. Where a primary source ex­ists it is used, and where only com­ment­ary could be reached the com­ment­ator is named.

Contents

01The approach
  1. 02The three patterns: comprehensive statute, sectoral and state-led, strategy-first
  2. 04The EU AI Act as the template the others draft against
  3. 06China: registration before deployment
  4. 08The soft layer that binds in practice: NIST, ISO/IEC 42001, OECD
  5. 11Where agents fall outside all of it
02The map
  1. 15The United States: no federal statute, and what fills the gap
  2. 17The states as the real regulator
  3. 19India: DPDP as the binding floor, synthetic content as the first duty
  4. 21India: the sector regulators, and the law held in reserve
  5. 24Saudi Arabia: SDAIA, and what actually binds
  6. 26The UAE: where the rule sits below the headline
  7. 28Qatar
  8. 30Kuwait, Oman and Bahrain: strategy without instrument
03The champions
  1. 34When the state owns the model, what changes
  2. 36Humain, and Saudi industrial policy for AI
  3. 38QAI, and Qatar's bet on sovereign compute
  4. 40G42, MGX, and the UAE's two-track position
  5. 42India: the AI Mission, and the sovereign model question
  6. 44Who regulates a national champion?
04The duties
  1. 48Inventory and classification: knowing what you run
  2. 50Documentation and evidence
  3. 52Human oversight, and what it costs
  4. 53Disclosure, labelling and provenance
  5. 55Procurement: governing what you did not build
  6. 58Buying from a sovereign champion: the contract questions
05The operator’s read
  1. 61What binds now, soonest first
  2. 62The agent gap, and what to do before it closes
  3. 64A single build to the highest applicable bar
  4. 66Method and scope
  5. 67Volge's expertise in AI governance
  6. •Sources
AI governanceRules of the Game01 •
01

How AI governance is approached globally

Three patterns, one template, and the gap they all share.
  • The three patterns: comprehensive statute, sectoral and state-led, strategy-first
  • The EU AI Act as the template the others draft against
  • China: registration before deployment
  • The soft layer that binds in practice: NIST, ISO/IEC 42001, OECD
  • Where agents fall outside all of it

The three patterns: comprehensive statute, sectoral and state-led, strategy-first

There is no single AI rulebook, and three patterns account for almost every national position a company now has to comply with.

No single AI rule­book ex­ists, and none is close. Three pat­terns ac­count for al­most every na­tional po­s­i­tion as at 6 Novem­ber 2026. The first is the com­pre­hens­ive bind­ing stat­ute, and the European Union holds the only work­ing ex­ample of it. Reg­u­la­tion (EU) 2024/1689 has been in force since 1 Au­gust 2024 (European Par­lia­ment and Coun­cil, 13 June 2024) and ap­plies in tranches run­ning to 2 Au­gust 2028. The second pat­tern is sec­toral and state-led, with no ho­ri­zontal stat­ute any­where in sight. That is the Amer­ican po­s­i­tion. The third is strategy be­fore law, and it is the com­mon­est of the three.

China sits out­side those three and is bet­ter read as a fourth po­s­i­tion, bind­ing and re­gis­tra­tion-led. Four de­part­mental in­stru­ments of the Cy­ber­space Ad­min­is­tra­tion of China are in force, with an agent-spe­cific im­ple­ment­a­tion opin­ion added on 8 May 2026 (CAC, NDRC and MIIT, 8 May 2026). The United King­dom is reg­u­lator-led by de­lib­er­ate choice, and its only bind­ing AI-spe­cific sec­ond­ary le­gis­la­tion re­quires a reg­u­lator to write a code rather than reg­u­lat­ing any sys­tem (SI 2026/425, in force 12 May 2026). India has pub­lished vol­un­tary guidelines over a floor of ex­ist­ing law. The Gulf states put strategy first. Stat­ute may fol­low.

Strategy is now near-uni­ver­sal, which is why it ex­plains so little. UN Trade and De­vel­op­ment coun­ted 89 na­tional AI policies world­wide by the end of 2023, of which only six came from least de­veloped coun­tries (UNCTAD, Au­gust 2025). About two thirds of de­veloped coun­tries had ad­op­ted one by then. The OECD.AI Policy Nav­ig­ator draws on more than 80 jur­is­dic­tions and re­cords over 900 na­tional AI policies and ini­ti­at­ives, of which 63 sit in the act­ive na­tional strategy cat­egory (OECD.AI, 6 Novem­ber 2026). That fig­ure counts ini­ti­at­ives, not jur­is­dic­tions. Japan alone holds two of the 63, so the true num­ber of coun­tries is lower.

En­acted law is far rarer. Across the G20, 11 AI-re­lated bills were passed into law in 2025, against none at all in 2016 (Stan­ford HAI, 2026 AI Index Re­port). The cu­mu­lat­ive count for 2016 to 2025 puts the United States at 25 and South Korea at 17, with France and Japan at 10 apiece and Italy at nine. The Index cau­tions that the smal­ler val­ues in that series should be read as in­dic­at­ive rather than exact. On its fig­ures China and India show one en­act­ment each over the dec­ade, which is a use­ful cor­rect­ive to the im­pres­sion that volume of law tracks volume of de­ploy­ment.

AI governanceRules of the Game02 •
Cumulative AI-related bills passed into law in G20 countries over the decade to 2025.
bills
United States25South Korea17France10Japan10Italy9United Kingdom6Germany5
Source: Stanford HAI, 2026 AI Index Report, 2026

No au­thor­it­at­ive count of jur­is­dic­tions hold­ing a bind­ing AI-spe­cific stat­ute is pub­lished by any­one. The IAPP Global AI Law and Policy Tracker, chapter 8 of the Stan­ford AI Index and the OECD.AI Policy Nav­ig­ator were each checked, and not one of them car­ries such a num­ber. The IAPP tracker mixes com­pre­hens­ive le­gis­la­tion with use-case le­gis­la­tion and na­tional strategies, and its own land­ing page con­cedes that it does not cover every AI ini­ti­at­ive in every jur­is­dic­tion (IAPP, up­dated 6 Oc­to­ber 2026). Third-party counts there­fore vary with the defin­i­tion of bind­ing. The ab­sence is the find­ing.

Strategy is now near-universal, which is why it explains so little.

Out­side the European Union the stat­ute route has been taken twice, and the pen­alty shows the am­bi­tion. South Korea's Frame­work Act on the De­vel­op­ment of Ar­ti­fi­cial In­tel­li­gence and Es­tab­lish­ment of Trust com­menced on 22 Janu­ary 2026 and caps ad­min­is­trat­ive fines at KRW 30,000,000 per vi­ol­a­tion, at­tached to three pro­ced­ural du­ties only, with no turnover per­cent­age any­where in the Act (Law No. 20676, pro­mul­gated 21 Janu­ary 2025). Japan's AI Pro­mo­tion Act, en­acted in May 2025, car­ries no pen­alty pro­vi­sion at all. The Amer­ican num­bers run the other way on volume and the same way on bite, with 150 state bills en­acted in 2025 and 58 fed­eral AI-re­lated reg­u­la­tions re­cor­ded in the same year, 28 of them is­sued by the Ex­ec­ut­ive Of­fice of the Pres­id­ent.

Key takeaways
  • The European Union holds the only com­pre­hens­ive bind­ing AI stat­ute in op­er­a­tion, and every other large jur­is­dic­tion is draft­ing against it rather than copy­ing it.
  • Strategy doc­u­ments are near-uni­ver­sal and there­fore tell a com­pany al­most noth­ing, while en­acted AI law across the G20 reached only 11 bills in 2025.
AI governanceRules of the Game03 •
  • Nobody pub­lishes an au­thor­it­at­ive count of jur­is­dic­tions with a bind­ing AI stat­ute, so any single num­ber offered for that fig­ure should be treated with sus­pi­cion.

The EU AI Act as the template the others draft against

The European statute is the reference text for every other drafter, and its hardest obligations have just been pushed into 2027 and 2028.

The European Union's AI Act is the text every other drafter now works against, and its high-risk tier has just moved. Reg­u­la­tion (EU) 2026/1744, the Di­gital Om­ni­bus on AI, was signed at Stras­bourg on 8 July 2026, pub­lished in the Of­fi­cial Journal on 24 July 2026 and entered into force on 27 July 2026 (European Par­lia­ment and Coun­cil, 8 July 2026). Two dates moved. The Annex III high-risk ob­lig­a­tions went from 2 Au­gust 2026 to 2 Decem­ber 2027, a delay of 16 months, and the Annex I ob­lig­a­tions from 2 Au­gust 2027 to 2 Au­gust 2028, a delay of 12 months. Re­cital (40) gives the reason as the delayed avail­ab­il­ity of stand­ards and the delayed es­tab­lish­ment of na­tional com­pet­ent au­thor­it­ies.

Plenty of the Act binds today. Chapters I and II, in­clud­ing the Art­icle 5 pro­hib­i­tions as ori­gin­ally ad­op­ted, have ap­plied since 2 Feb­ru­ary 2025. The chapters on gen­eral-pur­pose AI mod­els, on gov­ernance and on pen­al­ties, along with Art­icle 78, have ap­plied since 2 Au­gust 2025, with Art­icle 101 carved out of that tranche. The gen­eral date of ap­plic­a­tion was 2 Au­gust 2026, which brought in the Art­icle 50 trans­par­ency du­ties and Art­icle 101 it­self (Pub­lic­a­tions Of­fice, con­sol­id­ated text of 27 July 2026). Art­icles 102 to 110 have ap­plied since 27 July 2026. What is not yet in force is the sub­stance of the high-risk re­gime, and that is the part most com­ment­ary treats as the Act.

Two ob­lig­a­tions fall due on 2 Decem­ber 2026, and the om­ni­bus cre­ated both. The first is a leg­acy mark­ing duty under a new Art­icle 111(4), which re­quires pro­viders of sys­tems gen­er­at­ing syn­thetic audio, image, video or text con­tent placed on the mar­ket be­fore 2 Au­gust 2026 to com­ply with Art­icle 50(2) by that date. Re­cital (38) de­scribes the in­ter­val as a trans­itional period of four months. The second is a pair of new Art­icle 5 pro­hib­i­tions on non-con­sen­sual in­tim­ate im­agery and on child sexual abuse ma­ter­ial. A new Art­icle 5(1a) nar­rows them, so that they bite on a pro­vider only where such gen­er­a­tion is the in­ten­ded pur­pose or a reas­on­ably fore­see­able and re­pro­du­cible out­come. Neither ex­is­ted in the 2024 text.

AI governanceRules of the Game04 •

Ex­actly one har­mon­ised stand­ard sup­ports the Act. EN 18286:2026, on qual­ity man­age­ment sys­tems for AI Act reg­u­lat­ory pur­poses, was cited in the Of­fi­cial Journal by Com­mis­sion Im­ple­ment­ing De­cision (EU) 2026/2274 of 12 Oc­to­ber 2026, pub­lished two days later (European Com­mis­sion, 14 Oc­to­ber 2026). The Annex to that de­cision lists a single entry. CEN-CENELEC calls the stand­ard the first in sup­port of the Act and de­scribes it as one of the first build­ing blocks in a wider stand­ard­isa­tion land­scape still to come (CEN-CENELEC, 15 Oc­to­ber 2026). The cita­tion car­ries real weight, be­cause under Art­icle 40 con­form­ity with a cited stand­ard raises a pre­sump­tion of con­form­ity with the re­quire­ments that stand­ard cov­ers. One stand­ard, against a 16-month de­fer­ral jus­ti­fied by the ab­sence of stand­ards.

One harmonised standard stands behind a 16-month deferral that was justified by the absence of standards.

The guidelines that would tell a firm whether it is high-risk at all are still in draft. Art­icle 6(5) re­quired the Com­mis­sion to pub­lish them, with a com­pre­hens­ive list of prac­tical ex­amples of use cases in­side and out­side the high-risk cat­egory, no later than 2 Feb­ru­ary 2026. Draft guidelines ap­peared on 19 May 2026 in three down­load­able parts. The Com­mis­sion's policy page, last up­dated 19 Oc­to­ber 2026, still re­cords that con­sulta­tion feed­back is going to be in­cor­por­ated in the final ver­sion be­fore ad­op­tion (European Com­mis­sion, 19 Oc­to­ber 2026). Art­icle 6(5) was ex­pressly carved out of the de­fer­ral of Chapter III. The dead­line was missed.

The AI Act's absolute fine ceilings by tier, before the turnover alternative is applied.
EUR
010,000,00020,000,00030,000,00040,000,00035,000,000Article 5 bans15,000,000Operator duties15,000,000GPAI, Art 1017,500,000Misleading info
Source: European Parliament and Council, 2024

The pen­al­ties ex­plain why the rest of the world reads the Act so closely. Breach of the Art­icle 5 pro­hib­i­tions at­tracts up to EUR 35,000,000 or 7 per cent of total world­wide an­nual turnover, whichever is higher. Op­er­ator du­ties and the Art­icle 50 trans­par­ency ob­lig­a­tions sit at EUR 15,000,000 or 3 per cent, and sup­ply­ing in­cor­rect or mis­lead­ing in­form­a­tion to an au­thor­ity at

AI governanceRules of the Game05 •

EUR 7,500,000 or 1 per cent. Pro­viders of gen­eral-pur­pose AI mod­els face EUR 15,000,000 or 3 per cent under Art­icle 101, which the om­ni­bus left un­touched. Re­lief ex­ists for smal­ler firms. Small and me­dium-sized en­ter­prises pay the lower of the two fig­ures rather than the higher, and the om­ni­bus ex­ten­ded that treat­ment to small mid-caps.

Key takeaways
  • The Di­gital Om­ni­bus moved the Annex III high-risk ob­lig­a­tions by 16 months and the Annex I ob­lig­a­tions by 12, cit­ing the ab­sence of stand­ards.
  • Trans­par­ency, gov­ernance, gen­eral-pur­pose model du­ties and the pen­alty re­gime already apply, so a firm selling into Europe is reg­u­lated today whatever hap­pens to the high-risk tier.
  • Ex­actly one har­mon­ised stand­ard has been cited in the Of­fi­cial Journal, and the guidelines that clas­sify a sys­tem as high-risk re­main in draft.

China: registration before deployment

China regulates AI at the point of entry through filing and labelling, and the money penalty standing behind those duties rose tenfold in 2026.

China reg­u­lates AI be­fore de­ploy­ment rather than after harm. The mech­an­ism is the fil­ing. It pred­ates gen­er­at­ive AI. The In­ter­net In­form­a­tion Ser­vice Al­gorithmic Re­com­mend­a­tion Man­age­ment Pro­vi­sions, Order No. 9 of four is­su­ing bod­ies, have been in force since 1 March 2022 (CAC, 31 Decem­ber 2021). A pro­vider with pub­lic opin­ion prop­er­ties or so­cial mo­bil­isa­tion ca­pa­city must file within ten work­ing days of start­ing the ser­vice. What is filed cov­ers the pro­vider's name and ser­vice form, the al­gorithm type and ap­plic­a­tion field, the self-as­sess­ment re­port and the con­tent pro­posed for pub­lic­a­tion. The re­ceiv­ing de­part­ment then has thirty work­ing days to com­plete the fil­ing and issue a num­ber, or to re­fuse with reas­ons.

The gen­er­at­ive AI rules sit on top of that ma­chinery rather than re­pla­cing it. The In­terim Meas­ures for the Ad­min­is­tra­tion of Gen­er­at­ive AI Ser­vices, Order No. 15 of seven is­su­ing bod­ies, came into force on 15 Au­gust 2023 (CAC, 10 July 2023). Art­icle 17 is the com­pli­ance trig­ger. A pro­vider of gen­er­at­ive AI ser­vices with pub­lic opin­ion prop­er­ties or so­cial mo­bil­isa­tion ca­pa­city must carry out a se­cur­ity as­sess­ment under the rel­ev­ant state pro­vi­sions and must com­plete al­gorithm fil­ing under the 2022 Pro­vi­sions. Art­icle 2 fixes scope by where the ser­vice is offered, ap­ply­ing the Meas­ures to the pro­vi­sion of such ser­vices to the pub­lic within Chinese ter­rit­ory. There is no in­cor­por­a­tion test and no ex­emp­tion for an off­shore pro­vider.

AI governanceRules of the Game06 •

La­belling has been a legal duty since 1 Septem­ber 2025. The Meas­ures for La­belling AI-Gen­er­ated Syn­thetic Con­tent were cir­cu­lated by four bod­ies under a no­tice dated 7 March 2025 and pub­lished on 14 March 2025 (CAC, 14 March 2025). They sep­ar­ate ex­pli­cit la­belling, which the user can per­ceive, from im­pli­cit la­belling added by tech­nical means in the con­tent file data. Art­icle 5 re­quires the im­pli­cit label in file metadata to carry the con­tent at­trib­ute in­form­a­tion, the pro­vider's name or code and the con­tent num­ber. The sup­port­ing man­dat­ory na­tional stand­ard GB 45438-2025 took ef­fect on the same day (Stand­ard­isa­tion Ad­min­is­tra­tion of China and SAMR, 28 Feb­ru­ary 2025). Plat­forms carry du­ties too. Art­icle 6 re­quires a dis­sem­in­a­tion plat­form to check the metadata and to add a graded prompt around pub­lished con­tent.

A for­eign en­ter­prise serving the Chinese pub­lic needs four things, and only three of them are writ­ten in the AI rules. It needs a se­cur­ity as­sess­ment re­port lodged with the city-level or higher cy­ber­space de­part­ment and pub­lic se­cur­ity organ of its loc­al­ity, under the 2018 Se­cur­ity As­sess­ment Pro­vi­sions. It needs an al­gorithm fil­ing through the na­tional plat­form at beian.cac.gov.cn, dir­ec­ted for gen­er­at­ive AI to the pro­vin­cial cy­ber­space de­part­ment, with the fil­ing num­ber then dis­played prom­in­ently on the ser­vice. It needs whatever ad­min­is­trat­ive li­cence the law re­quires, in­side the for­eign-in­vest­ment lim­its set by the neg­at­ive list (NDRC and MOFCOM Order No. 23, in force 1 Novem­ber 2024). The fourth is prac­tice. A do­mestic en­tity hold­ing a PRC busi­ness li­cence and an ICP re­cord is re­por­ted con­sist­ently across prac­ti­tioner guid­ance. No CAC doc­u­ment stat­ing that re­quire­ment could be veri­fied.

Non-compliance by an offshore provider is handled by blocking rather than by a fine.

Non-com­pli­ance by an off­shore pro­vider is handled by block­ing rather than by en­force­ment against the en­tity. Art­icle 20 of the In­terim Meas­ures dir­ects the na­tional cy­ber­space au­thor­ity to no­tify the rel­ev­ant bod­ies to take tech­nical meas­ures and other ne­ces­sary meas­ures. No pub­lished en­force­ment ac­tion under the Meas­ures against a named for­eign pro­vider could be found. For­eign-brand fil­ings do hap­pen. Han Kun Law Of­fices re­ports as­sist­ing HP China in com­plet­ing the deep syn­thesis ser­vice al­gorithm fil­ing for the as­sist­ant built into its AI PC products. It de­scribes its cli­ent as among the first for­eign brands to com­plete such a fil­ing (Han Kun Law Of­fices, 7 Au­gust 2024).

AI governanceRules of the Game07 •
The statutory ceilings the Chinese AI rules refer out to, before and after the 2026 Cybersecurity Law amendment.
RMB
012,500,00025,000,00037,500,00050,000,0001,000,000Cyber Law to 202510,000,000Cyber Law, 202610,000,000Data Security Law50,000,000PIPL
Source: NPC Standing Committee, 2025

The money moved in 2026. The AI in­stru­ments mostly carry no fine of their own, re­fer­ring out in­stead to the Cy­ber­se­cur­ity Law and the Per­sonal In­form­a­tion Pro­tec­tion Law among oth­ers. The amend­ing de­cision on the Cy­ber­se­cur­ity Law was ad­op­ted on 28 Oc­to­ber 2025 and took ef­fect on 1 Janu­ary 2026, rais­ing the top tier from RMB 1,000,000 to RMB 10,000,000, with up to RMB 1,000,000 on re­spons­ible of­ficers (NPC Stand­ing Com­mit­tee, 28 Oc­to­ber 2025). The Per­sonal In­form­a­tion Pro­tec­tion Law re­mains heav­ier still at RMB 50,000,000 or 5 per cent of pre­vi­ous-year turnover. Only two of the AI in­stru­ments set their own fines. The newer of them, the An­thro­po­morphic AI Meas­ures in force since 15 July 2026, reaches RMB 200,000 where life and health safety is en­dangered with harm­ful con­sequences.

Key takeaways
  • China's con­trol point is the fil­ing, lodged within ten work­ing days of launch, with the fil­ing num­ber then dis­played prom­in­ently on the ser­vice it­self.
  • La­belling has been a bind­ing duty since 1 Septem­ber 2025, cov­er­ing both a label the user can per­ceive and an im­pli­cit one car­ried in file metadata.
  • The AI rules mostly carry no fine of their own, and the Cy­ber­se­cur­ity Law ceil­ing they refer out to rose ten­fold on 1 Janu­ary 2026.

The soft layer that binds in practice: NIST, ISO/IEC 42001, OECD

Voluntary frameworks matter only where a contract or a statute points at them, and the clearest American procurement pointer was deleted in 2025.

The soft layer binds only where some­thing harder points at it, and in 2025 the clearest pointer was cut rather than tightened. OMB Memor­andum M-24-18 of 24 Septem­ber

AI governanceRules of the Game08 •

2024 dir­ec­ted United States fed­eral AI ac­quis­i­tion to the NIST AI Risk Man­age­ment Frame­work by name (OMB, 24 Septem­ber 2024). It cited the frame­work's ap­pendix on how AI risk dif­fers from tra­di­tional soft­ware risk, and drew its gen­er­at­ive AI risk list from the frame­work's gen­er­at­ive pro­file. OMB Memor­andum M-25-22 of 3 April 2025 res­cin­ded and re­placed that memor­andum. NIST is not men­tioned once. Con­tracts must in­stead en­sure com­pli­ance with the min­imum risk man­age­ment prac­tices re­quired under OMB's own M-25-21 (OMB, 3 April 2025).

The frame­work that the memor­andum stopped cit­ing is it­self in sus­pen­sion. NIST re­leased AI RMF 1.0 on 26 Janu­ary 2023 as NIST AI 100-1, built around four func­tions and in­ten­ded through­out for vol­un­tary use (NIST, 26 Janu­ary 2023). Its de­vel­op­ment drew about 400 sets of formal com­ments from more than 240 or­gan­isa­tions, fol­low­ing a dir­ec­tion from Con­gress. Amer­ica's AI Ac­tion Plan of July 2025 then in­struc­ted NIST to re­vise the frame­work, spe­cific­ally to re­move ref­er­ences to mis­in­form­a­tion, to di­versity, equity and in­clu­sion, and to cli­mate change. No draft ex­ists. As at 6 Novem­ber 2026 there is no ver­sion 2.0, no pub­lished tar­get date and no ver­sion num­ber, and the NIST pages state twice that the frame­work is being re­vised.

ISO/IEC 42001:2023 is the one soft in­stru­ment that can be audited. Pub­lished in Decem­ber 2023 at 51 pages by ISO/IEC JTC 1/SC 42, it sets re­quire­ments for an AI man­age­ment sys­tem rather than guid­ance about one (ISO, Decem­ber 2023). What makes it con­tract­able is ISO/IEC 42006:2025, pub­lished in July 2025, which sets re­quire­ments for the bod­ies that audit and cer­tify against it. Up­take can­not be stated. The ISO Sur­vey is no longer pub­lished by ISO and now sits be­hind a login on IAF CertSearch, and the In­ter­na­tional Ac­cred­it­a­tion Forum it­self ceased op­er­a­tions on 1 Janu­ary 2026, so no au­thor­it­at­ive count of cer­ti­fic­ates could be veri­fied. Sec­ond­ary es­tim­ates cir­cu­late and are not prin­ted here.

The OECD Re­com­mend­a­tion on Ar­ti­fi­cial In­tel­li­gence is the widest-signed AI in­stru­ment and the least bind­ing. OECD/LEGAL/0449 was ad­op­ted by the Coun­cil at min­is­terial level on 22 May 2019 and re­vised twice, on 8 Novem­ber 2023 and again on 3 May 2024 (OECD, 3 May 2024). The second re­vi­sion added the hand­ling of mis­in­form­a­tion and in­form­a­tion in­teg­rity in gen­er­at­ive AI, to­gether with safe over­ride, re­pair or de­com­mis­sion­ing of sys­tems. It has 47 ad­her­ents. A Coun­cil Re­com­mend­a­tion is not leg­ally bind­ing, and the OECD de­scribes it as a polit­ical com­mit­ment car­ry­ing an ex­pect­a­tion that ad­her­ents will do their best to im­ple­ment it. Its defin­i­tion of an AI sys­tem is non­ethe­less the one the European Union, the Coun­cil of Europe and the United Na­tions have ad­op­ted.

AI governanceRules of the Game09 •
Signature against ratification for the Council of Europe AI Convention, and the threshold it must clear to apply.
states and organisations
06.212.518.82521Signatories1Ratifications5Needed to apply
Source: Council of Europe, 2026

The treaty writ­ten to be bind­ing is the one that binds nobody. The Coun­cil of Europe Frame­work Con­ven­tion on Ar­ti­fi­cial In­tel­li­gence was ad­op­ted on 17 May 2024 and opened for sig­na­ture at Vil­nius on 5 Septem­ber 2024, with ten sig­nat­or­ies on the day (Coun­cil of Europe, 5 Septem­ber 2024). It now has 21 sig­nat­or­ies. It has one Party, the European Union, which rat­i­fied on 15 May 2026. Art­icle 30 sets entry into force three months after five sig­nat­or­ies have rat­i­fied, at least three of them Coun­cil of Europe mem­ber states. The Con­ven­tion is there­fore not in force, and the United King­dom, which signed on the open­ing day, has not rat­i­fied.

The treaty written to be binding is the one that binds nobody.

Three mech­an­isms give the soft layer real force in 2026, and not one of them is a frame­work. The first is the AI Act's Art­icle 40 pre­sump­tion of con­form­ity, which turns a stand­ard cited in the Of­fi­cial Journal into legal cover. The second is ISO/IEC 42006:2025, which makes cer­ti­fic­a­tion against ISO/IEC 42001 sale­able and there­fore spe­cifiable in a con­tract. The third is a con­tract clause. Pro­posed GSAR clause 552.239-7001 would define its sup­ply-chain roles by ref­er­ence to NIST AI RMF 1.0 and would re­quire sys­tem doc­u­ment­a­tion con­sist­ent with the frame­work's guidelines (Gen­eral Ser­vices Ad­min­is­tra­tion, 17 June 2026). Its final text and ef­fect­ive date could not be veri­fied. Brit­ish pro­cure­ment guid­ance points at none of the three, cit­ing only do­mestic in­stru­ments.

Key takeaways
  • Amer­ican fed­eral pro­cure­ment stopped point­ing at the NIST frame­work in April 2025, and a Gen­eral Ser­vices Ad­min­is­tra­tion clause has since pro­posed to point at it again.
  • The OECD Re­com­mend­a­tion has 47 ad­her­ents and binds none of them, while the Coun­cil of Europe treaty holds one rat­i­fic­a­tion against a threshold of five.
AI governanceRules of the Game10 •
  • ISO/IEC 42001 is the one soft in­stru­ment that can be audited, but no au­thor­it­at­ive count of cer­ti­fic­ates is­sued against it could be veri­fied.

Where agents fall outside all of it

Autonomous agents are almost absent from binding AI law, and what does reach them was written before agentic AI existed.

Autonom­ous agents sit out­side nearly all of the AI rule­books, and the counts are the proof. The word agent, in any form, ap­pears zero times in Reg­u­la­tion (EU) 2024/1689. For cal­ib­ra­tion, the same 591,474-char­ac­ter text car­ries 484 in­stances of high-risk. The pat­tern re­peats. The pub­lished ETSI baseline for AI cyber se­cur­ity con­tains no in­stance of the word either, in the tech­nical spe­cific­a­tion or the European Stand­ard that su­per­seded it. Not one of the 97 titles in the ISO/IEC JTC 1/SC 42 work pro­gramme con­tains agent or agen­tic. The Texas Re­spons­ible Ar­ti­fi­cial In­tel­li­gence Gov­ernance Act re­turns zero for agent and zero for autonom­ous across its 43,308-char­ac­ter en­rolled text.

What the European stat­ute uses in­stead is autonomy, and it uses it el­even times. Art­icle 3(1) defines an AI sys­tem as a ma­chine-based sys­tem de­signed to op­er­ate with vary­ing levels of autonomy. The human over­sight pro­vi­sions re­quire meas­ures com­men­sur­ate with the level of autonomy. The pen­alty cri­teria take ac­count of the ex­tent to which a sys­tem acts autonom­ously and whether a human can over­ride a de­cision. The Di­gital Om­ni­bus then pro­duced the first ap­pear­ance of the phrase Agen­tic AI in bind­ing European law, ex­actly once, as code AIH 0401 in a new Annex XIV list­ing the des­ig­na­tion codes for no­ti­fied bod­ies (European Par­lia­ment and Coun­cil, 8 July 2026). It im­poses noth­ing.

Agent-relevant instruments counted by status as at 6 November 2026, from published voluntary texts to binding law.
instruments
02.557.51010Voluntary8Announced/draft3Binding1Model law
Source: OWASP, IMDA, ETSI and CSA, 2026
AI governanceRules of the Game11 •

One in­stru­ment any­where in the world reg­u­lates agent autonomy as such, and it is Chinese. The Im­ple­ment­a­tion Opin­ions on the Stand­ard­ised Ap­plic­a­tion and In­nov­at­ive De­vel­op­ment of Agents were pub­lished on 8 May 2026 (CAC, NDRC and MIIT, 8 May 2026). Item 6 is the op­er­at­ive rule. It re­quires three de­cision modes to be de­lin­eated, cov­er­ing de­cisions re­served to the user per­son­ally, de­cisions re­quir­ing the user's au­thor­isa­tion and de­cisions the agent takes on its own. It then provides that the user must re­tain the right to know and the final de­cision, and that an agent's ex­e­cu­tion of op­er­a­tions must not ex­ceed the scope of the user's au­thor­isa­tion. The caveat mat­ters. This is an ad­min­is­trat­ive im­ple­ment­a­tion opin­ion giv­ing ef­fect to the State Coun­cil's AI Plus policy. It binds state or­gans and steers in­dustry. It car­ries no pen­al­ties of its own.

The clearest binding rule reaching an autonomous agent is twenty-six years old, and it was written for electronic data interchange.

The clearest bind­ing rule reach­ing an autonom­ous agent is twenty-six years old. Sec­tion 14 of the Uni­form Elec­tronic Trans­ac­tions Act, veri­fied here from the Michigan en­act­ment of Act 305 of 2000, im­me­di­ately ef­fect­ive on 16 Oc­to­ber 2000, provides that a con­tract may be formed by the in­ter­ac­tion of elec­tronic agents of the parties. That holds even if no in­di­vidual was aware of or re­viewed the agents' ac­tions or the res­ult­ing terms (Michigan Le­gis­lature, 16 Oc­to­ber 2000). An elec­tronic agent is defined as a com­puter pro­gram or auto­mated means act­ing without re­view by an in­di­vidual. So a con­tract an agent forms in the United States without human ap­proval already binds the prin­cipal, who can­not es­cape it on the ground that nobody saw it. It was writ­ten for elec­tronic data in­ter­change.

Two fur­ther hooks exist. Neither was draf­ted for AI. Com­mis­sion Del­eg­ated Reg­u­la­tion (EU) 2017/589 has been in force since 3 Janu­ary 2018 and gov­erns sys­tems that trans­act without per-order human ap­proval, im­pos­ing pre-trade con­trols and a man­dat­ory abil­ity to can­cel all un­ex­ecuted or­ders (European Com­mis­sion, 19 July 2016). It is framed around al­gorithmic trad­ing by in­vest­ment firms. The UNCITRAL Model Law on Auto­mated Con­tract­ing, ad­op­ted on 11 July 2024, is the only in­ter­na­tional text writ­ten de­lib­er­ately for non-de­term­in­istic autonom­ous sys­tems in con­tract. Its Art­icle 5(1) denies any ob­jec­tion to a con­tract on the sole ground that no nat­ural per­son in­ter­vened in its form­a­tion (UNCITRAL, 11 July 2024). It is a model law, so it binds nowhere until a state en­acts it.

AI governanceRules of the Game12 •

Everything else on agents is vol­un­tary or un­fin­ished. Ten agent-rel­ev­ant in­stru­ments have been pub­lished and every one of them is vol­un­tary. They in­clude the OWASP Top 10 for Agen­tic Ap­plic­a­tions of Decem­ber 2025 and the OWASP Agent Con­trol Stand­ard v0.1.0 of 1 Septem­ber 2026. Singa­pore's IMDA frame­work for agen­tic AI be­longs there too, in a ver­sion 1.5 of 20 May 2026 that su­per­sedes the launch edi­tion (IMDA, 20 May 2026). Eight more in­stru­ments are an­nounced or in draft, in­clud­ing all three ETSI agent work items. Only three bind. Of those, only the Chinese opin­ion was writ­ten for AI agents at all. The sharpest reg­u­lat­ory state­ment of 2026 came from a com­pet­i­tion au­thor­ity rather than a le­gis­lature, with the UK Com­pet­i­tion and Mar­kets Au­thor­ity telling busi­nesses that if an AI agent you use does some­thing il­legal, you are re­spons­ible (CMA, 9 March 2026).

Key takeaways
  • The European AI Act con­tains the word agent zero times, against 484 in­stances of high-risk in the same text, and the om­ni­bus adds only a tax­onomy code.
  • China's Im­ple­ment­a­tion Opin­ions on Agents of 8 May 2026 is the only in­stru­ment any­where that caps agent autonomy, and it car­ries no pen­al­ties of its own.
  • What binds an autonom­ous agent today is con­tract law writ­ten for elec­tronic data in­ter­change in 2000 and trad­ing rules writ­ten for in­vest­ment firms.
AI governanceRules of the Game13 •

The part in brief

  1. 01Only the European Union runs a comprehensive binding AI statute, with Regulation (EU) 2024/1689 effective from 1 August 2024 and phasing in until 2 August 2028.
  2. 02Strategy documents now say little, since the OECD navigator logs over 900 national AI policies while only 11 AI bills passed into law across the G20 during 2025.
  3. 03No authoritative count exists of jurisdictions holding a binding AI statute, since the IAPP tracker, the Stanford AI Index and the OECD navigator each omit such a figure.
  4. 04China controls deployment through filing, requiring a lodgement within ten working days of launch and a displayed filing number, with content labelling binding since 1 September 2025.
  5. 05Agents sit outside almost every rulebook: the European statute never uses the word agent, and only a Chinese implementation opinion of 8 May 2026 caps agent autonomy, without penalties.
AI governanceRules of the Game14 •
How AI governance is approached globally

The clearest binding rule reaching an autonomous agent is twenty-six years old, and it was written for electronic data interchange.

02

The map

The United States, India and the six Gulf states, instrument by instrument.
  • The United States: no federal statute, and what fills the gap
  • The states as the real regulator
  • India: DPDP as the binding floor, synthetic content as the first duty
  • India: the sector regulators, and the law held in reserve
  • Saudi Arabia: SDAIA, and what actually binds
  • The UAE: where the rule sits below the headline
  • Qatar
  • Kuwait, Oman and Bahrain: strategy without instrument

The United States: no federal statute, and what fills the gap

One free-standing federal AI statute exists, and its subject is intimate imagery. Executive orders and budget guidance carry the rest.

The United States has one free-stand­ing fed­eral stat­ute on ar­ti­fi­cial in­tel­li­gence, and its sub­ject is in­tim­ate im­agery. Everything else is ex­ec­ut­ive ac­tion. The TAKE IT DOWN Act, S. 146 of the 119th Con­gress, be­came Pub­lic Law 119-12 on 19 May 2025. It crim­in­al­ises the know­ing pub­lic­a­tion of non-con­sen­sual in­tim­ate visual de­pic­tions and of di­gital for­ger­ies, mean­ing de­pic­tions cre­ated through soft­ware or ma­chine learn­ing that are in­dis­tin­guish­able from au­then­tic ones. From 19 May 2026 a covered plat­form must hon­our a valid re­moval re­quest within 48 hours, and fail­ure is an un­fair or de­cept­ive prac­tice en­force­able by the Fed­eral Trade Com­mis­sion.

What was res­cin­ded mat­ters, be­cause the re­cord is often re­por­ted wrongly. Ex­ec­ut­ive Order 14110 of 30 Oc­to­ber 2023 was re­voked by Ex­ec­ut­ive Order 14148 of 20 Janu­ary 2025, at sec­tion 2(ggg), pub­lished at 90 FR 8237. EO 14179 did not re­voke it. Its own sec­tion 5(a) refers to ac­tions taken under the already re­voked order, so com­ment­ary cred­it­ing EO 14179 with the re­voc­a­tion should be set aside. EO 14179 of 23 Janu­ary 2025 re­mains in force. It de­clares a policy of sus­tain­ing Amer­ican AI dom­in­ance and re­quired OMB to re­vise its two AI memor­anda within 60 days.

The or­ders that fill the gap are about build­ing, ex­port­ing and pro­cur­ing rather than li­cens­ing. Three were signed on 23 July 2025, cov­er­ing fed­eral per­mit­ting of data centre in­fra­struc­ture (EO 14318), pro­cure­ment of large lan­guage mod­els under two Un­biased AI Prin­ciples (EO 14319) and ex­port of the Amer­ican AI tech­no­logy stack (EO 14320). The same day brought the AI Ac­tion Plan, a policy doc­u­ment. EO 14409 of 2 June 2026 went fur­ther into se­cur­ity, re­quir­ing a clas­si­fied bench­mark­ing pro­cess for covered fron­tier mod­els. Then it closed the door. Sec­tion 3(c) states that noth­ing in the sec­tion au­thor­ises a man­dat­ory li­cens­ing, pre­clear­ance or per­mit­ting re­quire­ment for new AI mod­els.

Nothing in the section authorises a mandatory licensing, preclearance or permitting requirement for new AI models.
AI governanceRules of the Game15 •

One order of 2026 changed the vocab­u­lary rather than the law. Ex­ec­ut­ive Order 14434, "In­aug­ur­at­ing the Era of Super In­tel­li­gence", was signed on 29 Septem­ber 2026 and pub­lished on 2 Oc­to­ber 2026 at 91 FR 63129. The ex­ec­ut­ive branch is to use "Super In­tel­li­gence" and "SI" in place of "Ar­ti­fi­cial In­tel­li­gence" and "AI" in non-stat­utory doc­u­ments, while pre­vi­ously is­sued reg­u­la­tions and con­tracts need not be altered. Agen­cies have com­plied. OCC Bul­letin 2026-54 car­ries a foot­note de­fin­ing the new term by ref­er­ence to 15 USC 9401(3) and cit­ing the order.

What binds fed­eral agen­cies is budget-of­fice guid­ance. OMB Memor­andum M-25-21 of 3 April 2025 re­placed M-24-10 and reaches all ex­ec­ut­ive branch de­part­ments and agen­cies, in­clud­ing the in­de­pend­ent reg­u­lat­ory agen­cies. It defines high-im­pact AI as AI whose out­put is a prin­cipal basis for de­cisions with sig­ni­fic­ant ef­fect on mat­ters such as civil rights, credit or em­ploy­ment, and lists fif­teen cat­egor­ies pre­sumed to qual­ify. Seven min­imum prac­tices fol­low for those uses, among them pre-de­ploy­ment test­ing, an im­pact as­sess­ment with in­de­pend­ent re­view and a signed risk ac­cept­ance, and human over­sight with a fail-safe. M-25-22 of the same date gov­erns ac­quis­i­tion, and re­quires con­tracts per­man­ently to pro­hibit the use of non-pub­lic agency data and res­ults to train com­mer­cially avail­able AI al­gorithms ab­sent agency con­sent.

Pre-emption of state law has been at­temp­ted and has not suc­ceeded. The Sen­ate voted 99 to 1 in the early hours of 1 July 2025 to strike a pro­posed ten-year morator­ium from the budget re­con­cili­ation bill. Ex­ec­ut­ive Order 14365 of 11 Decem­ber 2025 took the ex­ec­ut­ive route, re­quir­ing an AI Lit­ig­a­tion Task Force, a Com­merce eval­u­ation of oner­ous state AI laws by 11 March 2026 and a broad­band policy no­tice on the same date. Neither de­liv­er­able was pub­lished. The Fed­eral Trade Com­mis­sion did de­liver, is­su­ing a pro­posed policy state­ment on the sup­pres­sion of ac­cur­acy in AI sys­tems on 1 July 2026 under docket FTC-2026-0859. No final state­ment was loc­ated, and no stat­ute pre-empt­ing state AI law has been en­acted.

Key takeaways
  • The only free-stand­ing fed­eral AI stat­ute is the TAKE IT DOWN Act of 19 May 2025, which gave plat­forms a 48-hour re­moval duty from 19 May 2026.
  • EO 14409 of 2 June 2026 ex­pressly de­clines to au­thor­ise li­cens­ing, pre­clear­ance or per­mit­ting for new mod­els, so there is no fed­eral ap­proval re­gime.
  • OMB M-25-21 and M-25-22, both of 3 April 2025, bind agen­cies and their sup­pli­ers. Neither reaches a com­pany that sells to nobody in gov­ern­ment.
AI governanceRules of the Game16 •

The states as the real regulator

Five states carry the operative American rulebook, and the headline Colorado statute never came into force at all.

Col­or­ado's AI Act, the law most often cited as Amer­ica's first, never took ef­fect. The se­quence is worth fol­low­ing. SB 24-205 was due to com­mence on 1 Feb­ru­ary 2026, and SB 25B-004 of 28 Au­gust 2025 pushed that to 30 June 2026. In April 2026 the At­tor­ney Gen­eral un­der­took, in a mo­tion filed jointly with xAI and the De­part­ment of Justice, not to en­force it pending the con­clu­sion of rule­mak­ing. It was re­pealed on 14 May 2026, six weeks be­fore the de­ferred date ar­rived. The re­place­ment, SB 26-189 on auto­mated de­cision-mak­ing tech­no­logy, was en­acted the same day and com­mences on 1 Janu­ary 2027.

The one state law that stopped operating was stopped by its own legislature, not by Washington.

Cali­for­nia le­gis­lates faster than any other state, and in two dir­ec­tions at once. The first threshold is com­pute. SB 53, the Trans­par­ency in Fron­tier Ar­ti­fi­cial In­tel­li­gence Act, was signed on 29 Septem­ber 2025 and took ef­fect on 1 Janu­ary 2026, bind­ing de­velopers of found­a­tion mod­els trained with more than 10^26 op­er­a­tions and im­pos­ing the heav­ier du­ties on groups with an­nual gross rev­en­ues above USD 500 mil­lion. A crit­ical safety in­cid­ent must be re­por­ted to the Of­fice of Emer­gency Ser­vices within 15 days of dis­cov­ery, and within 24 hours to law en­force­ment where there is im­min­ent risk of death. Civil pen­al­ties reach USD 1,000,000 per vi­ol­a­tion.

The proven­ance re­gime moved more slowly. SB 942, the Cali­for­nia AI Trans­par­ency Act, was signed on 19 Septem­ber 2024, and AB 853 of 13 Oc­to­ber 2025 moved its op­er­at­ive date from 1 Janu­ary 2026 to 2 Au­gust 2026. Large on­line plat­form du­ties fol­low on 1 Janu­ary 2027, and cap­ture-device du­ties from 1 Janu­ary 2028, at USD 5,000 per vi­ol­a­tion. The pri­vacy agency's auto­mated de­cision-mak­ing reg­u­la­tions took ef­fect on 1 Janu­ary 2026, with com­pli­ance for sig­ni­fic­ant de­cisions from 1 Janu­ary 2027.

Texas took the op­pos­ite ap­proach to proof. HB 149, the Texas Re­spons­ible Ar­ti­fi­cial In­tel­li­gence Gov­ernance Act, states in sec­tion 10 that it takes ef­fect on 1 Janu­ary 2026. Sec­tion

AI governanceRules of the Game17 •

552.056(b) reaches a de­veloper or de­ployer act­ing with the in­tent to dis­crim­in­ate un­law­fully against a pro­tec­ted class, and sec­tion 552.056(c) adds that a dis­par­ate im­pact is not suf­fi­cient by it­self to demon­strate in­tent. En­force­ment is ex­clus­ive to the At­tor­ney Gen­eral and there is no private right of ac­tion. Pen­al­ties run from USD 10,000 to 12,000 for cur­able vi­ol­a­tions left un­cured and from USD 80,000 to 200,000 for un­cur­able ones.

Utah is the one state whose cross-sec­tor AI stat­ute car­ries an ex­piry date. The sun­set is the point. SB 149, the Ar­ti­fi­cial In­tel­li­gence Policy Act, took ef­fect on 1 May 2024 and is co­di­fied at Utah Code Title 13 Chapter 72. Its gen­eral duty is dis­clos­ure on re­quest rather than dis­clos­ure by de­fault, and a heav­ier duty to dis­close prom­in­ently ap­plies to reg­u­lated oc­cu­pa­tions. SB 226 nar­rowed it from 7 May 2025 to a high-risk ar­ti­fi­cial in­tel­li­gence in­ter­ac­tion. SB 332 ex­ten­ded the Act's sun­set from 7 May 2025 to 1 July 2027.

New York's hir­ing rule is mu­ni­cipal, and the state's own fron­tier law is not yet live. NYC Local Law 144 be­came op­er­at­ive on 1 Janu­ary 2023, with en­force­ment from 5 July 2023 under rules at 6 RCNY 5-300 and fol­low­ing. En­force­ment rests with two at­tor­neys. The State Comp­troller's audit, Re­port 2024-N-6 of 2 Decem­ber 2025, found two com­plaints in two years and a single in­stance of non-com­pli­ance among 32 com­pan­ies sur­veyed, where aud­it­ors re­view­ing the same com­pan­ies found at least 17 po­ten­tial in­stances. The RAISE Act, signed as Chapter 699 of 2025, was re­pealed and re-en­acted by Chapter 96 of 2026 on 27 March 2026, com­men­cing 1 Janu­ary 2027.

Any count of state AI law needs its count­ing rule pub­lished along­side it. The Trans­par­ency Co­ali­tion's 2026 State AI Le­gis­la­tion Re­port of 27 Oc­to­ber 2026 re­cords 115 new AI-re­lated laws passed or en­acted across 30 states between 1 Janu­ary and 15 Oc­to­ber 2026, ex­clud­ing data-centre meas­ures. MultiState's tracker re­cords 1,561 AI-re­lated bills in­tro­duced in 45 states as of March 2026. The NYU Cen­ter on Tech­no­logy Policy coun­ted 109 AI laws plus 28 sep­ar­ate data-centre laws to 1 July 2026. The Cen­ter for Demo­cracy and Tech­no­logy, count­ing pub­lic-sec­tor AI bills alone, found three en­acted out of 69 in­tro­duced. The range is three to 115.

Key takeaways
  • Col­or­ado SB 24-205 was delayed twice, left un­en­forced by un­der­tak­ing, then re­pealed on 14 May 2026. Its re­place­ment com­mences on 1 Janu­ary 2027.
  • Texas tests in­tent and Cali­for­nia tests ef­fect, so the same hir­ing tool can be law­ful in one state and ac­tion­able in the other.
AI governanceRules of the Game18 •
The same calendar year yields three to 115 state AI laws, depending on who is counting.
laws enacted in 2026
All AI laws115Centres apart109Cross-sector acts7Public sector3
Source: Transparency Coalition, NYU Center on Technology Policy and CDT, 2026
  • New York's audit found two com­plaints in two years against at least 17 po­ten­tial breaches, which is the clearest meas­ure of paper rules.

India: DPDP as the binding floor, synthetic content as the first duty

India's data law commences in stages to May 2027, while the labelling and takedown duties on synthetic content already bind.

India's bind­ing floor for AI is its data pro­tec­tion law, and most of it is not yet in force. The Di­gital Per­sonal Data Pro­tec­tion Act, 2023 (22 of 2023) was en­acted by Par­lia­ment on 11 Au­gust 2023. The Rules fol­lowed as G.S.R. 846(E) of 13 Novem­ber 2025, made under sec­tions 40(1) and 40(2). Com­mence­ment is staged. Rules 1, 2 and 17 to 21 took ef­fect on no­ti­fic­a­tion, cov­er­ing defin­i­tions and the con­sti­tu­tion of the Data Pro­tec­tion Board. Rule 4, on Con­sent Man­agers, com­mences on 13 Novem­ber 2026, and the sub­stant­ive du­ties at Rules 3, 5 to 16, 22 and 23 com­mence eight­een months after no­ti­fic­a­tion, on 13 May 2027.

The May 2027 date car­ries al­most everything a model builder would care about. It brings no­tice, se­cur­ity safe­guards, breach in­tim­a­tion, re­ten­tion lim­its, chil­dren's data rules, the ob­lig­a­tions of a Sig­ni­fic­ant Data Fi­du­ciary, data prin­cipal rights and the cross-bor­der trans­fer power. The pen­al­ties are stiff. They reach Rs 250 crore for fail­ure to main­tain reas­on­able se­cur­ity safe­guards, Rs 200 crore each for an un­no­ti­fied breach and for chil­dren's-data breaches, and Rs 50 crore oth­er­wise. Ap­peals from the Board lie to the Tele­com Dis­putes Set­tle­ment and Ap­pel­late Tribunal.

One clause reaches al­gorithms dir­ectly. Rule 13(3) re­quires a Sig­ni­fic­ant Data Fi­du­ciary to ob­serve due di­li­gence to verify that tech­nical meas­ures in­clud­ing al­gorithmic soft­ware ad­op­ted by it for host­ing, dis­play, pub­lish­ing, trans­mis­sion, stor­age or shar­ing of per­sonal data are not

AI governanceRules of the Game19 •

likely to pose a risk to the rights of Data Prin­cipals. Rule 13(1) and (2) add an im­pact as­sess­ment and an audit once every twelve months from des­ig­na­tion, with sig­ni­fic­ant ob­ser­va­tions re­por­ted to the Board. There is no AI-spe­cific ex­emp­tion and no text-and-data-min­ing ex­cep­tion. The clause was read in a spe­cial­ist re­pro­duc­tion rather than in the Gaz­ette file, which was trun­cated on re­trieval.

The ad­ju­dic­ator is a body in law and an ab­sence in fact. The Data Pro­tec­tion Board of India is es­tab­lished under sec­tion 18 of the Act as a body cor­por­ate and an in­de­pend­ent ad­ju­dic­at­ory au­thor­ity. A MeitY va­cancy cir­cu­lar of 6 May 2026 in­vited ap­plic­a­tions for one post of Chair­per­son and four posts of Mem­ber, with con­sol­id­ated monthly salar­ies of Rs 4,50,000 and Rs 4,00,000. No ap­point­ment could be veri­fied. A dated check by Pri­vacy Law Net­work on 21 Septem­ber 2026 found no Gaz­ette no­ti­fic­a­tion, min­istry doc­u­ment or press re­lease an­noun­cing one.

Syn­thetic con­tent is where India moved first and moved hard. G.S.R. 120(E) of 10 Feb­ru­ary 2026 amended the In­form­a­tion Tech­no­logy (In­ter­me­di­ary Guidelines and Di­gital Media Eth­ics Code) Rules, 2021, with ef­fect from 20 Feb­ru­ary 2026. New Rule 2(1)(wa) defines syn­thet­ic­ally gen­er­ated in­form­a­tion as audio, visual or audio-visual in­form­a­tion ar­ti­fi­cially or al­gorith­mic­ally cre­ated or altered using a com­puter re­source so that it ap­pears real and is likely to be per­ceived as in­dis­tin­guish­able from a nat­ural per­son or a real-world event. Text is out­side the defin­i­tion. Carve-outs cover routine good-faith edit­ing and the or­din­ary cre­ation of doc­u­ments and train­ing ma­ter­i­als where no false re­cord res­ults.

Pre­ven­tion comes first, la­belling second. Rule 3(3)(a)(i) re­quires reas­on­able and ap­pro­pri­ate tech­nical meas­ures, in­clud­ing auto­mated tools, to pre­vent users gen­er­at­ing syn­thetic con­tent that breaks the law, nam­ing the IT Act, the Bhar­atiya Nyaya San­hita 2023, the POCSO Act 2012 and the Ex­plos­ive Sub­stances Act 1908. Rule 3(3)(a)(ii) then re­quires all other syn­thetic con­tent to be prom­in­ently la­belled in the visual dis­play, or pre­fixed with an audio dis­clos­ure, and em­bed­ded with per­man­ent metadata in­clud­ing a unique iden­ti­fier for the ser­vice used. Rule 3(3)(b) for­bids en­abling re­moval of the label or the iden­ti­fier.

The same in­stru­ment cut four clocks. On ac­tual know­ledge from a court order or a reasoned writ­ten in­tim­a­tion, Rule 3(1)(d) now re­quires re­moval within three hours where it al­lowed thirty-six, and the in­tim­a­tion must come from an of­ficer not below Joint Sec­ret­ary, or Deputy In­spector Gen­eral where is­sued by the po­lice. A com­plaint about con­tent ex­pos­ing private areas, nud­ity or im­per­son­a­tion runs on two hours under Rule 3(2)(b), down from twenty-four.

AI governanceRules of the Game20 •

A re­moval com­plaint under the pro­viso to Rule 3(2)(a)(i) runs on thirty-six hours, down from sev­enty-two, and gen­eral griev­ance res­ol­u­tion falls from fif­teen days to seven. No en­force­ment ac­tion under these pro­vi­sions could be loc­ated.

The February 2026 amendment cut each hour-based takedown clock under the IT Rules.
hours allowed
02040608036Order, was3Order, now24Imagery, was2Imagery, now72Removal, was36Removal, now
Source: MeitY, G.S.R. 120(E), 2026
Three hours from a reasoned written intimation, where the Rules had allowed thirty-six until February 2026.
Key takeaways
  • The DPDP Rules com­mence in three stages. The du­ties that mat­ter to a model builder ar­rive on 13 May 2027, not now.
  • Rule 13(3) is the only In­dian pro­vi­sion in force that speaks to al­gorithms, and it binds des­ig­nated Sig­ni­fic­ant Data Fi­du­ciar­ies alone.
  • La­belling and proven­ance metadata on syn­thetic audio and video bind today, under rules in force since 20 Feb­ru­ary 2026. Gen­er­ated text does not.

India: the sector regulators, and the law held in reserve

Six Indian regulators are at work on AI, and only one of them has issued an instrument that binds anybody.

India de­cided not to le­gis­late ho­ri­zont­ally, and said so plainly. MeitY un­veiled the India AI Gov­ernance Guidelines on 5 Novem­ber 2025 under the IndiaAI Mis­sion, built on seven

AI governanceRules of the Game21 •

guid­ing prin­ciples, re­com­mend­a­tions across six pil­lars and an ac­tion plan on short, me­dium and long-term timelines. The Sec­ret­ary, MeitY, Shri S. Krish­nan, put the policy in one line: our focus re­mains on using ex­ist­ing le­gis­la­tion wherever pos­sible. The Guidelines are non-bind­ing and carry no pen­al­ties. They pro­pose an AI Gov­ernance Group sup­por­ted by a Tech­no­logy and Policy Ex­pert Com­mit­tee, and neither body is re­por­ted as con­sti­tuted.

Our focus remains on using existing legislation wherever possible, said the Secretary, MeitY, on 5 November 2025.

The Re­serve Bank of India has pro­duced a re­port and a draft, and no cir­cu­lar. Its FREE-AI com­mit­tee re­port, on re­spons­ible and eth­ical en­able­ment of AI in the fin­an­cial sec­tor, was re­leased on 13 Au­gust 2025 and runs to seven sut­ras and twenty-six re­com­mend­a­tions across six pil­lars. It is a com­mit­tee re­port. It cre­ates no ob­lig­a­tion on a reg­u­lated en­tity. The draft Guid­ance on Reg­u­lat­ory Prin­ciples for Model Risk Man­age­ment, 2026 was is­sued for com­ment on 24 June 2026, with com­ments clos­ing on 24 July 2026. It re­mains a draft, and no final cir­cu­lar could be loc­ated.

The se­cur­it­ies reg­u­lator has an­nounce­ments, not rules. No SEBI ar­ti­fi­cial in­tel­li­gence or ma­chine learn­ing in­stru­ment could be loc­ated. On 10 Oc­to­ber 2026 the Chair­man, Tuhin Kanta Pandey, said guidelines for cap­ital mar­kets would issue shortly, de­scrib­ing a tiered ap­proach with ac­count­ab­il­ity, data con­trols, kill switches and human over­sight, and stat­ing that every reg­u­lated en­tity re­mains re­spons­ible for any AI tool it uses. The old­est AI-touch­ing In­dian in­stru­ment found is a 2019 SEBI cir­cu­lar re­quir­ing mar­ket in­fra­struc­ture in­sti­tu­tions to make quarterly dis­clos­ures on AI and ma­chine learn­ing ap­plic­a­tions.

Med­ical device soft­ware is the one place an In­dian reg­u­lator has is­sued a final AI-aware doc­u­ment. The Cent­ral Drugs Stand­ard Con­trol Or­gan­isa­tion is­sued its Guid­ance Doc­u­ment on Med­ical Device Soft­ware, Doc No. CDSCO/MD/GD/MDSW/01/2026, by cir­cu­lar dated 21 July 2026 under the Med­ical Devices Rules, 2017. It re­places the earlier split between stan­dalone and em­bed­ded soft­ware with one term, which takes the risk class of the hard­ware it drives. It re­cog­nises AI and ma­chine learn­ing based soft­ware with a med­ical pur­pose and re­quires post-mar­ket sur­veil­lance for per­form­ance, bias and drift. An Al­gorithm Change Pro­tocol may be de­vised, wherever ap­plic­able. So it is dis­cre­tion­ary. Whether such guid­ance binds at all is dis­puted in the sources, and the CDSCO text could not be opened.

AI governanceRules of the Game22 •

In­sur­ance and tele­coms are earlier still. IRDAI con­sti­tuted a seven-mem­ber work­ing group on AI gov­ernance in in­sur­ance by of­fice order dated 17 June 2026, re­port­ing in three months, and sep­ar­ately dir­ec­ted in­surers to file an ac­tion taken re­port on AI cyber read­i­ness by 22 May 2026. No IRDAI AI reg­u­la­tion has is­sued. TRAI re­com­men­ded in July 2023 that an in­de­pend­ent stat­utory Ar­ti­fi­cial In­tel­li­gence and Data Au­thor­ity of India be cre­ated, and it was not. On 30 Oc­to­ber 2026 TRAI re­leased draft guidelines on the re­spons­ible use of AI in tele­com­mu­nic­a­tion net­works, with com­ments due on 27 Novem­ber 2026.

A ho­ri­zontal law is held in re­serve rather than ruled out. On 8 Oc­to­ber 2026 the Union Min­is­ter for Elec­tron­ics and In­form­a­tion Tech­no­logy, Shri Ash­wini Vaish­naw, an­nounced that MeitY would re­lease a reg­u­la­tion-fo­cused con­sulta­tion paper within a month, built on safety, a human-first ap­proach, skilling and leav­ing no per­son be­hind. He called the ap­proach techno-legal. The paper could not be veri­fied as is­sued. Re­ports dis­agree on whether it leads to a stan­dalone AI law, amend­ments to ex­ist­ing law, or fur­ther re­fine­ment of the techno-legal frame­work.

No In­dian rule reaches an autonom­ous agent. No defin­i­tion of an AI agent ap­pears in the DPDP Act or Rules, in the IT Rules, or in any loc­ated in­stru­ment of the fin­an­cial and tele­com reg­u­lat­ors. The nearest ana­logue in a bind­ing in­stru­ment is the DPDP Con­sent Man­ager, a per­son re­gistered with the Board who lets a Data Prin­cipal give, man­age and with­draw con­sent through an in­ter­op­er­able plat­form, and that duty com­mences on 13 Novem­ber 2026. The second nearest is the Ac­count Ag­greg­ator, a class cre­ated by RBI Mas­ter Dir­ec­tion of 2 Septem­ber 2016, with sev­en­teen com­pan­ies re­gistered.

Most Indian AI-touching measures are advisory, draft or merely announced.
instruments located, November 2026
Advisory4Binding, in force2Draft, closed2Announced only2Binding, pending1Draft, open1
Source: MeitY, RBI, SEBI, CDSCO, IRDAI and TRAI instruments, 2026
Key takeaways
  • The Re­serve Bank has an ad­vis­ory re­port of 13 Au­gust 2025 and a draft guid­ance whose com­ments closed on 24 July 2026. No cir­cu­lar binds yet.
AI governanceRules of the Game23 •
  • SEBI has said guidelines will come shortly, which it said on 10 Oc­to­ber 2026. Until then li­ab­il­ity for the tool rests with the reg­u­lated en­tity.
  • CDSCO's guid­ance of 21 July 2026 is the only final AI-aware sec­tor doc­u­ment, and com­ment­at­ors dis­agree on whether it binds at all.

Saudi Arabia: SDAIA, and what actually binds

Saudi Arabia has issued six AI documents in under three years. None is a law, and what binds is the data regime.

In Saudi Ar­a­bia the en­force­able in­stru­ment is the data law, and the AI in­stru­ments sit above it as guid­ance. The Per­sonal Data Pro­tec­tion Law entered into force on 14 Septem­ber 2023, with a one-year grace period for con­trol­lers end­ing 14 Septem­ber 2024, and its im­ple­ment­ing reg­u­la­tions were pub­lished by SDAIA on 7 Septem­ber 2023. SDAIA is the com­pet­ent au­thor­ity. Com­ment­ary re­cords that SDAIA en­force­ment com­mit­tees is­sued 48 formal de­cisions dur­ing 2025 against or­gan­isa­tions found in vi­ol­a­tion, a fig­ure that could not be traced to an SDAIA pub­lic­a­tion. The po­s­i­tion is the in­verse of the UAE's, where the data law waits on its reg­u­la­tions.

The data protection law is in force and enforced. The AI instruments sitting above it are advisory.

SDAIA's AI out­put is ad­vis­ory. The AI Eth­ics Prin­ciples were is­sued in final form as Ver­sion 1.0 in Septem­ber 2023, with a Ver­sion 2.0 since, and they are not leg­ally bind­ing, al­though non-com­pli­ance may still at­tract en­force­ment under other in­stru­ments such as the data law. Gen­er­at­ive AI Guidelines for gov­ern­ment en­tit­ies fol­lowed on 10 Janu­ary 2024. Deep­fakes Guidelines were is­sued on 8 May 2026, cov­er­ing pro­tec­tion of per­sonal data, ex­pli­cit con­sent, di­gital wa­ter­marks and source veri­fic­a­tion. The Saudi Press Agency de­scribed that doc­u­ment as man­dat­ing ad­her­ence to strict stand­ards, yet named no en­force­ment mech­an­ism.

The most re­cent frame­work is the na­tional AI risk man­age­ment frame­work, launched on 14 July 2026. It is de­scribed as a uni­fied na­tional meth­od­o­logy for identi­fy­ing, as­sess­ing, treat­ing and mon­it­or­ing AI risks, in­ten­ded to en­able pub­lic and private en­tit­ies to adopt AI safely. Its

AI governanceRules of the Game24 •

meth­od­o­logy runs to four phases, end­ing in con­tinu­ous mon­it­or­ing and re­view. Risk levels are cal­cu­lated on a mat­rix link­ing prob­ab­il­ity of oc­cur­rence to scale of im­pact. The frame­work rests on seven core prin­ciples and clas­si­fies AI risks into seven main types. The an­nounce­ment nowhere states that the frame­work is man­dat­ory. It uses the lan­guage of en­able­ment.

One policy might have changed the ad­vis­ory char­ac­ter. It re­mains un­fin­ished. SDAIA opened a con­sulta­tion on a draft Re­spons­ible AI Policy on 3 April 2026 and closed it on 3 May 2026. No source opened re­ports a fi­nal­ised policy or an ad­op­tion date, and the risk tiers re­por­ted by com­ment­at­ors dur­ing the con­sulta­tion win­dow could not be con­firmed against an SDAIA doc­u­ment. The ref­er­ence num­ber and doc­u­ment date cir­cu­lat­ing for the risk man­age­ment frame­work rest on sec­ond­ary data­bases alone, be­cause the SDAIA site re­jec­ted dir­ect re­trieval.

Sec­tor reg­u­lat­ors have moved in nar­row places. SAMA's rule­book states, in its fraud risk pro­vi­sions, that where ma­chine learn­ing or ar­ti­fi­cial in­tel­li­gence is used the sys­tem should not be a black box and should be cap­able of being audited. There is no stan­dalone uni­fied Saudi frame­work gov­ern­ing AI use in fin­an­cial in­sti­tu­tions, so banks fall under ex­ist­ing rules on data gov­ernance, cy­ber­se­cur­ity, risk man­age­ment and cus­tomer data pro­tec­tion. The Cap­ital Mar­ket Au­thor­ity con­sul­ted on reg­u­lat­ing robo-ad­vis­ory in Au­gust 2025 and ap­proved the frame­work in March 2026. Robo-ad­vis­ory port­fo­lios reached SAR 6.4 bil­lion, up 86 per cent year on year, on the Au­thor­ity's 2025 an­nual re­port.

The med­ical device pic­ture is sharpest. The Saudi Food and Drug Au­thor­ity's guid­ance on re­view and ap­proval of ar­ti­fi­cial in­tel­li­gence and big data based med­ical devices, as pos­ted on the na­tional con­sulta­tion plat­form, car­ries the doc­u­ment num­ber MDS - G00, the ver­sion num­ber 0.0 and a ver­sion date of 00/00/2020. It is, on its face, an un­numbered and un­dated draft. It would set a four-tier ver­sion con­trol scheme, under which a change in op­er­at­ing prin­ciples, in­ten­ded use or per­form­ance bey­ond the ap­proved ac­cur­acy range re­quires change ap­proval, while a train­ing data change that stays within that range is left to the man­u­fac­turer's own con­trol.

This is not for want of time. Vis­ion 2030 was launched in 2016, SDAIA was cre­ated in 2019 and the Na­tional Strategy for Data and AI was an­nounced in 2020, so the policy ar­chi­tec­ture pred­ates most of the world's AI stat­utes. The In­ter­na­tional Centre for Ar­ti­fi­cial In­tel­li­gence Re­search and Eth­ics is af­fil­i­ated with SDAIA, and the GCC Gen­eral Sec­ret­ariat names it as a

AI governanceRules of the Game25 •

part­ner in Gulf AI co­oper­a­tion. Six years on, the bind­ing layer is still the data law. An or­gan­isa­tion de­ploy­ing a model in the King­dom an­swers for its per­sonal data, not for its model.

Key takeaways
  • No Saudi in­stru­ment spe­cific to AI binds a private com­pany. The Per­sonal Data Pro­tec­tion Law of 14 Septem­ber 2023 and its reg­u­la­tions do.
  • SDAIA has is­sued eth­ics prin­ciples, gen­er­at­ive AI guidelines, deep­fake guidelines and a risk man­age­ment frame­work. Each is guid­ance, not law.
  • The draft Re­spons­ible AI Policy closed con­sulta­tion on 3 May 2026 and could not be veri­fied as ad­op­ted, so plan on the data law for now.

The UAE: where the rule sits below the headline

There is no federal UAE AI law. The binding obligations sit in a financial free zone and in two emirate health regulators.

There is no fed­eral AI law. The UAE's new­est na­tional body is drawn in terms of gov­ern­ment it­self. On 14 June 2026 Sheikh Mo­hammed bin Rashid Al Mak­toum ap­proved the es­tab­lish­ment of the Ar­ti­fi­cial In­tel­li­gence and Data Au­thor­ity, de­scribed as the single na­tional body for data, ar­ti­fi­cial in­tel­li­gence and di­gital gov­ern­ment, re­port­ing dir­ectly to the Cab­inet. It con­sol­id­ates the former AI of­fice, the di­gital gov­ern­ment sec­tor of the tele­com­mu­nic­a­tions reg­u­lator and the UAE Data Of­fice, under Omar Sul­tan Al Olama as Chair­man. Its man­date cov­ers pro­pos­ing policies and stand­ards, and en­sur­ing com­pli­ance across fed­eral en­tit­ies. The Cab­inet state­ment con­fers no rule-mak­ing power over the private sec­tor.

What sits at the fed­eral level is a charter and an un­fin­ished data law. The UAE Charter for the De­vel­op­ment and Use of Ar­ti­fi­cial In­tel­li­gence was pub­lished in July 2024 and sets out twelve gen­eral prin­ciples. It de­scribes it­self as a guid­ing frame­work. It con­tains no ob­lig­a­tions, no ad­dress­ees, no thresholds and no en­force­ment pro­vi­sion, and its twelfth prin­ciple refers com­pli­ance back to other in­stru­ments. Fed­eral De­cree-Law No. 45 of 2021, the data pro­tec­tion law, re­mains in force but its Ex­ec­ut­ive Reg­u­la­tions have still not been is­sued, which one ad­viser sum­mar­ised in June 2026 as in force but un­en­force­able.

The one in­stru­ment in the re­gion draf­ted for autonomy rather than auto­ma­tion is a free zone rule. Reg­u­la­tion 10 on the pro­cessing of per­sonal data through autonom­ous and

AI governanceRules of the Game26 •

semi-autonom­ous sys­tems was en­acted with amend­ments to the DIFC Data Pro­tec­tion Reg­u­la­tions an­nounced on 7 Septem­ber 2023, and guid­ance dated 1 Septem­ber 2023. Purely auto­mated sys­tems are de­lib­er­ately ex­cluded, be­cause sys­tems with no de­gree of autonomy whose op­er­a­tion is de­term­in­ist­ic­ally con­trolled by hu­mans were not meant to be caught. Autonomy is the trig­ger, not the tech­no­logy.

Two new roles re­place con­trol­ler and pro­cessor. A De­ployer is the per­son or en­tity that au­thor­ises or be­ne­fits from the op­er­a­tion of a Sys­tem and its out­put, car­ry­ing the gen­eral re­spons­ib­il­it­ies of a con­trol­ler. An Op­er­ator is the tech­nical ser­vice pro­vider act­ing on the De­ployer's in­struc­tions. The guid­ance states the ana­logy out­right, that a Sys­tem op­er­at­ing autonom­ously under the au­thor­ity and for the be­ne­fit of its De­ployer is sub­stan­tially sim­ilar to an em­ployee. Pur­poses the Sys­tem gen­er­ates for it­self must rest on an ex­haust­ive set of prin­ciples hard coded by hu­mans, which the Sys­tem can­not change.

A System operating autonomously for its Deployer is substantially similar to an employee, and the Deployer is liable for its actions.

The high-risk gate is closed and the key has not been cut. Reg­u­la­tion 10.3.3 pro­hib­its mak­ing a Sys­tem avail­able for com­mer­cial use in high-risk pro­cessing un­less audit and cer­ti­fic­a­tion re­quire­ments have been es­tab­lished, the Sys­tem com­plies with them, it pro­cesses per­sonal data solely for human-defined or human-ap­proved pur­poses, and an Autonom­ous Sys­tems Of­ficer is ap­poin­ted. On DIFC's own ac­count, no Sys­tem may be used for high-risk pro­cessing until the Com­mis­sioner pro­mul­gates those cer­ti­fic­a­tion re­quire­ments. Two sub-com­mit­tees were re­quired to re­com­mend the cri­teria by 30 April 2024. Con­sulta­tion Paper No. 3 of June 2026 was still pro­pos­ing the ma­chinery, roughly 33 months after en­act­ment. Whether those amend­ments have been en­acted could not be veri­fied.

ADGM has gone no fur­ther. It has no AI-spe­cific in­stru­ment, al­though sec­tion 20 of its Data Pro­tec­tion Reg­u­la­tions 2021 cov­ers auto­mated de­cision-mak­ing and pro­fil­ing much as the DIFC's do. Di­gital Dubai pub­lished AI Eth­ics Prin­ciples and Guidelines with a self-as­sess­ment tool in Janu­ary 2019, and the of­fi­cial doc­u­ment states that the guidelines are non-bind­ing. En­tit­ies in either free zone must still com­ply with fed­eral health ob­lig­a­tions under Fed­eral Law No. 2 of 2019, which re­quires local stor­age and pro­cessing of health data sub­ject to stated ex­cep­tions.

AI governanceRules of the Game27 •

Health is where the ob­lig­a­tion bites. The Abu Dhabi De­part­ment of Health pub­lished its policy on the use of ar­ti­fi­cial in­tel­li­gence in the emir­ate's health­care sec­tor on 30 April 2018, the re­gion's first emir­ate-level AI gov­ernance frame­work, reach­ing li­censed pro­viders, in­surers, re­search­ers and any end user work­ing with Abu Dhabi pa­tient data. It sets six guid­ing prin­ciples, re­quires reg­u­lar audits with re­port­ing, and re­tains the power to im­pose sanc­tions. A Re­spons­ible Ar­ti­fi­cial In­tel­li­gence Stand­ard fol­lowed in 2025. The Dubai Health Au­thor­ity is­sued its own policy in Au­gust 2021 as cir­cu­lar CIR-2021-00000141, en­forced through its Health Reg­u­la­tion Sec­tor. Both re­quire grace­ful de­grad­a­tion on mal­func­tion.

Key takeaways
  • The fed­eral layer is a charter with no ob­lig­a­tions and a data law whose Ex­ec­ut­ive Reg­u­la­tions are still out­stand­ing five years after en­act­ment.
  • DIFC Reg­u­la­tion 10 as­signs li­ab­il­ity for an autonom­ous sys­tem to its De­ployer on an em­ploy­ment ana­logy, and it ap­plies only within the free zone.
  • The two emir­ate health policies, though called policies, op­er­ate as bind­ing reg­u­la­tion with audits and sanc­tions be­hind them.

Qatar

One Qatari AI instrument plausibly binds, and it reaches only entities licensed by the central bank.

Only one Qatari AI in­stru­ment plaus­ibly binds. The Qatar Cent­ral Bank is­sued its Ar­ti­fi­cial In­tel­li­gence Guideline on 4 Septem­ber 2024, reg­u­lat­ing the use of AI by li­censed en­tit­ies, and it is re­cor­ded as in force from that date. Qatar News Agency re­por­ted the same day that the guideline was is­sued under the Third Fin­an­cial Sec­tor Strategy and the FinTech Strategy. It ap­plies to all li­censed en­tit­ies that de­velop, pro­cure, out­source, op­er­ate or provide AI sys­tems, so the trig­ger is the li­cence rather than the tech­no­logy. Sec­ond­ary sources de­scribe it as leg­ally bind­ing on those en­tit­ies. That char­ac­ter­isa­tion rests on com­ment­ary, be­cause the text it­self could not be reached.

The ob­lig­a­tions read con­ven­tion­ally. Prior cent­ral bank ap­proval is re­quired be­fore launch­ing a new AI sys­tem as a pro­vider or ma­ter­i­ally modi­fy­ing an ex­ist­ing one, and again be­fore sign­ing a high-risk pur­chase, li­cens­ing or out­sourcing agree­ment. A li­censed en­tity must keep an

AI governanceRules of the Game28 •

up­dated re­gister of AI sys­tems, run risk as­sess­ments that clas­sify high-risk sys­tems, adopt board-ap­proved AI gov­ernance policies, main­tain human over­sight of high-risk sys­tems and give cus­tom­ers a mech­an­ism to query and re­quest re­view of an AI-driven de­cision af­fect­ing them. The clause num­bers and the ob­lig­a­tion list come from a reg­u­lat­ory data­base, not from the cent­ral bank, and that data­base marks its own entry as un­con­firmed.

Pen­al­ties are not sep­ar­ately provided. The po­s­i­tion is de­scribed as sub­ject to the cent­ral bank's gen­eral su­per­vis­ory en­force­ment powers rather than a ded­ic­ated AI pen­alty re­gime. A li­censed en­tity faces su­per­vis­ory con­sequences rather than a stated civil pen­alty per vi­ol­a­tion. For a bank already in­side a pruden­tial re­la­tion­ship, the prac­tical ef­fect may be stronger than a fine. For any­body out­side that re­la­tion­ship, there is no ef­fect at all.

An announced intention is not an instrument, and in Qatar's capital markets the AI rules remain exactly that.

The cap­ital mar­kets reg­u­lator has an in­ten­tion. In May 2025, dur­ing the 50th An­nual Meet­ing of IOSCO held in Qatar from 12 to 14 May, the Qatar Fin­an­cial Mar­kets Au­thor­ity said a draft reg­u­la­tion on AI use in the fin­an­cial mar­ket would be is­sued soon, framed around trans­par­ency and con­sist­ency and tied to its Stra­tegic Plan 2023 to 2027. No source opened con­firms that the draft has been pub­lished for con­sulta­tion or ad­op­ted. The most re­cent cov­er­age loc­ated is from May 2025, and it de­scribes an in­ten­tion rather than an in­stru­ment.

The data law pred­ates the AI de­bate. Qatar's Per­sonal Data Pri­vacy Pro­tec­tion Law is Law No. 13 of 2016, pub­lished in the Of­fi­cial Gaz­ette on 29 Decem­ber 2016 and ef­fect­ive in 2017. It af­fords extra pro­tec­tion to data re­lat­ing to chil­dren, health, eth­nic ori­gin and re­li­gious be­liefs, and pro­cessing those cat­egor­ies re­quires per­mis­sion from the Com­pet­ent De­part­ment. The re­por­ted max­imum pen­alty is QAR 5 mil­lion. Sources dis­agree on which body reg­u­lates on­shore, nam­ing either a di­vi­sion of the na­tional cyber se­cur­ity agency or a na­tional data pro­tec­tion of­fice, and neither ac­count could be traced to a primary source.

As in the UAE, a free zone runs its own re­gime. The Qatar Fin­an­cial Centre op­er­ates under the QFC Data Pro­tec­tion Reg­u­la­tions 2021, ap­ply­ing to en­tit­ies in­cor­por­ated or re­gistered in the Centre, su­per­vised by the QFC Data Pro­tec­tion Of­fice and re­quir­ing a data pro­tec­tion im­pact as­sess­ment for high-risk pro­cessing. So an or­gan­isa­tion in Qatar must es­tab­lish which

AI governanceRules of the Game29 •

of two data re­gimes ap­plies be­fore it es­tab­lishes what ap­plies to its mod­els. The an­swer turns on re­gis­tra­tion. Neither re­gime is AI-spe­cific, and neither sets a la­belling or proven­ance duty of the kind India brought into force in Feb­ru­ary 2026.

Above that, everything is strategy or draft. Qatar's Na­tional Ar­ti­fi­cial In­tel­li­gence Strategy was pub­lished by the then Min­istry of Trans­port and Com­mu­nic­a­tions and is built on six pil­lars, provid­ing re­com­mend­a­tions rather than bind­ing rules. Its date is dis­puted, given as 2019 by one source and early 2022 by an­other, and the dis­crep­ancy is un­re­solved. A draft Na­tional AI Policy ex­ists and is ex­pressly a con­sulta­tion draft, with an ad­viser writ­ing in Novem­ber 2025 de­scrib­ing Qatar as mov­ing to­ward es­tab­lish­ing both the policy and a na­tional centre. A draft Cab­inet de­cision ap­proved in Septem­ber 2025 would cre­ate that centre.

Key takeaways
  • The cent­ral bank guideline of 4 Septem­ber 2024 is the only Qatari AI in­stru­ment that plaus­ibly binds, and only li­censed fin­an­cial en­tit­ies are caught.
  • The cap­ital mar­kets reg­u­lator said in May 2025 that a draft would fol­low soon. No draft, con­sulta­tion or ad­op­ted rule could be loc­ated since.
  • Two data re­gimes run in par­al­lel, the 2016 on­shore law and the 2021 free zone reg­u­la­tions, and neither is spe­cific to ar­ti­fi­cial in­tel­li­gence.

Kuwait, Oman and Bahrain: strategy without instrument

All three have published AI strategies. None has a binding AI instrument that reaches a private company.

Bahrain's AI bill was not en­acted. The draft­ing ap­pears to have star­ted again. A thirty-eight art­icle bill was ap­proved un­an­im­ously by the Shura Coun­cil in April 2024, with a re­quire­ment that the gov­ern­ment draft it prop­erly and refer it to Par­lia­ment within six months. It then met res­ist­ance in the Coun­cil of Rep­res­ent­at­ives, where the gov­ern­ment pro­posed a reg­u­lat­ory sand­box in­stead, and it had not been en­acted as of July 2026. A re­port of 19 June 2026 re­cords the Shura Coun­cil's le­gis­lat­ive com­mit­tee as still draft­ing le­gis­la­tion on ar­ti­fi­cial in­tel­li­gence, draw­ing on both the European frame­work and the Arab model law. No source opened re­cords a pro­mul­gated Bahraini AI law.

AI governanceRules of the Game30 •

In Bahrain a gov­ern­ment policy binds. The In­form­a­tion and eGovernment Au­thor­ity is­sued the Gen­eral Policy for the Use of Ar­ti­fi­cial In­tel­li­gence, Ver­sion 1.0, dated 20 May 2025, ap­proved by the Min­is­terial Com­mit­tee for In­form­a­tion and Com­mu­nic­a­tion Tech­no­logy that day, from which date all gov­ern­ment en­tit­ies must com­ply. Its scope does not reach the private sec­tor. It sets el­even prin­ciples, the first re­quir­ing that the final de­cision in im­port­ant mat­ters re­main under human con­trol. Neg­at­ive ef­fects of an AI solu­tion are at­trib­uted to the re­spons­ible in­di­vidu­als, not to the ma­chine.

Any negative effects resulting from these solutions are attributed to the responsible individuals, not to the machine.

The policy points at other laws. Its first pil­lar re­quires com­pli­ance with a named set of ex­ist­ing in­stru­ments, among them the Per­sonal Data Pro­tec­tion Law of 2018, the Cy­ber­crime Law of 2014, the de­cree on cloud com­put­ing ser­vices to for­eign parties, the Open Data Policy and the GCC Ar­ti­fi­cial In­tel­li­gence Eth­ics Guideline. The data law was en­acted on 12 July 2018 and came into force on 1 Au­gust 2019. The policy re­cords that it in­cludes pro­vi­sions on pro­cessing by auto­mated means and the con­di­tions that make such pro­cessing law­ful.

Oman's po­s­i­tion is settled once the two in­stru­ments are read apart. The Min­istry of Trans­port, Com­mu­nic­a­tions and In­form­a­tion Tech­no­logy is­sued the Pub­lic Policy for Safe and Eth­ical Use of Ar­ti­fi­cial In­tel­li­gence on 1 April 2025, pre­pared in line with Is­lamic prin­ciples, with over­sight shared between the min­istry and sec­tor reg­u­lat­ors. UNESCO de­scribes a risk-based frame­work ap­plic­able to pub­lic en­tit­ies and rel­ev­ant private-sec­tor act­ors, and notes that en­force­ment pro­ced­ures re­main in early op­er­a­tional stages. Pub­lic-sec­tor pro­cure­ment of AI re­quires pre-de­ploy­ment as­sess­ments, formal ap­proval by the com­pet­ent au­thor­ity and li­fe­cycle mon­it­or­ing. No AI-spe­cific li­ab­il­ity re­gime has been ad­op­ted.

The Omani royal de­cree that is often read as AI reg­u­la­tion is not. Royal De­cree No. 50/2026 of 30 April 2026 es­tab­lished an Ar­ti­fi­cial In­tel­li­gence Spe­cial Zone in Mus­cat, ex­tend­ing spe­cial eco­nomic zone in­cent­ives in­clud­ing tax breaks to pro­jects in the zone. Ad­visers writ­ing in June 2026 state plainly that the de­cree does not in­tro­duce spe­cific ar­ti­fi­cial in­tel­li­gence reg­u­la­tion. A bind­ing in­stru­ment about AI there­fore ex­ists in Oman, and it is an in­vest­ment meas­ure. Sep­ar­ately, the Omani data law was pro­mul­gated by Royal De­cree No. 6/2022 on 9

AI governanceRules of the Game31 •

Feb­ru­ary 2022 and entered into force a year later, with the com­pli­ance win­dow ex­ten­ded to 5 Feb­ru­ary 2026.

Kuwait has neither an AI law nor an AI reg­u­lator. Its data in­stru­ment is the Data Pri­vacy Pro­tec­tion Reg­u­la­tion of April 2021, ad­min­istered by the tele­com­mu­nic­a­tions reg­u­lator, which car­ries a 72-hour breach no­ti­fic­a­tion duty and fines up to KWD 50,000. The cita­tion is con­tested. DLA Piper cites a 2024 ad­min­is­trat­ive de­cision and states that Kuwait has no ded­ic­ated per­sonal data pro­tec­tion law ap­ply­ing to all nat­ural and jur­istic per­sons. Cham­bers re­cords that the Kuwaiti reg­u­la­tion con­tains no pro­vi­sions spe­cific­ally ad­dress­ing AI. A Na­tional Ar­ti­fi­cial In­tel­li­gence Strategy for 2025 to 2028 is re­cor­ded, and the reg­u­lator has is­sued a non-bind­ing guide on AI eth­ics.

The bloc has named the gap it­self. The GCC Gen­eral Sec­ret­ariat ap­proved the form­a­tion of an ar­ti­fi­cial in­tel­li­gence and emer­ging tech­no­lo­gies team in June 2021, which has pro­duced three named frame­works. A uni­fied Gulf guid­ing strategy was still in pre­par­a­tion as of 7 Oc­to­ber 2025. The Sec­ret­ary Gen­eral named the ab­sence of uni­fied le­gis­lat­ive stand­ards as the first re­main­ing chal­lenge, and said that mem­ber states' AI in­vest­ments have reached tens of bil­lions of dol­lars over the past few years, with plans to in­crease them to hun­dreds of bil­lions by 2030. No GCC state has en­acted an om­ni­bus AI stat­ute.

In the Gulf the binding AI rule sits below the national line, where it exists at all.
instruments located, by level
Advisory national6Sector regulator3Government only2Free zone1National, private0
Source: Volge review of Gulf instruments, 2026
Key takeaways
  • Bahrain's 2024 bill was never en­acted, and the Shura Coun­cil's com­mit­tee was still draft­ing in June 2026, draw­ing on European and Arab mod­els.
  • Oman's Royal De­cree 50/2026 is bind­ing and is about AI, but it cre­ates an in­vest­ment zone rather than reg­u­lat­ing any AI sys­tem.
  • Kuwait has a strategy for 2025 to 2028, a con­tested data reg­u­la­tion with no AI pro­vi­sions, and a non-bind­ing eth­ics guide. Noth­ing else.
AI governanceRules of the Game32 •

The part in brief

  1. 01The United States has one free-standing federal AI statute, the TAKE IT DOWN Act of 19 May 2025, and executive orders decline to authorise any licensing of new models.
  2. 02American state law is where the duties live, with Colorado's Act repealed on 14 May 2026 before commencing and Texas testing intent while California tests effect.
  3. 03India's binding floor is its data protection law, whose substantive duties arrive on 13 May 2027, while synthetic audio and video labelling has bound since 20 February 2026.
  4. 04The UAE's federal layer is a charter without obligations and a data law still lacking Executive Regulations, while Qatar's only plausibly binding AI rule reaches licensed financial entities.
  5. 05In Saudi Arabia only the data protection law binds a private company, and no Gulf state has enacted an omnibus AI statute, Bahrain's 2024 bill included.
AI governanceRules of the Game33 •
The map

A System operating autonomously for its Deployer is substantially similar to an employee, and the Deployer is liable for its actions.

03

The national champions

When the state owns the model, and also writes the rule.
  • When the state owns the model, what changes
  • Humain, and Saudi industrial policy for AI
  • QAI, and Qatar's bet on sovereign compute
  • G42, MGX, and the UAE's two-track position
  • India: the AI Mission, and the sovereign model question
  • Who regulates a national champion?

When the state owns the model, what changes

Every Gulf government is at once the author of its AI rules and the sponsor of the systems those rules would govern.

No Gulf in­stru­ment ad­dresses the case where the state owns the model. Re­search to 6 Novem­ber 2026 loc­ated six na­tional AI strategies across the six states of the Gulf Co­oper­a­tion Coun­cil and no na­tional AI frame­work law reach­ing the private sec­tor. The gap is struc­tural. Where a bind­ing rule ex­ists it sits below the na­tional line, in a free zone or in­side a sec­tor reg­u­lator, and it speaks to li­censees rather than to own­ers.

Saudi Ar­a­bia shows the shape of the prob­lem. The Saudi Data and Ar­ti­fi­cial In­tel­li­gence Au­thor­ity writes the na­tional AI in­stru­ments and also ad­min­is­ters en­force­ment of the Per­sonal Data Pro­tec­tion Law, under which, on com­ment­ary from Cap­tain Com­pli­ance, its en­force­ment com­mit­tees is­sued 48 formal de­cisions dur­ing 2025. Its AI Eth­ics Prin­ciples do not bind. Non-com­pli­ance may still at­tract en­force­ment under the data law, which is an in­dir­ect route and not a rule about mod­els.

The Emir­ates went fur­ther. On 14 June 2026 the Cab­inet ap­proved an Ar­ti­fi­cial In­tel­li­gence and Data Au­thor­ity, de­scribed in its own an­nounce­ment as the single na­tional body re­spons­ible for data, ar­ti­fi­cial in­tel­li­gence and di­gital gov­ern­ment, re­port­ing dir­ectly to the Cab­inet. The man­date cov­ers na­tional policy and le­gis­la­tion as well as stand­ards for data and AI man­age­ment, with a duty of en­sur­ing com­pli­ance across fed­eral en­tit­ies. One body pro­poses the rule and then checks the en­tity ap­ply­ing it.

In several Gulf states the body that writes the AI rule also sponsors the programme the rule would govern.

The same state­ment does not con­fer rule-mak­ing power over the private sec­tor. That mat­ters to a buyer, be­cause a state-backed de­veloper in­cor­por­ated as a com­pany sits out­side the fed­eral-en­tity peri­meter the Au­thor­ity is drawn around. Noth­ing loc­ated fills the space. The UAE Charter of July 2024 sets twelve prin­ciples and con­tains neither ad­dress­ees nor thresholds, and no en­force­ment pro­vi­sion at all.

AI governanceRules of the Game34 •
Six Gulf states have published an AI strategy and none has enacted a national AI framework law.
GCC states, count
01.534.566AI strategy0Framework law3Binding AI rule
Source: SPA, UAE Cabinet, QCB, MTCIT and Bahrain iGA, 2019 to 2026

Where a Gulf in­stru­ment does reach the state, it reaches the state as a user. Bahrain's Gen­eral Policy for the Use of Ar­ti­fi­cial In­tel­li­gence, ap­proved on 20 May 2025, ap­plies to all gov­ern­ment en­tit­ies and not to the private sec­tor. It at­trib­utes any neg­at­ive ef­fect of an AI solu­tion to the re­spons­ible in­di­vidu­als rather than to the ma­chine. Oman's Pub­lic Policy of 1 April 2025 works the same way, with com­pli­ance and over­sight sit­ting in the min­istry and in sec­tor reg­u­lat­ors. Neither speaks to own­er­ship.

The ma­chinery be­hind those policies is young. UNESCO's 2026 coun­try pro­file of Oman re­cords that eval­u­ation and en­force­ment pro­ced­ures re­main in early op­er­a­tional stages. It also re­cords that no cer­ti­fied re­gistry of AI vendors is main­tained and that no AI-spe­cific li­ab­il­ity re­gime has been ad­op­ted. A buyer can­not lean on ap­par­atus that is not yet run­ning.

In­ter­na­tional law of­fers a ref­er­ence point the Gulf has not taken up. The Coun­cil of Europe Frame­work Con­ven­tion, opened for sig­na­ture on 5 Septem­ber 2024, cov­ers the use of AI sys­tems by pub­lic au­thor­it­ies and by private act­ors act­ing on their be­half. Its sig­nat­or­ies num­ber 21. No Gulf state is among them, and the treaty has one Party, the European Union, so the nearest thing to a rule about the state as owner has not entered into force.

Key takeaways
  • Six Gulf states have an AI strategy and none has a na­tional AI frame­work law, so the own­er­ship ques­tion is left un­asked rather than re­solved any­where.
  • Saudi Ar­a­bia's AI au­thor­ity writes the ad­vis­ory frame­works and en­forces the data pro­tec­tion law, with 48 en­force­ment de­cisions re­cor­ded by com­ment­ary for 2025.
  • The Coun­cil of Europe treaty would reach pub­lic au­thor­it­ies and their con­tract­ors, but no Gulf state has signed it and it has not entered into force.
AI governanceRules of the Game35 •

Humain, and Saudi industrial policy for AI

Saudi Arabia's champion is barely documented in the public record, and the rules it would answer to are advisory by design.

The re­search be­hind this book could not es­tab­lish the basic facts about Hu­main. No primary source opened for the Gulf fact base re­cords its own­er­ship, its gov­ernance or any com­mit­ment it has made. The in­vest­ment totals in cir­cu­la­tion, re­por­ted as USD 100 bil­lion and 2.2 gigawatts, were traced only to low-qual­ity sec­ond­ary sources and set aside. That ab­sence is the find­ing. A buyer can­not con­tract against a bal­ance sheet it can­not source.

What can be sourced is the bloc's own ac­count of scale. The Sec­ret­ary Gen­eral of the Gulf Co­oper­a­tion Coun­cil stated on 7 Oc­to­ber 2025 that mem­ber states' in­vest­ments in AI have reached tens of bil­lions of dol­lars, with plans to raise them to hun­dreds of bil­lions by 2030. That is the fig­ure to use. It fixes the order of mag­nitude without pre­tend­ing to a com­pany-level num­ber that no primary source sup­ports.

Against that scale, the rules are soft. The Saudi Data and Ar­ti­fi­cial In­tel­li­gence Au­thor­ity has is­sued five in­stru­ments since Septem­ber 2023, be­gin­ning with AI Eth­ics Prin­ciples ver­sion 1.0 and end­ing with a na­tional AI risk man­age­ment frame­work launched on 14 July 2026. Latham & Watkins re­cords that the Prin­ciples are not leg­ally bind­ing. The Saudi Press Agency an­nounce­ment of the risk frame­work uses the lan­guage of en­able­ment and nowhere states that the frame­work is man­dat­ory.

The Saudi AI authority issued five instruments in under three years and none of them is a law.
instruments issued
00.511.5212023220240202522026
Source: Saudi Press Agency and Digital Policy Alert, 2023 to 2026

One in­stru­ment strains that de­scrip­tion. The au­thor­ity's deep­fakes guidelines of 8 May 2026 are re­por­ted by the Saudi Press Agency to man­date that de­velopers and cre­at­ors ad­here to

AI governanceRules of the Game36 •

strict eth­ical and reg­u­lat­ory stand­ards, in­clud­ing wa­ter­mark­ing and source veri­fic­a­tion. No en­force­ment mech­an­ism is named in the an­nounce­ment. A duty with no ad­dressee and no pen­alty re­mains a state­ment of ex­pect­a­tion, whatever verb the press re­lease reaches for.

A duty with no addressee and no penalty remains a statement of expectation, whatever verb the press release reaches for.

The bind­ing layer is the data law. The Per­sonal Data Pro­tec­tion Law entered into force on 14 Septem­ber 2023 and be­came fully en­force­able a year later, with the AI au­thor­ity as com­pet­ent reg­u­lator. Com­ment­ary re­cords 48 en­force­ment de­cisions by its com­mit­tees dur­ing 2025. A Saudi de­veloper train­ing on per­sonal data an­swers to a live su­per­visor, though for data hand­ling rather than for model be­ha­viour.

Sec­tor rules reach fur­ther. The Saudi Cent­ral Bank rule­book provides that where ma­chine learn­ing or ar­ti­fi­cial in­tel­li­gence is used in fraud risk, the sys­tem should not be a black box and should be cap­able of being audited. The Cap­ital Mar­ket Au­thor­ity con­sul­ted on robo-ad­vis­ory from Au­gust 2025 and ap­proved the frame­work in March 2026. Robo-ad­vis­ory port­fo­lios had reached SAR 6.4 bil­lion, up 86 per cent on the year, on the reg­u­lator's 2025 an­nual re­port.

The med­ical device pic­ture is weaker still. The Food and Drug Au­thor­ity's guid­ance on re­view­ing AI and big data based med­ical devices, as pos­ted to the pub­lic con­sulta­tion plat­form, car­ries the doc­u­ment num­ber MDS-G00, ver­sion num­ber 0.0 and a ver­sion date of 00/00/2020. On its face it is an un­dated draft. A cham­pion selling clin­ical soft­ware in Saudi Ar­a­bia would be as­sessed against a doc­u­ment that has never been fi­nal­ised.

So the an­swer on ap­plic­able rules is short and un­sat­is­fy­ing. A Saudi state-backed de­veloper an­swers to the data pro­tec­tion law and to whichever sec­tor rule­book li­censes its cus­tomer, be­neath an ad­vis­ory na­tional stack writ­ten by the same au­thor­ity that en­forces the data law. Own­er­ship changes none of it. No loc­ated in­stru­ment at­taches a single ad­di­tional ob­lig­a­tion to a model the state hap­pens to own.

Key takeaways
  • Hu­main's own­er­ship, gov­ernance and com­mit­ments are not doc­u­mented in any primary source opened, and the widely quoted cap­ital and power fig­ures were re­jec­ted as un­sourced.
AI governanceRules of the Game37 •
  • Five AI in­stru­ments is­sued by the Saudi AI au­thor­ity since Septem­ber 2023 are ad­vis­ory, and the bind­ing ob­lig­a­tions come from the 2023 data pro­tec­tion law in­stead.
  • Sec­tor rule­books bite harder than na­tional ones, with the cent­ral bank re­quir­ing that an AI fraud sys­tem be audit­able rather than a black box.

QAI, and Qatar's bet on sovereign compute

No Qatari state-owned developer appears in the fact base, so the question becomes which Qatari rule would reach one at all.

Qatar's cham­pion can­not be de­scribed from the re­cord as­sembled for this book. Noth­ing in the Gulf fact base names QAI, or any other Qatari state-owned model de­veloper, or any com­mit­ment made by one. The en­tity is simply ab­sent. What fol­lows is the rule set that would apply to such a de­veloper, and it is nar­row and sits al­most wholly in fin­ance.

One Qatari AI in­stru­ment plaus­ibly binds. The Qatar Cent­ral Bank is­sued its Ar­ti­fi­cial In­tel­li­gence Guideline on 4 Septem­ber 2024, and it ap­plies to li­censed en­tit­ies that de­velop, pro­cure, out­source, op­er­ate or provide AI sys­tems. Its bind­ing status rests on com­ment­ary. The bank's own text, re­por­ted to be pub­lished as Annex 160 to the In­struc­tions to Banks 2024, could not be reached dur­ing the re­search pass.

The re­por­ted ob­lig­a­tions mat­ter more to the buyer than to the builder. Com­ment­ary re­cords a re­quire­ment of prior ap­proval from the cent­ral bank be­fore a li­censed en­tity signs a high-risk AI pur­chase, li­cens­ing or out­sourcing agree­ment, and be­fore it launches or ma­ter­i­ally mod­i­fies an AI sys­tem as a pro­vider. A state-owned vendor is not the ad­dressee. Its cus­tomer is, and that is the prac­tical con­straint on any sale into a Qatari bank.

Pen­al­ties fol­low the gen­eral su­per­vis­ory route. Com­ment­ary de­scribes the po­s­i­tion as sub­ject to the cent­ral bank's or­din­ary en­force­ment powers rather than to any ded­ic­ated AI pen­alty re­gime. There is no AI fine to quote. The on­shore data pro­tec­tion law, Law No. 13 of 2016, car­ries a re­por­ted max­imum pen­alty of QAR 5 mil­lion, and it is the only Qatari money fig­ure the fact base sup­ports.

Everything above it ad­vises. The Na­tional Ar­ti­fi­cial In­tel­li­gence Strategy of­fers re­com­mend­a­tions rather than bind­ing rules, and sources dis­agree on whether it dates from 2019 or early

AI governanceRules of the Game38 •
Qatar has one AI instrument that plausibly binds and five that advise or await adoption.
instruments located
00.81.52.231Binding2Advisory3Draft
Source: Digital Policy Alert, regulations.ai and Eversheds Sutherland, 2024 to 2026

2022. Min­istry guid­ance on AI eth­ics is de­scribed as vol­un­tary. A draft Na­tional AI Policy ex­ists and is ex­pressly a con­sulta­tion draft, with Ever­sheds Suth­er­land pla­cing its fi­nal­isa­tion some time after Novem­ber 2025.

A state-owned vendor is not the addressee of Qatar's one binding AI rule. Its customer is.

One re­por­ted fea­ture of that draft goes straight to the gov­ernance ques­tion. The draft is said to ex­clude pub­lic bod­ies that already re­quire spe­cial­ised ar­range­ments, nam­ing the Na­tional Cyber Se­cur­ity Agency and the Qatar Cent­ral Bank. The claim comes from a search ex­tract. It was not traced to a primary source and can­not be re­lied upon. If it sur­vives draft­ing, Qatar's ho­ri­zontal policy will carve parts of the state out of scope at the mo­ment of ad­op­tion.

The su­per­vis­ory map is un­settled as well. Sources dis­agree on which body ad­min­is­ters the on­shore data pro­tec­tion law, one nam­ing a di­vi­sion of the Na­tional Cyber Se­cur­ity Agency and an­other a na­tional data pro­tec­tion of­fice under the min­istry. Neither ac­count reaches a primary source. A buyer can­not name its coun­ter­party's su­per­visor from the pub­lic re­cord, which is an odd sen­tence to have to write about a reg­u­lated mar­ket.

The in­sti­tu­tional build-out is still under way. A draft Cab­inet de­cision ap­proved in Septem­ber 2025 would es­tab­lish a Na­tional Cen­ter for Ar­ti­fi­cial In­tel­li­gence under the min­istry. Until that body ex­ists, Qatari AI gov­ernance runs through a fin­an­cial reg­u­lator and a data law of 2016. Com­pute policy sits out­side both of them, which is where the gov­ernance ques­tion for a sov­er­eign com­pute pro­gramme ac­tu­ally lives.

AI governanceRules of the Game39 •
Key takeaways
  • No Qatari state-owned model de­veloper is named in any of the four fact bases, so noth­ing here de­scribes QAI as an en­tity or its com­mit­ments.
  • The 2024 cent­ral bank guideline is the only Qatari AI rule that plaus­ibly binds, and it ad­dresses li­censed fin­an­cial in­sti­tu­tions rather than vendors.
  • Qatar's draft na­tional policy is re­por­ted to ex­clude the cent­ral bank and the cyber se­cur­ity agency, which would place parts of the state out­side scope.

G42, MGX, and the UAE's two-track position

Neither G42 nor MGX appears in the fact base, so the only documented constraint on a two-track position is American procurement.

The re­cord as­sembled for this book con­tains noth­ing on G42 or MGX. Neither en­tity is named in the Gulf, United States, India or global fact bases, and the Star­gate UAE ca­pa­city fig­ures in cir­cu­la­tion were traced to low-qual­ity sec­ond­ary sources and ex­cluded. The align­ment story can­not be evid­enced here. What can be evid­enced is the mech­an­ism through which Amer­ican align­ment is en­forced, and that mech­an­ism is pro­cure­ment.

The United States con­di­tions fed­eral pur­chase on prop­er­ties of the model it­self. Ex­ec­ut­ive Order 14319 of 23 July 2025 per­mits agency heads to pro­cure only large lan­guage mod­els de­veloped in ac­cord­ance with two prin­ciples, truth-seek­ing and ideo­lo­gical neut­ral­ity. The im­ple­ment­ing guid­ance fol­lowed. OMB Memor­andum M-26-04 is dated 11 Decem­ber 2025, and agen­cies must write com­pli­ance terms into any so­li­cit­a­tion or order for a large lan­guage model is­sued after that date.

The dis­clos­ure annex is the part that bites a for­eign cham­pion. Where an agency needs more, M-26-04 reaches any train­ing or de­vel­op­ment that oc­curred out­side the United States, in­clud­ing the type of activ­ity and the coun­try in which it happened. It also reaches any modi­fic­a­tions made to com­ply with a reg­u­la­tion from a gov­ern­ment other than the Amer­ican fed­eral gov­ern­ment. This is for­eign-in­flu­ence dis­clos­ure by con­tract. It asks a state-backed vendor to de­scribe its own home ob­lig­a­tions.

The bind­ing clause is milder. On 28 Septem­ber 2026 the Gen­eral Ser­vices Ad­min­is­tra­tion is­sued GSAR 552.239-7001 as a reg­u­la­tion de­vi­ation, ef­fect­ive 19 Oc­to­ber 2026, and

AI governanceRules of the Game40 •
American procurement asks a state-backed vendor to disclose the modifications it made to satisfy another government's rules.

com­ment­ary from Crow­ell and Mor­ing re­cords that it re­placed the de­tailed un­biased-AI ob­lig­a­tions with a single reas­on­able-ef­forts stand­ard. De­com­mis­sion­ing li­ab­il­ity is capped at 25 per cent of the af­fected order value. The lever loses force at each han­dover from order to memor­andum to clause.

Re­voc­ab­il­ity is the sharper con­straint. On 27 Feb­ru­ary 2026 the Gen­eral Ser­vices Ad­min­is­tra­tion an­nounced that it was re­mov­ing An­thropic from USAi.gov and from the Mul­tiple Award Sched­ule, in sup­port of a pres­id­en­tial dir­ect­ive to cease all use of that com­pany's tech­no­logy. No stat­ute and no rule­mak­ing were in­volved. Ac­cess to a vehicle sit­ting in­side more than USD 116 bil­lion of fed­eral con­tracts was with­drawn by dir­ec­tion, which is the pre­ced­ent any for­eign cham­pion should price.

In the UAE the binding AI rules sit in a free zone and in health regulators, not at federal level.
instruments located
00.81.52.230Federal, private1Free zone3Health sector2Advisory
Source: DIFC, Abu Dhabi DoH, Dubai DHA and UAE AI Office, 2018 to 2026

The do­mestic track is thin­ner. There is no fed­eral AI frame­work law, and Fed­eral De­cree-Law No. 45 of 2021 on per­sonal data re­mains in force without its Ex­ec­ut­ive Reg­u­la­tions, which Trow­ers & Ham­lins de­scribed in June 2026 as leav­ing the law un­en­force­able. Breach no­ti­fic­a­tion, trans­fer con­trols and of­ficer thresholds are un­fixed. A vendor's home ob­lig­a­tions are hard to dis­close when they have not yet been writ­ten down.

One Emir­ati in­stru­ment is draf­ted for autonomy rather than for auto­ma­tion. Reg­u­la­tion 10 of the Dubai In­ter­na­tional Fin­an­cial Centre as­signs con­trol­ler-style re­spons­ib­il­ity to the

AI governanceRules of the Game41 •

De­ployer that au­thor­ises or be­ne­fits from a sys­tem, and pro­cessor-style re­spons­ib­il­ity to the Op­er­ator. The centre's own guid­ance com­pares the sys­tem's po­s­i­tion to that of an em­ployee. Its high-risk gate is shut. No sys­tem may be used for high-risk pro­cessing until the Com­mis­sioner pro­mul­gates cer­ti­fic­a­tion re­quire­ments, which were still at con­sulta­tion in mid-2026, roughly 33 months after en­act­ment.

The two-track po­s­i­tion is gov­erned asym­met­ric­ally. The Amer­ican side is writ­ten down, con­di­tioned on prop­er­ties of the model and with­draw­able at will. The Emir­ati side con­sists of an Au­thor­ity ap­proved on 14 June 2026 and a free-zone rule whose ma­chinery re­mains un­built. A buyer weigh­ing an Emir­ati cham­pion should read the Amer­ican clause first, be­cause it is the only text in this chapter with a work­ing en­force­ment his­tory.

Key takeaways
  • G42 and MGX ap­pear in none of the four fact bases, so the UAE's align­ment with Amer­ican cap­ital can­not be de­scribed from this re­search at all.
  • Amer­ican pro­cure­ment con­di­tions pur­chase on model prop­er­ties and re­quires dis­clos­ure of de­vel­op­ment con­duc­ted abroad and of changes made for for­eign reg­u­lat­ors.
  • The An­thropic del­ist­ing of 27 Feb­ru­ary 2026 shows ex­clu­sion achieved by dir­ect­ive, without stat­ute or rule­mak­ing, from a vehicle cov­er­ing USD 116 bil­lion of con­tracts.

India: the AI Mission, and the sovereign model question

India funds AI capability generously and attaches no obligation to the funding, which is very nearly the whole answer.

India has a mis­sion rather than a cham­pion, at least in the re­cord as­sembled here. The India fact base names no state-owned or state-fun­ded model. That is the first find­ing. The sov­er­eign model ques­tion can­not be answered on this evid­ence, and what the base does doc­u­ment is the money and the ab­sence of any con­di­tion at­tached to it. The IndiaAI Mis­sion was ap­proved by the Union Cab­inet on 7 March 2024 with an out­lay of Rs 10,371.92 crore over five years.

The Mis­sion funds cap­ab­il­ity. It is im­ple­men­ted by the IndiaAI In­de­pend­ent Busi­ness Di­vi­sion under the Di­gital India Cor­por­a­tion and im­poses no ob­lig­a­tion on any private party. Its Safe

AI governanceRules of the Game42 •

and Trus­ted AI pil­lar funds work rather than reg­u­lat­ing it. No in­stru­ment loc­ated at­taches a safety or audit con­di­tion to a model built with that pub­lic money.

Sovereignty is a funding question in India, and not yet a regulatory one.

The ob­lig­a­tions that exist are gen­eral. The Di­gital Per­sonal Data Pro­tec­tion Rules 2025 were no­ti­fied on 13 Novem­ber 2025, with con­sent man­ager re­gis­tra­tion com­men­cing on 13 Novem­ber 2026 and the sub­stant­ive du­ties on 13 May 2027. Rule 13(3) is the pro­vi­sion that reaches a model dir­ectly. A Sig­ni­fic­ant Data Fi­du­ciary must verify that al­gorithmic soft­ware it ad­opts is not likely to pose a risk to the rights of data prin­cipals.

One In­dian rule does reach a model builder dir­ectly. The amended In­form­a­tion Tech­no­logy Rules, in force from 20 Feb­ru­ary 2026, re­quire an in­ter­me­di­ary of­fer­ing a tool that can gen­er­ate syn­thetic audio or visual con­tent to label the out­put prom­in­ently and to embed per­man­ent proven­ance metadata car­ry­ing a unique iden­ti­fier. Text is out­side the defin­i­tion. A state-fun­ded In­dian image or speech model would carry that duty on the same terms as a com­mer­cial one.

India's data protection penalties are large and say nothing about artificial intelligence.
Rs crore, maximum penalty
062.5125187.5250250Security lapse200Breach notice200Children's data50Other breach
Source: Press Information Bureau backgrounder, 2025

Pen­al­ties under the Act are large and gen­eric. Fail­ure to main­tain reas­on­able se­cur­ity safe­guards car­ries up to Rs 250 crore, fail­ures on breach no­ti­fic­a­tion and on chil­dren's data up to Rs 200 crore each, and any other breach up to Rs 50 crore. None of them is AI-spe­cific. A state-fun­ded de­veloper is ex­posed on ex­actly the terms a private one is, which is con­sist­ent and still not a gov­ernance an­swer.

AI governanceRules of the Game43 •

The en­force­ment body is not yet staffed. The min­istry ad­vert­ised one chair­per­son post and four mem­ber posts for the Data Pro­tec­tion Board on 6 May 2026, and a check on 21 Septem­ber 2026 found no ap­point­ment no­ti­fied. The Board ex­ists in law alone. The sub­stant­ive du­ties it would su­per­vise take ef­fect on 13 May 2027, which leaves a short win­dow in which to con­sti­tute an ad­ju­dic­ator.

The ho­ri­zontal choice is on re­cord. The min­istry un­veiled non-bind­ing India AI Gov­ernance Guidelines on 5 Novem­ber 2025, and its Sec­ret­ary stated that the focus re­mains on using ex­ist­ing le­gis­la­tion wherever pos­sible. The gov­ernance group and ex­pert com­mit­tee those guidelines pro­pose are not re­por­ted as con­sti­tuted. A con­sulta­tion paper on reg­u­la­tion was an­nounced on 8 Oc­to­ber 2026 with a one-month ho­ri­zon, and noth­ing loc­ated shows that it has is­sued.

So the hon­est find­ing for India is a neg­at­ive one. Pub­lic money is spe­cific, year-stamped and trace­able to a Cab­inet de­cision of March 2024. The du­ties that would gov­ern what the money builds are ho­ri­zontal, and they draw no dis­tinc­tion between a state-fun­ded model and any other. The chapter is short for a reason.

Key takeaways
  • The IndiaAI Mis­sion car­ries an out­lay of Rs 10,371.92 crore over five years and im­poses no ob­lig­a­tion on any private party that re­ceives or uses it.
  • No In­dian in­stru­ment loc­ated at­taches a safety, dis­clos­ure or audit con­di­tion to a pub­licly fun­ded model, so the du­ties that apply are the ho­ri­zontal ones.
  • The Data Pro­tec­tion Board had no chair­per­son or mem­ber ap­poin­ted as at 21 Septem­ber 2026, against sub­stant­ive du­ties com­men­cing on 13 May 2027.

Who regulates a national champion?

In every Gulf state the body that writes the AI rule also sponsors the programme, which leaves the contract doing the work.

No na­tional reg­u­lator in the Gulf an­swers for a state-owned model, on the evid­ence as­sembled here. Re­search to 6 Novem­ber 2026 loc­ated no bind­ing AI in­stru­ment at fed­eral or na­tional level reach­ing the private sec­tor in any of the six mem­ber states. One

AI governanceRules of the Game44 •

free-zone rule was found. Three sec­tor-reg­u­lator in­stru­ments were found along­side it, two fur­ther policies bind gov­ern­ment en­tit­ies only, and at least eight na­tional frame­works are ad­vis­ory.

Every binding Gulf AI instrument located sits below the national level.
instruments located
024680National1Free zone3Sector regulator2Government only8Advisory
Source: DIFC, QCB, Abu Dhabi DoH, Bahrain iGA and SDAIA instruments, 2018 to 2026

The con­flict is struc­tural. The Saudi Data and Ar­ti­fi­cial In­tel­li­gence Au­thor­ity writes the na­tional AI frame­works and ad­min­is­ters en­force­ment of the data pro­tec­tion law. The new Emir­ati Au­thor­ity pro­poses le­gis­la­tion and sets stand­ards, and it car­ries a duty of en­sur­ing com­pli­ance across fed­eral en­tit­ies. Oman's min­istry runs the na­tional AI pro­gramme and also holds com­pli­ance and over­sight for the policy gov­ern­ing it.

The bind­ing rules ad­dress the buyer. The Qatar Cent­ral Bank guideline is re­por­ted to re­quire prior ap­proval be­fore a li­censed en­tity signs a high-risk AI pur­chase or out­sourcing agree­ment. Reg­u­la­tion 10 of the Dubai In­ter­na­tional Fin­an­cial Centre places con­trol­ler-style li­ab­il­ity on the De­ployer, with guid­ance com­par­ing the sys­tem's po­s­i­tion to that of an em­ployee. Bahrain's policy at­trib­utes neg­at­ive ef­fects to re­spons­ible in­di­vidu­als in­side the gov­ern­ment en­tity rather than to the ma­chine.

That asym­metry is not an ac­ci­dent of draft­ing. A re­gime built on data pro­tec­tion con­cepts reaches who­ever pro­cesses per­sonal data, and a state-owned de­veloper selling a model is fre­quently not that per­son. The ob­lig­a­tion lands down­stream. A cham­pion can be com­mer­cially dom­in­ant and reg­u­lat­or­ily peri­pheral at the same time, and in the Gulf of 2026 sev­eral of them are.

A champion can be commercially dominant and regulatorily peripheral at the same time.
AI governanceRules of the Game45 •

So the buyer car­ries the gov­ernance that the state has not writ­ten. The first ques­tion is which in­stru­ment binds the vendor, named and dated, as dis­tinct from the na­tional strategy the vendor will cite. The second is whether the vendor's su­per­visor is a dif­fer­ent body from its spon­sor. Both ques­tions are fac­tual. Six strategies were loc­ated across the Gulf and not one of them an­swers either.

Three con­tract ques­tions fol­low. Ask who is De­ployer and who is Op­er­ator, be­cause Reg­u­la­tion 10 al­loc­ates li­ab­il­ity on that dis­tinc­tion alone. Ask whether high-risk pro­cessing is law­ful at all, given that the Dubai gate stays shut until cer­ti­fic­a­tion re­quire­ments are pro­mul­gated. Ask what hap­pens when the model changes, which the Saudi med­ical device draft an­swers in four tiers and the In­dian device guid­ance leaves to dis­cre­tion.

Dis­clos­ure terms are avail­able off the shelf. OMB Memor­andum M-26-04 reaches train­ing or de­vel­op­ment con­duc­ted out­side the United States and modi­fic­a­tions made to sat­isfy a for­eign gov­ern­ment's reg­u­la­tion, which is a us­able tem­plate for any­one buy­ing a state-backed model. The European Com­mis­sion's model con­trac­tual AI clauses were up­dated on 5 March 2025. The United King­dom's Pro­cure­ment Policy Note 02/24 sup­plies op­tional ques­tions for identi­fy­ing AI use in a sup­ply chain.

Two gaps de­serve stat­ing plainly. The fact bases name none of the Gulf cham­pi­ons, so noth­ing in this part de­scribes Hu­main, QAI, G42 or MGX as en­tit­ies, and noth­ing here re­ports their com­mit­ments. No Gulf in­stru­ment loc­ated uses the word agent or agen­tic in an op­er­at­ive pro­vi­sion. Both ab­sences are find­ings a buyer can act on, be­cause they de­scribe pre­cisely what the con­tract has to sup­ply.

Key takeaways
  • No bind­ing Gulf AI in­stru­ment reach­ing the private sec­tor sits at na­tional level, so no na­tional reg­u­lator an­swers for a state-owned model today.
  • The rules that do bind ad­dress the buyer rather than the vendor, which places the gov­ernance bur­den on pro­cure­ment and on the con­tract it­self.
  • Ask which named in­stru­ment binds the vendor, whether its su­per­visor dif­fers from its spon­sor, and who holds li­ab­il­ity when the model changes.
AI governanceRules of the Game46 •

The part in brief

  1. 01In the Gulf no national regulator answers for a state-owned model: research located six AI strategies and no national framework law reaching the private sector.
  2. 02Four Gulf champions cannot be described from this research, because no primary source opened names Humain's ownership, QAI as an entity, or G42 and MGX at all.
  3. 03The conflict is structural, since the Saudi authority that writes the advisory AI frameworks also enforces the data protection law, with 48 decisions recorded for 2025.
  4. 04American procurement is the only lever with an enforcement history, requiring disclosure of foreign development and, on 27 February 2026, removing a named vendor by directive alone.
  5. 05India funds capability without conditions, committing Rs 10,371.92 crore over five years while attaching no safety or audit duty to a publicly funded model.
AI governanceRules of the Game47 •
The national champions

American procurement asks a state-backed vendor to disclose the modifications it made to satisfy another government's rules.

04

What the mandates make you do

Inventory, evidence, oversight, disclosure, procurement.
  • Inventory and classification: knowing what you run
  • Documentation and evidence
  • Human oversight, and what it costs
  • Disclosure, labelling and provenance
  • Procurement: governing what you did not build
  • Buying from a sovereign champion: the contract questions

Inventory and classification: knowing what you run

Every regime's first duty is the same, because nothing binds until a system has been found and classified.

Every AI rule­book starts in the same place. Be­fore any duty can at­tach, a sys­tem has to be found and sor­ted into a class. The European Union makes clas­si­fic­a­tion the hinge of its whole stat­ute, fix­ing ob­lig­a­tions by ref­er­ence to Annex I and Annex III of Reg­u­la­tion (EU) 2024/1689. Art­icle 6(5) re­quired the Com­mis­sion to pub­lish prac­tical clas­si­fic­a­tion guidelines, with worked ex­amples of what is and is not high-risk, no later than 2 Feb­ru­ary 2026. It did not.

Draft guidelines ap­peared on 19 May 2026 in three parts, and a tar­geted con­sulta­tion closed on 23 July 2026. The Com­mis­sion's policy page, last up­dated 19 Oc­to­ber 2026, still re­cords that feed­back will be in­cor­por­ated be­fore ad­op­tion. So the test by which a European firm de­cides whether it runs a high-risk sys­tem is it­self un­adop­ted. The du­ties have moved too. Reg­u­la­tion (EU) 2026/1744 of 8 July 2026 pushed the Annex III ob­lig­a­tions to 2 Decem­ber 2027 and the Annex I ob­lig­a­tions to 2 Au­gust 2028, and re­cital (40) at­trib­utes the move to delayed stand­ards and delayed na­tional com­pet­ent au­thor­it­ies.

Other re­gimes ask for the re­gister first and argue about classes later. OMB Memor­andum M-25-21 of 3 April 2025 re­quires every United States fed­eral agency to pub­lish an AI use case in­vent­ory an­nu­ally. It defines high-im­pact AI by the con­sequence of its out­put, and lists fif­teen cat­egor­ies pre­sumed to fall in­side that defin­i­tion. The Qatar Cent­ral Bank's Ar­ti­fi­cial In­tel­li­gence Guideline of 4 Septem­ber 2024 re­quires li­censed en­tit­ies to main­tain an up­dated AI sys­tem re­gister and to clas­sify high-risk sys­tems by risk as­sess­ment. The pat­tern re­peats. DIFC Reg­u­la­tion 10, in force since Septem­ber 2023, calls its own re­gister a basic map for warp speed pro­cessing.

The most use­ful clas­si­fic­a­tion test pub­lished any­where is two days older than this edi­tion's cut-off. OCC Bul­letin 2026-54 of 4 Novem­ber 2026 sets out four il­lus­trat­ive tiers for Amer­ican banks, and says a sys­tem's tier turns on what it could cause and how much agency has been trans­ferred to it, rather than on the tech­nical soph­ist­ic­a­tion of the model. An autonom­ous pay­ment agent sits in Tier 1 Crit­ical. De­cision sup­port with a man­dat­ory human in the loop sits in Tier 3. The test is be­ha­vi­oural, not ar­chi­tec­tural.

What a us­able in­vent­ory con­tains is there­fore settled by con­ver­gence rather than by any single in­stru­ment. Col­or­ado's in­sur­ance reg­u­la­tion 10-1-1, amended with ef­fect from 15

AI governanceRules of the Game48 •
Each instrument names its own count of governance items, and the counts land close together.
named items in the instrument
051015207OMB practices14Colorado reg15OMB categories16OCC framework
Source: OMB, OCC and Colorado Division of Insurance, 2025 and 2026
The test by which a European firm decides whether it runs a high-risk system is itself still unadopted.

Oc­to­ber 2025, re­quires a ver­sion-con­trolled model in­vent­ory in­side a four­teen-com­pon­ent gov­ernance frame­work. The New York De­part­ment of Fin­an­cial Ser­vices, in In­sur­ance Cir­cu­lar Let­ter No. 7 of 11 July 2024, re­quires a model in­vent­ory and an­nual out­put test­ing for drift. China asks the same, as a fil­ing. Art­icle 24 of the Al­gorithmic Re­com­mend­a­tion Pro­vi­sions re­quires the pro­vider's name, ser­vice form, ap­plic­a­tion field, al­gorithm type and self-as­sess­ment re­port within ten work­ing days of launch.

What a usable inventory carries
  • Owner, pur­pose and the de­cision the out­put feeds, be­cause every clas­si­fic­a­tion test in the European, Amer­ican and Gulf in­stru­ments turns on con­sequence rather than on tech­nique.
  • Train­ing and eval­u­ation proven­ance, in­clud­ing whether a third-party model sits un­der­neath, which both the Chinese fil­ing re­gime and the GSA con­tract clause re­quire to be named.
  • The tier as­signed, the test ap­plied and the date it was ap­plied, so that a clas­si­fic­a­tion can be de­fen­ded to a su­per­visor two years after it was made.
  • Human over­sight ac­tu­ally in place, named by per­son or by role, since OMB M-25-21 treats over­sight and a fail-safe as one of seven min­imum prac­tices.

One in­stru­ment that used to catch AI mod­els in banks no longer does. SR 11-7 of 4 April 2011 re­quired a main­tained model in­vent­ory and in­de­pend­ent val­id­a­tion on con­cep­tual sound­ness. It was su­per­seded on 17 April 2026 by SR 26-2, is­sued along­side OCC Bul­letin 2026-13.

AI governanceRules of the Game49 •

The re­place­ment is nar­rower. That bul­letin states that gen­er­at­ive and agen­tic AI mod­els are novel and rap­idly evolving and are there­fore not within its scope. A firm that built its in­vent­ory on model risk man­age­ment now finds its new­est sys­tems out­side the thing it built.

Documentation and evidence

A regulator does not ask what you believe about your system. It asks for the file you kept while building it.

Reg­u­lat­ors ask for files, not as­sur­ances. The European tech­nical doc­u­ment­a­tion du­ties and the In­dian im­pact as­sess­ment and audit du­ties are writ­ten in dif­fer­ent legal tra­di­tions and ar­rive at nearly the same list. Both want a re­cord made at the time, by a named per­son, that would let some­body else re­con­struct the de­cision. Neither ac­cepts a sum­mary writ­ten af­ter­wards. The real dif­fer­ence is who holds the file and who may call for it.

In Europe the doc­u­ment­a­tion duty is the thing the de­fer­ral did not touch. EN 18286:2026, on qual­ity man­age­ment sys­tems for EU AI Act reg­u­lat­ory pur­poses, was cited in the Of­fi­cial Journal by Com­mis­sion Im­ple­ment­ing De­cision (EU) 2026/2274 of 12 Oc­to­ber 2026, pub­lished on 14 Oc­to­ber 2026. CEN-CENELEC de­scribes it as the first har­mon­ised stand­ard sup­port­ing the Act, and says it turns the qual­ity man­age­ment re­quire­ments of Art­icle 17 into prac­tical, audit­able re­quire­ments for pro­viders of high-risk sys­tems. It is also the only one. The Annex to that im­ple­ment­ing de­cision lists a single entry.

The legal con­sequence of the cita­tion is the reason to read the stand­ard now. Re­cital (1) of the im­ple­ment­ing de­cision re­cords that sys­tems con­form­ing to a har­mon­ised stand­ard whose ref­er­ence has been pub­lished in the Of­fi­cial Journal are pre­sumed to con­form with the re­quire­ments that stand­ard cov­ers. The pre­sump­tion runs from pub­lic­a­tion. So one doc­u­ment, cited in Oc­to­ber 2026, is the whole of the audit­able evid­ence route into a stat­ute whose high-risk du­ties were put back six­teen months on the ground that stand­ards were not ready.

One standard, cited in October 2026, is the whole auditable evidence route into the European statute.
AI governanceRules of the Game50 •

India states the evid­ence duty as a ca­dence rather than as a stand­ard. Rule 13 of the Di­gital Per­sonal Data Pro­tec­tion Rules 2025, no­ti­fied as G.S.R. 846(E) on 13 Novem­ber 2025, re­quires a Sig­ni­fic­ant Data Fi­du­ciary to un­der­take a Data Pro­tec­tion Im­pact As­sess­ment and an audit once every twelve months from des­ig­na­tion, and to cause the as­sessor or aud­itor to fur­nish a re­port of sig­ni­fic­ant ob­ser­va­tions to the Data Pro­tec­tion Board. Rule 13(3) reaches AI dir­ectly. It re­quires due di­li­gence to verify that tech­nical meas­ures in­clud­ing al­gorithmic soft­ware are not likely to pose a risk to the rights of Data Prin­cipals.

Two dates gov­ern any­one build­ing that file. Rule 4, on the re­gis­tra­tion and ob­lig­a­tions of Con­sent Man­agers, com­mences one year after pub­lic­a­tion, on 13 Novem­ber 2026. The sub­stant­ive rules, being Rules 3, 5 to 16, 22 and 23, com­mence eight­een months after pub­lic­a­tion, on 13 May 2027. Three years and nine months will then have passed since the Act was en­acted on 11 Au­gust 2023. The Board that re­ceives those audit re­ports was con­sti­tuted in law on 13 Novem­ber 2025, had its Chair­per­son and four Mem­ber posts ad­vert­ised by MeitY cir­cu­lar of 6 May 2026, and had no ap­pointee re­cor­ded as at 21 Septem­ber 2026.

India prices the evidence duty through the penalty tiers in its Act.
Rs crore, maximum penalty
062.5125187.5250250Security lapse200Breach notice200Children's data50Other breach
Source: Press Information Bureau, 2025

The pen­al­ties make the file worth keep­ing. Fail­ure to main­tain reas­on­able se­cur­ity safe­guards at­tracts up to Rs 250 crore under the Act. Breach no­ti­fic­a­tion fail­ures and chil­dren's data fail­ures at­tract up to Rs 200 crore each, and any other breach by a Data Fi­du­ciary up to Rs 50 crore. Ap­peals from the Board lie to the Tele­com Dis­putes Set­tle­ment and Ap­pel­late Tribunal. A firm with no dated im­pact as­sess­ment and no audit trail has noth­ing to put in front of either body.

The European Union frames over­sight by autonomy rather than by ac­tion. Reg­u­la­tion (EU) 2024/1689 con­tains no in­stance of the word agent, and uses autonomy in­stead, el­even times. Art­icle 3(1) defines an AI sys­tem as one de­signed to op­er­ate with vary­ing levels of

AI governanceRules of the Game51 •

autonomy. The over­sight pro­vi­sions re­quire meas­ures com­men­sur­ate with the risks, level of autonomy and con­text of use, and the pen­alty cri­teria in­clude the ex­tent to which a sys­tem acts autonom­ously and whether a human could over­ride its de­cision. The fram­ing is de­lib­er­ate. Those du­ties sit in Chapter III, which does not apply until 2 Decem­ber 2027.

Human oversight, and what it costs

Human oversight appears in every instrument and means something different in each, from a job title to approval of a single action.

Human over­sight is the one clause every rule­book con­tains, and no two mean the same thing by it. At one end sits a named of­ficer. At the other sits ap­proval be­fore a single ac­tion may pro­ceed. The cost of com­pli­ance runs on that dif­fer­ence, be­cause a firm can ap­point an of­ficer in a week and can­not staff a per-ac­tion gate at all without re­design­ing the work­flow the sys­tem was bought to re­move.

The DIFC turned over­sight into a post. Reg­u­la­tion 10 on the Pro­cessing of Per­sonal Data through Autonom­ous and Semi-Autonom­ous Sys­tems, en­acted with ef­fect from 1 Septem­ber 2023, re­quires a De­ployer or Op­er­ator en­gaged in high-risk pro­cessing to ap­point an Autonom­ous Sys­tems Of­ficer with com­pet­en­cies sub­stan­tially sim­ilar to those of a Data Pro­tec­tion Of­ficer. The Com­mis­sioner's own guid­ance states the em­ploy­ment ana­logy plainly. A sys­tem op­er­at­ing under the au­thor­ity and for the be­ne­fit of its De­ployer, it says, sits in a po­s­i­tion sub­stan­tially sim­ilar to that of an em­ployee in­side the De­ployer's or­gan­isa­tion.

A firm can appoint an officer in a week. It cannot staff a per-action gate without redesigning the workflow.

The gate is non­ethe­less closed. Reg­u­la­tion 10.3.3 bars com­mer­cial use of a sys­tem for high-risk pro­cessing until audit and cer­ti­fic­a­tion re­quire­ments have been es­tab­lished, and the guid­ance re­cords the reg­u­lat­ory in­tent that no sys­tem may be so used until the Com­mis­sioner has pro­mul­gated them. The two sub-com­mit­tees charged with re­com­mend­ing the cri­teria were to re­port by 30 April 2024. DIFC Con­sulta­tion Paper No. 3 of June 2026, which closed on 18 July 2026, was still pro­pos­ing that ma­chinery, roughly thirty-three months after the dead­line passed.

AI governanceRules of the Game52 •
What the clause asks for, instrument by instrument
  • Under the EU AI Act, over­sight meas­ures com­men­sur­ate with risk, autonomy and con­text of use, with the du­ties de­ferred to 2 Decem­ber 2027 for Annex III sys­tems.
  • Under DIFC Reg­u­la­tion 10, a named Autonom­ous Sys­tems Of­ficer for high-risk pro­cessing, with the cer­ti­fic­a­tion the pro­hib­i­tion de­pends on still un­pub­lished in 2026.
  • Under OMB M-25-21, human over­sight and a fail-safe as one of seven min­imum risk man­age­ment prac­tices for high-im­pact AI across United States fed­eral agen­cies.
  • Under Bahrain's iGA Gen­eral Policy of 20 May 2025, the final de­cision in im­port­ant mat­ters must re­main under human con­trol, bind­ing gov­ern­ment en­tit­ies only.

The closest thing to a per-ac­tion ap­proval ex­pect­a­tion is vol­un­tary and Singa­por­ean. The IMDA Model AI Gov­ernance Frame­work for Agen­tic AI, launched on 22 Janu­ary 2026 and now served as ver­sion 1.5 of 20 May 2026, says or­gan­isa­tions should define check­points or ac­tion bound­ar­ies re­quir­ing human ap­proval, nam­ing ir­re­vers­ible ac­tions such as per­man­ently de­let­ing data or mak­ing pay­ments. It asks for denial by de­fault. The trig­ger is a fail­ure of the ap­proval in­fra­struc­ture, in­clud­ing a su­per­visor who can­not be reached. It also asks firms to audit the over­rid­ing it­self, track­ing the human over­ride rate, be­cause a low rate may sig­nal rub­ber-stamp­ing and a short re­sponse time may sig­nal re­view fa­tigue.

Bind­ing ver­sions of the same idea are nar­row and late. Cali­for­nia SB 947 was chaptered on 30 Septem­ber 2026 and be­comes op­er­at­ive on 1 July 2027, bar­ring an em­ployer from using an auto­mated de­cision sys­tem out­put for dis­cip­line or ter­min­a­tion where that out­put can­not be cor­rob­or­ated. Col­or­ado SB 26-189, signed on 14 May 2026 and in force from 1 Janu­ary 2027, gives a con­sumer the right to re­quest human re­view and re­con­sid­er­a­tion fol­low­ing an ad­verse out­come. One is a duty be­fore the de­cision, con­fined to em­ploy­ment. The other ar­rives after it.

Disclosure, labelling and provenance

Three labelling regimes now bind, and they disagree about what counts as content and about where the mark must go.

Three la­belling re­gimes bind today and they do not agree on the ba­sics. Europe marks at the pro­vider. It la­bels vis­ibly at the de­ployer. India la­bels prom­in­ently and em­beds proven­ance metadata, but only for audio and visual ma­ter­ial. China does both and then puts

AI governanceRules of the Game53 •

a veri­fic­a­tion duty on the plat­form that dis­trib­utes the file. A firm serving all three can­not sat­isfy them with one im­ple­ment­a­tion of one stand­ard.

The European route is Art­icle 50, and it has ap­plied since 2 Au­gust 2026. Art­icle 50(2) re­quires pro­viders of sys­tems gen­er­at­ing syn­thetic audio, image, video or text to mark out­puts in a ma­chine-read­able format de­tect­able as ar­ti­fi­cially gen­er­ated, with tech­nical solu­tions that are ef­fect­ive and in­ter­op­er­able as far as tech­nic­ally feas­ible. Art­icle 50(4) binds de­ploy­ers. They must dis­close deep fakes, and must dis­close AI-gen­er­ated text pub­lished to in­form the pub­lic on mat­ters of pub­lic in­terest un­less a per­son held ed­it­or­ial re­spons­ib­il­ity for it.

One European date is close. New Art­icle 111(4), in­ser­ted by Reg­u­la­tion (EU) 2026/1744, re­quires pro­viders of sys­tems gen­er­at­ing syn­thetic con­tent that were placed on the mar­ket be­fore 2 Au­gust 2026 to take the ne­ces­sary steps to com­ply with Art­icle 50(2) by 2 Decem­ber 2026. Re­cital (38) calls that a trans­itional period of four months. The same day brings the two new Art­icle 5 pro­hib­i­tions. The Com­mis­sion ad­op­ted guidelines on the Art­icle 50 ob­lig­a­tions on 20 July 2026, and about 190 com­pan­ies and or­gan­isa­tions had signed the Code of Prac­tice on Trans­par­ency of AI-gen­er­ated Con­tent by the end of that month.

India's route is nar­rower and much faster. G.S.R. 120(E), dated 10 Feb­ru­ary 2026 and in force from 20 Feb­ru­ary 2026, in­ser­ted Rule 2(1)(wa) into the In­form­a­tion Tech­no­logy Rules 2021, de­fin­ing syn­thet­ic­ally gen­er­ated in­form­a­tion as audio, visual or audio-visual ma­ter­ial cre­ated or altered so as to ap­pear real. Text gen­er­ated by AI falls out­side the defin­i­tion. Rule 3(3)(a)(ii) re­quires all other such ma­ter­ial to be prom­in­ently la­belled in the visual dis­play, or pre­fixed by an audio dis­clos­ure, and em­bed­ded with per­man­ent metadata or other proven­ance mech­an­isms in­clud­ing a unique iden­ti­fier for the com­puter re­source used.

Text generated by AI falls outside the Indian definition, and squarely inside the European one.

The same In­dian in­stru­ment moved the clocks, which is where the com­pli­ance cost ac­tu­ally sits. It cut the take­down period on a court or gov­ern­ment order from thirty-six hours to three, the period for a com­plaint about in­tim­ate im­agery or im­per­son­a­tion from twenty-four hours to two, and the period for a re­moval com­plaint under rule 3(1)(b) from sev­enty-two hours to thirty-six. Griev­ance res­ol­u­tion fell to seven days. Sig­ni­fic­ant so­cial media in­ter­me­di­ar­ies

AI governanceRules of the Game54 •

carry a fur­ther duty under Rule 4(1A) to re­quire a user de­clar­a­tion and then to verify it by tech­nical means.

The Indian takedown clocks as they now stand, after G.S.R. 120(E).
hours allowed after notice
0102030403Court order2Intimate imagery36Removal complaint
Source: MeitY, G.S.R. 120(E), 2026

China's route is the most pre­script­ive and the longest in op­er­a­tion. The Meas­ures for La­belling AI-Gen­er­ated Syn­thetic Con­tent were is­sued on 7 March 2025 and took ef­fect on 1 Septem­ber 2025, along­side the man­dat­ory na­tional stand­ard GB 45438-2025. They re­quire ex­pli­cit la­bels by mod­al­ity and an im­pli­cit label in the file metadata car­ry­ing the con­tent at­trib­ute in­form­a­tion, the pro­vider's name or code and the con­tent num­ber. Art­icle 6 then moves veri­fic­a­tion down­stream. The dis­sem­in­a­tion plat­form must verify that metadata and add a graded prompt, dis­tin­guish­ing con­tent that is gen­er­ated, that may be gen­er­ated and that is merely sus­pec­ted of being gen­er­ated.

Procurement: governing what you did not build

Where legislatures have not acted, buyers have, and the contract clause has become the most reliable instrument of AI regulation.

The sharpest AI rules now in force in the United States are con­tract terms, not stat­utes. One free-stand­ing fed­eral AI stat­ute has been en­acted, the TAKE IT DOWN Act, ap­proved on 19 May 2025 as Pub­lic Law 119-12. Al­most everything else that binds a sup­plier binds it be­cause a buyer wrote it down. The design is de­lib­er­ate. The hook into the vol­un­tary stand­ard was cut in 2025 and re­stored in 2026 by a dif­fer­ent hand.

OMB Memor­andum M-24-18 of 24 Septem­ber 2024 dir­ec­ted fed­eral AI ac­quis­i­tion to the NIST AI Risk Man­age­ment Frame­work by name. M-25-22 of 3 April 2025 res­cin­ded and re­placed it, and con­tains no ref­er­ence to NIST or to the frame­work at all, hook­ing in­stead into

AI governanceRules of the Game55 •

OMB's own M-25-21 for high-im­pact use cases. The frame­work it­self is under re­vi­sion on the in­struc­tion of Amer­ica's AI Ac­tion Plan of July 2025, which asked NIST to strip out ref­er­ences to mis­in­form­a­tion, di­versity and cli­mate change. No draft has been pub­lished.

The Gen­eral Ser­vices Ad­min­is­tra­tion then put it back, by clause. GSAR 552.239-7001, on basic safe­guard­ing of data within large lan­guage model AI sys­tems, was is­sued as a GSA Reg­u­la­tion De­vi­ation on 28 Septem­ber 2026 and took ef­fect on 19 Oc­to­ber 2026. It defines its sup­ply-chain roles by ref­er­ence to the frame­work's Ap­pendix A task cat­egor­ies, and flows down to any sub­con­tractor per­form­ing those tasks that handles gov­ern­ment data. A second route reaches the same place. Sec­tion 552.105 of Texas HB 149, ef­fect­ive 1 Janu­ary 2026, makes sub­stan­tial com­pli­ance with the frame­work's Gen­er­at­ive AI Pro­file a de­fence to a civil pen­alty.

A voluntary framework became binding through a contract clause and a state penalty defence, never through a statute.

What the clause ac­tu­ally asks for is the use­ful part for any buyer. Ap­plic­ab­il­ity turns on a two-part test, re­quir­ing that lan­guage model func­tion­al­ity be a sig­ni­fic­ant fea­ture and that gov­ern­ment data be sub­mit­ted to or pro­duced by it. In­cid­ent no­ti­fic­a­tion is nar­rowed to the model used in con­tract per­form­ance, with a safe har­bour where a FedRAMP or CISA re­port car­ries sub­stan­tially the same in­form­a­tion. The gov­ern­ment may sus­pend use. De­com­mis­sion­ing li­ab­il­ity is capped at a quarter of the af­fected task or de­liv­ery order value.

M-25-22 sup­plies the di­li­gence and doc­u­ment­a­tion list sit­ting be­hind that clause. Agency pro­cesses must en­sure con­tracts per­man­ently pro­hibit the use of non-pub­lic in­put­ted agency data and out­put­ted res­ults to fur­ther train pub­licly or com­mer­cially avail­able al­gorithms ab­sent ex­pli­cit con­sent. Man­dat­ory award terms cover in­tel­lec­tual prop­erty and gov­ern­ment data, pri­vacy, pro­tec­tion against vendor lock-in, com­pli­ance with M-25-21 for high-im­pact uses and on­go­ing mon­it­or­ing. Con­tracts must also de­tail the vendor's test­ing pro­ced­ures, and must not pro­hibit agen­cies from in­tern­ally dis­clos­ing how the test­ing was done or what it found. The test­ing term is the sharpest.

What to put in the agreement
  • Di­li­gence on the model un­der­neath, named by pro­vider and ver­sion, since the Chinese fil­ing re­gime and the GSA clause both re­quire the third party to be iden­ti­fied.
AI governanceRules of the Game56 •
M-25-22 set its implementation clock in days rather than in years.
days from 3 April 2025
075150225300100GSA guide plan180Agency compliance200Tools repository270Procedure update
Source: OMB Memorandum M-25-22, 2025
  • Doc­u­ment­a­tion de­livered as a con­di­tion of ac­cept­ance, not on re­quest, so that the evid­ence pack ex­ists be­fore the first reg­u­lator's let­ter ar­rives.
  • Audit rights with a co­oper­a­tion clause, which the New York in­sur­ance cir­cu­lar of 11 July 2024 says it drew from the NAIC model bul­letin of Decem­ber 2023.
  • A sus­pen­sion right ex­er­cis­able at any time, fol­low­ing the GSA clause, with de­com­mis­sion­ing cost al­loc­ated rather than left to be ar­gued later.
  • Exit on defined no­tice with data and out­put re­turn, be­cause vendor lock-in is one of the five man­dat­ory award terms under M-25-22.

Buy­ers else­where do con­sid­er­ably less. The European Com­mis­sion's model con­trac­tual AI clauses, up­dated on 5 March 2025 with a factsheet of 23 Janu­ary 2026, are tem­plates and guid­ance rather than re­quire­ments. The United King­dom's Guidelines for AI Pro­cure­ment of 8 June 2020 have not been sub­stant­ively re­vised in six years, and the Crown Com­mer­cial Ser­vice buyer's guid­ance cites only Brit­ish in­stru­ments, men­tion­ing neither the NIST frame­work nor ISO/IEC 42001. Oman asks for more than either. Its pub­lic policy of 1 April 2025 re­quires pre-de­ploy­ment as­sess­ment and formal ap­proval by the com­pet­ent au­thor­ity be­fore a pub­lic body pro­cures an AI sys­tem.

The stand­ards gap is ad­mit­ted by the bloc it­self. On 7 Oc­to­ber 2025 the GCC Sec­ret­ary Gen­eral named the ab­sence of uni­fied le­gis­lat­ive stand­ards as the first of the re­main­ing chal­lenges, while stat­ing that mem­ber states' AI in­vest­ments have reached tens of bil­lions of dol­lars with plans to raise them to hun­dreds of bil­lions by 2030. The con­sequence for a buyer is that the coun­ter­party's ob­lig­a­tions are largely self-de­clared. Oman is ex­pli­cit about it. UNESCO's 2026 coun­try pro­file re­cords that Oman main­tains no cer­ti­fied re­gistry of AI

AI governanceRules of the Game57 •

vendors and that no sys­tem­atic mon­it­or­ing mech­an­ism has yet as­sessed how con­sist­ently its li­fe­cycle gov­ernance re­quire­ments are im­ple­men­ted.

Buying from a sovereign champion: the contract questions

When the counterparty is also the regulator's shareholder, diligence has to ask questions a standard vendor assessment never reaches.

Buy­ing from a na­tional cham­pion changes the di­li­gence, not the law. Sev­eral Gulf states have built their AI cap­ab­il­ity through bod­ies that also write the guid­ance those bod­ies' cus­tom­ers fol­low, and in one case through the body that en­forces the data pro­tec­tion stat­ute. The Saudi Data and Ar­ti­fi­cial In­tel­li­gence Au­thor­ity ad­min­is­ters the Per­sonal Data Pro­tec­tion Law, in force since 14 Septem­ber 2023. It also is­sued the na­tional AI risk man­age­ment frame­work launched on 14 July 2026. The roles sit to­gether.

The Emir­ates con­sol­id­ated their ma­chinery in 2026. On 14 June 2026 Sheikh Mo­hammed bin Rashid Al Mak­toum ap­proved the es­tab­lish­ment of the Ar­ti­fi­cial In­tel­li­gence and Data Au­thor­ity. The Cab­inet de­scribes it as the single na­tional body re­spons­ible for data, ar­ti­fi­cial in­tel­li­gence and di­gital gov­ern­ment, re­port­ing dir­ectly to the Cab­inet. It ab­sorbs the Of­fice of Ar­ti­fi­cial In­tel­li­gence, the Di­gital Gov­ern­ment Sec­tor at the tele­com­mu­nic­a­tions reg­u­lator and the UAE Data Of­fice. Omar Sul­tan Al Olama, Min­is­ter of State for Ar­ti­fi­cial In­tel­li­gence, was ap­poin­ted its Chair­man.

Read the man­date care­fully, be­cause its lim­its are the point. The Au­thor­ity pro­poses na­tional policies and le­gis­la­tion, and sets stand­ards and guidelines for data and AI man­age­ment, car­ry­ing the duty of en­sur­ing com­pli­ance across fed­eral en­tit­ies. The Cab­inet state­ment con­fers no rule-mak­ing power over the private sec­tor. Mean­while Fed­eral De­cree-Law No. 45 of 2021 re­mains in force with its Ex­ec­ut­ive Reg­u­la­tions still un­is­sued, so breach no­ti­fic­a­tion tim­ing, cross-bor­der trans­fer con­trols and data pro­tec­tion of­ficer thresholds are not yet fixed in bind­ing de­tail.

A body that issues the guidance can change the guidance. Diligence has to ask which hat is being worn.
AI governanceRules of the Game58 •
Across the Gulf, the binding AI rule sits below the national line.
binding instruments located, by level
01.534.560National, private1Free zone2Government only3Sector regulator6Advisory national
Source: DIFC, QCB, Abu Dhabi DoH, Dubai DHA and Bahrain iGA instruments, 2026

Three ques­tions fol­low from that struc­ture. The first is which ca­pa­city the coun­ter­party acts in when it makes a rep­res­ent­a­tion, be­cause a body that is­sues guid­ance can re­vise it. The second is what hap­pens to the con­tract if state policy turns, which is not hy­po­thet­ical. On 27 Feb­ru­ary 2026 the Gen­eral Ser­vices Ad­min­is­tra­tion an­nounced the re­moval of a named AI vendor from USAi.gov and from the Mul­tiple Award Sched­ule in sup­port of a pres­id­en­tial dir­ect­ive, with no stat­ute and no rule­mak­ing be­hind it, on a sched­ule that ac­coun­ted for over USD 52.5 bil­lion in sales in the 2025 fin­an­cial year.

The contract questions worth asking
  • Which ca­pa­city the coun­ter­party acts in, since the au­thor­ity is­su­ing the guid­ance may also own, fund or su­per­vise the vendor you are sign­ing with.
  • Whether com­pli­ance rep­res­ent­a­tions rest on a pub­lished in­stru­ment or on self-de­clar­a­tion, which UNESCO re­cords as Oman's po­s­i­tion on vendor cer­ti­fic­a­tion in 2026.
  • What hap­pens on a policy re­versal, in­clud­ing whether a del­ist­ing or a dir­ec­tion to cease use ter­min­ates, sus­pends or merely var­ies the agree­ment.
  • Who holds the cer­ti­fic­a­tion key, be­cause DIFC Reg­u­la­tion 10.3.3 bars high-risk use until the Com­mis­sioner pub­lishes re­quire­ments still at con­sulta­tion in mid-2026.

The third ques­tion is where the gate sits, and whether any­body holds the key. DIFC an­nounced in April 2026 that it in­ten­ded to be­come the world's first AI-nat­ive fin­an­cial centre, while the pro­hib­i­tion in its own Reg­u­la­tion 10.3.3 still de­pends on cer­ti­fic­a­tion re­quire­ments the Com­mis­sioner has not pro­mul­gated. Oman's Royal De­cree No. 50/2026 of 30 April 2026 es­tab­lished an Ar­ti­fi­cial In­tel­li­gence Spe­cial Zone in Mus­cat with tax breaks and cus­toms ex­emp­tions, and the de­cree in­tro­duces no AI reg­u­la­tion at all. In­cent­ive and ob­lig­a­tion issue from the same hand, at very dif­fer­ent speeds.

AI governanceRules of the Game59 •

The part in brief

  1. 01Classification comes first in every rulebook, yet Europe's classification guidelines missed their 2 February 2026 deadline and stayed in draft as high-risk duties moved to 2 December 2027.
  2. 02Evidence is the duty the European deferral left intact, and EN 18286:2026 is the single harmonised standard cited in the Official Journal, by a decision of 12 October 2026.
  3. 03Human oversight means different things in each instrument, from a named Autonomous Systems Officer in the DIFC to a per-action approval expectation that only Singapore's voluntary framework states.
  4. 04Three labelling regimes bind today and disagree, with Europe marking at the provider, India covering audio and visual material alone, and China pushing verification onto the distributing platform.
  5. 05The sharpest rules are contract terms, so a buyer of a sovereign champion's model should ask which capacity the counterparty acts in and who holds the certification key.
AI governanceRules of the Game60 •
What the mandates make you do

A voluntary framework became binding through a contract clause and a state penalty defence, never through a statute.

05

The operator's read

What binds now, the agent gap, and the single build.
  • What binds now, soonest first
  • The agent gap, and what to do before it closes
  • A single build to the highest applicable bar
  • Method and scope
  • Volge's expertise in AI governance
  • Sources

What binds now, soonest first

A dated list of what already applies and what lands next, counted forward from 6 November 2026.

Four things bind today that did not bind a year ago, and the next twenty-two months carry more com­mence­ments than the pre­vi­ous four years. The pace has changed. In Europe the gen­eral date of ap­plic­a­tion passed on 2 Au­gust 2026, bring­ing the Art­icle 50 trans­par­ency du­ties and the Art­icle 101 fines on gen­eral-pur­pose model pro­viders into force. In India the syn­thetic con­tent rules took ef­fect on 20 Feb­ru­ary 2026. In China the amended Cy­ber­se­cur­ity Law raised the ceil­ing ten­fold on 1 Janu­ary 2026. In Texas the Re­spons­ible Ar­ti­fi­cial In­tel­li­gence Gov­ernance Act com­menced the same day.

The nearest date is four weeks away. On 2 Decem­ber 2026 two European du­ties fall due to­gether. New Art­icle 111(4) re­quires pro­viders of sys­tems gen­er­at­ing syn­thetic con­tent placed on the mar­ket be­fore 2 Au­gust 2026 to com­ply with the Art­icle 50(2) mark­ing duty, a trans­itional period re­cital (38) of Reg­u­la­tion (EU) 2026/1744 de­scribes as four months. The same day the two new Art­icle 5 pro­hib­i­tions apply, reach­ing non-con­sen­sual in­tim­ate im­agery and ma­ter­ial within Art­icle 2 points (c) and (e) of Dir­ect­ive 2011/93/EU.

Counted forward from this edition, the commencements cluster early and then thin out.
months from 6 November 2026
06.212.518.825013 Nov 202612 Dec 202621 Jan 2027613 May 2027117 Oct 2027132 Dec 2027212 Aug 2028
Source: EUR-Lex, MeitY, MAS and state legislatures, 2026

India's own nearest date will be a week old by the time this is read. Rule 4 of the Di­gital Per­sonal Data Pro­tec­tion Rules 2025, on the re­gis­tra­tion and ob­lig­a­tions of Con­sent Man­agers, com­mences on 13 Novem­ber 2026, one year after the Rules were no­ti­fied. Sub­stant­ive du­ties fol­low on 13 May 2027. In between, the Tele­com Reg­u­lat­ory Au­thor­ity of India closes com­ments on 27 Novem­ber 2026 on its draft guidelines on re­spons­ible use of AI in tele­com­mu­nic­a­tion net­works, re­leased on 30 Oc­to­ber 2026.

AI governanceRules of the Game61 •

1 Janu­ary 2027 is the heav­iest single date on the list. Four re­gimes land at once. Col­or­ado's SB 26-189 takes ef­fect, hav­ing been signed on 14 May 2026 to re­place an Act that never com­menced at all. New York's RAISE Act com­mences as sub­sti­tuted by Chapter 96 of the Laws of 2026, signed on 27 March 2026. The Illinois Ar­ti­fi­cial In­tel­li­gence Safety Meas­ures Act com­mences, ap­proved on 6 July 2026. Cali­for­nia's auto­mated de­cision-mak­ing tech­no­logy com­pli­ance date ar­rives under the Pri­vacy Pro­tec­tion Agency reg­u­la­tions, as do the plat­form du­ties added by AB 853.

Almost every near date is a labelling duty. Almost every far date needs a standard nobody has written yet.

After that the dates thin out. The Mon­et­ary Au­thor­ity of Singa­pore is­sued its Guidelines on Ar­ti­fi­cial In­tel­li­gence Risk Man­age­ment on 7 Oc­to­ber 2026, with sec­tions 3 and 4 tak­ing ef­fect on 7 Oc­to­ber 2027 and sec­tions 5 and 6 by 7 Oc­to­ber 2028. The European high-risk ob­lig­a­tions apply to Annex III sys­tems from 2 Decem­ber 2027 and to Annex I sys­tems from 2 Au­gust 2028, with leg­acy pub­lic-au­thor­ity sys­tems run­ning to 2 Au­gust 2030. Utah's Ar­ti­fi­cial In­tel­li­gence Policy Act ex­pires on 1 July 2027 un­less ex­ten­ded again.

Two things are worth no­ti­cing about the shape of that list. Al­most every near date is a trans­par­ency or la­belling duty, and al­most every far date is a con­form­ity duty. En­gin­eer­ing can meet the near ones. The far ones can­not be met without the stand­ards nobody has yet writ­ten. One har­mon­ised stand­ard has been cited in the Of­fi­cial Journal, and the de­fer­ral that bought six­teen months for the Annex III du­ties was jus­ti­fied by the ab­sence of the rest.

The agent gap, and what to do before it closes

No statute anywhere is framed around AI agents, and the law that already reaches them was written for electronic data interchange.

The word is miss­ing from the stat­utes. Reg­u­la­tion (EU) 2024/1689 con­tains no in­stance of agent in any sense, against 484 in­stances of high-risk in the same text. Texas HB 149 con­tains none, and no in­stance of autonom­ous or autonomy either. The pub­lished ETSI baseline con­tains none. The ninety-seven titles in the ISO/IEC JTC 1/SC 42 work pro­gramme con­tain none. The single ex­cep­tion in bind­ing European law is a con­form­ity-as­sess­ment code.

AI governanceRules of the Game62 •

That code is worth nam­ing be­cause it is so small. Annex XIV, in­ser­ted into the AI Act by Art­icle 1 point 43 of Reg­u­la­tion (EU) 2026/1744, fixes the scope of des­ig­na­tion for no­ti­fied bod­ies, and code AIH 0401 cov­ers sys­tems based on other emer­ging AI tech­no­lo­gies not covered by other codes, in­clud­ing Agen­tic AI. It im­poses no ob­lig­a­tion on any­body. On the evid­ence gathered it is the first ap­pear­ance of that phrase in bind­ing European law, and it is a tax­onomy entry rather than a duty.

Count the word agent in the instruments most often cited as governing agents.
occurrences in the full text
00.20.50.810EU AI Act0ETSI EN 304 2230SC 42 titles0Texas HB 1491Digital Omnibus
Source: EUR-Lex, ETSI, ISO and Texas Legislature, 2026

One in­stru­ment any­where reg­u­lates agent autonomy as such, and it is Chinese. The Im­ple­ment­a­tion Opin­ions on the Stand­ard­ised Ap­plic­a­tion and In­nov­at­ive De­vel­op­ment of Agents, pub­lished by the Cy­ber­space Ad­min­is­tra­tion of China with the Na­tional De­vel­op­ment and Re­form Com­mis­sion and the Min­istry of In­dustry and In­form­a­tion Tech­no­logy on 8 May 2026, define an agent as an in­tel­li­gent sys­tem pos­sess­ing autonom­ous per­cep­tion, memory, de­cision-mak­ing, in­ter­ac­tion and ex­e­cu­tion cap­ab­il­it­ies. Item 6 is the au­thor­ity rule. It re­quires three de­cision modes to be de­lin­eated, and states that an agent's ex­e­cu­tion of op­er­a­tions must not ex­ceed the scope of the user's au­thor­isa­tion.

Its legal char­ac­ter mat­ters too. It is an im­ple­ment­a­tion opin­ion giv­ing ef­fect to the State Coun­cil's AI Plus Opin­ions of 21 Au­gust 2025, so it binds state or­gans and steers in­dustry rather than car­ry­ing pen­al­ties of its own. Item 11 sets tiered gov­ernance, with cy­ber­space de­part­ments and sec­tor reg­u­lat­ors fix­ing which scen­arios are open in sens­it­ive fields, and self-test­ing and in­dustry self-dis­cip­line in lower-risk areas. Item 4 pro­poses an agent re­gis­tra­tion plat­form with di­gital iden­tity man­age­ment and cap­ab­il­ity de­clar­a­tion.

E-SIGN presupposes legal attribution and does not supply it. Attribution is the whole of the agent gap.
AI governanceRules of the Game63 •

What binds an agent today in the West is older than the tech­no­logy. Sec­tion 14 of the Uni­form Elec­tronic Trans­ac­tions Act, en­acted in Michigan as Act 305 of 2000 and im­me­di­ately ef­fect­ive on 16 Oc­to­ber 2000, provides that a con­tract may be formed by the in­ter­ac­tion of elec­tronic agents even if no in­di­vidual was aware of or re­viewed their ac­tions. E-SIGN is nar­rower still. It is a valid­ity rule that pre­sup­poses legal at­tri­bu­tion rather than sup­ply­ing it, and the at­tri­bu­tion ques­tion is the gap.

What to do before the gap closes
  • Clas­sify by trans­fer of agency, fol­low­ing OCC Bul­letin 2026-54 of 4 Novem­ber 2026, which places an autonom­ous pay­ment agent in its most crit­ical tier.
  • Set ac­tion bound­ar­ies re­quir­ing human ap­proval be­fore ir­re­vers­ible steps, as the IMDA agen­tic frame­work asks, and deny by de­fault when ap­proval in­fra­struc­ture fails.
  • Fix au­thor­ity by con­tract and by con­fig­ur­a­tion, using the Chinese three-mode split between user-only, user-au­thor­ised and agent-autonom­ous de­cisions as the draft­ing tem­plate.
  • Read the OWASP Agen­tic Top 10 of 9 Decem­ber 2025 against your own tool per­mis­sions, and treat ex­cess­ive agency as a design de­fect rather than a policy ques­tion.
  • Note that the Agent Con­trol Stand­ard pub­lished on 1 Septem­ber 2026 stands at ver­sion 0.1.0 and de­faults to fail-open, so it is a dir­ec­tion of travel.

Two more bind­ing re­gimes already reach sys­tems that trans­act without per-ac­tion ap­proval, and both are sec­toral. Com­mis­sion Del­eg­ated Reg­u­la­tion (EU) 2017/589, in force since 3 Janu­ary 2018, im­poses or­gan­isa­tional re­quire­ments on in­vest­ment firms en­gaged in al­gorithmic trad­ing, in­clud­ing pre-trade con­trols and the abil­ity to can­cel un­ex­ecuted or­ders. FINRA has solved the ac­count­ab­il­ity ques­tion once, by re­gis­ter­ing a named human re­spons­ible for the design of an al­gorithmic trad­ing strategy, man­dat­ory since 30 Janu­ary 2017. Neither was writ­ten for con­sumer agents. Both show what a work­able an­swer looks like.

A single build to the highest applicable bar

Build once to the strictest version of each duty, because the strictest version is rarely in the same jurisdiction twice.

Build­ing five com­pli­ance pro­grammes for five rule­books costs more than build­ing one to the highest bar, and the arith­metic is not close. The du­ties over­lap heav­ily in sub­stance and di­verge in de­tail, so the ex­pens­ive part is never the con­trol it­self but the five ver­sions of the evid­ence. The harder ques­tion is where the highest bar sits for each duty, be­cause it moves. No jur­is­dic­tion holds every column.

AI governanceRules of the Game64 •

On doc­u­ment­a­tion the highest bar is European and audit­able. EN 18286:2026 turns the Art­icle 17 qual­ity man­age­ment re­quire­ments into audit­able re­quire­ments for pro­viders, and Art­icle 40 of Reg­u­la­tion (EU) 2024/1689 gives con­form­ity with a cited har­mon­ised stand­ard a pre­sump­tion of con­form­ity run­ning from pub­lic­a­tion in the Of­fi­cial Journal. ISO/IEC 42001:2023 is cer­ti­fi­able, which ISO's own cata­logue es­tab­lishes through the ex­ist­ence of ISO/IEC 42006:2025 set­ting re­quire­ments for cer­ti­fic­a­tion bod­ies. No stat­ute ex­amined de­mands either. Both are the cheapest way to pro­duce one evid­ence set rather than five.

On la­belling the bar is Chinese. It is the only one with a man­dat­ory tech­nical stand­ard be­hind it. GB 45438-2025 was is­sued on 28 Feb­ru­ary 2025 and took ef­fect on 1 Septem­ber 2025 as a man­dat­ory na­tional stand­ard, spe­cify­ing meth­ods for ex­pli­cit la­bels and for im­pli­cit la­bels em­bed­ded in file metadata. Art­icle 6 of the La­belling Meas­ures then pushes veri­fic­a­tion down­stream to the dis­sem­in­a­tion plat­form. A firm that builds to the Chinese metadata re­quire­ment has sat­is­fied the European ma­chine-read­able duty and the In­dian proven­ance duty on the way.

A firm that builds to the Chinese metadata requirement satisfies the European and Indian duties on the way.

On human over­sight the highest bar is vol­un­tary and Singa­por­ean, which is awk­ward but true. The IMDA agen­tic frame­work asks for defined ac­tion bound­ar­ies re­quir­ing human ap­proval be­fore ir­re­vers­ible ac­tions. It asks for denial by de­fault when the ap­proval in­fra­struc­ture fails, and for audit of the over­ride rate it­self. Noth­ing bind­ing goes that far. Cali­for­nia SB 947 re­quires cor­rob­or­a­tion be­fore dis­cip­line or ter­min­a­tion from 1 July 2027, and Col­or­ado SB 26-189 gives a right to human re­view after an ad­verse out­come from 1 Janu­ary 2027.

The European tiers set the ceil­ing. They are what should size the pro­gramme. Non-com­pli­ance with the Art­icle 5 pro­hib­i­tions at­tracts up to EUR 35,000,000 or 7 per cent of world­wide an­nual turnover, whichever is higher. Op­er­ator du­ties and the Art­icle 50 trans­par­ency ob­lig­a­tions at­tract up to EUR 15,000,000 or 3 per cent, and the same tier ap­plies to gen­eral-pur­pose model pro­viders under Art­icle 101, in force since 2 Au­gust 2026. China's Per­sonal In­form­a­tion Pro­tec­tion Law reaches RMB 50,000,000 or 5 per cent of turnover, and the In­dian Act reaches Rs 250 crore.

Two cau­tions be­long with the re­com­mend­a­tion. The first is that build­ing to the highest bar does not dis­charge the low­est, be­cause some du­ties are pro­ced­ural rather than sub­stant­ive,

AI governanceRules of the Game65 •
The European tiers remain the ceiling that should size the programme.
EUR million, maximum fine
01020304035Article 5 bans15Operator duties15GPAI providers7.5Misleading info
Source: Regulation (EU) 2024/1689, Article 99

and a Chinese al­gorithm fil­ing or an In­dian Con­sent Man­ager re­gis­tra­tion can­not be sat­is­fied by a bet­ter con­trol some­where else. The second is that the highest bar can be lowered. The European high-risk ob­lig­a­tions moved out by six­teen and twelve months in July 2026, and the Col­or­ado duty of care was re­pealed on 14 May 2026, six weeks be­fore the de­ferred date on which it was due to apply.

Method and scope

What this edition covers, how a fact was accepted, and what was deliberately left out of it.

This edi­tion states the po­s­i­tion as at 6 Novem­ber 2026. Noth­ing later is in­cluded. Every date and fig­ure rests on a primary in­stru­ment or on a named, dated sec­ond­ary source, and where only com­ment­ary could be reached the com­ment­ator is named in the text. Where two sources dis­agree, the primary one gov­erns and the dis­agree­ment is re­cor­ded rather than re­solved si­lently. Four fact bases sit be­hind these chapters, cov­er­ing the global pic­ture, the United States, India and the Gulf.

Three words carry a fixed sense through­out. In force means a pro­vi­sion ap­plies to some­body today. Sched­uled means a com­mence­ment date is fixed in an in­stru­ment that has already been made, as with the In­dian sub­stant­ive rules on 13 May 2027 or the European Annex III du­ties on 2 Decem­ber 2027. An­nounced means a body has said it will do some­thing without is­su­ing the in­stru­ment, as with the Se­cur­it­ies and Ex­change Board of India's state­ment of 10 Oc­to­ber 2026 that guidelines on AI and ma­chine learn­ing would fol­low shortly.

Every count car­ries its rule. The num­ber of Amer­ican state AI laws en­acted dur­ing 2026 ranges from three to 115 de­pend­ing on defin­i­tion, so every fig­ure used here names its source and its win­dow. No count of jur­is­dic­tions hold­ing a bind­ing AI-spe­cific stat­ute ap­pears

AI governanceRules of the Game66 •

any­where in this edi­tion. None is pub­lished by the In­ter­na­tional As­so­ci­ation of Pri­vacy Pro­fes­sion­als tracker, by the OECD Policy Nav­ig­ator or by the Stan­ford AI Index.

Four things were left out de­lib­er­ately. Data centre and com­pute policy sits out­side scope, as does ex­port con­trol. Copy­right and train­ing data ap­pear only where an in­stru­ment im­poses a dis­clos­ure duty, which is why the Brit­ish re­port of 18 March 2026 is noted for drop­ping a broad ex­cep­tion with opt-out and not dis­cussed fur­ther. Lit­ig­a­tion is noted where it changed what ap­plies and oth­er­wise left alone. Fig­ures that could not be traced to a page ac­tu­ally opened were ex­cluded, in­clud­ing cer­ti­fic­a­tion counts for ISO/IEC 42001, for which no pub­lic au­thor­it­at­ive fig­ure ex­ists.

Volge's expertise in AI governance

The work a buyer should have in place, how an engagement of this kind is run, and what it produces.

A buyer fa­cing four rule­books needs four things in place be­fore it needs ad­vice. The first is an in­vent­ory nam­ing every sys­tem, its owner and the de­cision its out­put feeds. The second is a clas­si­fic­a­tion de­cision for each one, re­cor­ded with the test ap­plied and the date. The third is a doc­u­ment­a­tion set that would sur­vive a reg­u­lator's re­quest without re­con­struc­tion. The fourth is a named human ac­count­able for over­sight of every sys­tem above the low­est tier. That is the floor. Volge builds those four, in that order, and then main­tains them against the dates in the pre­ced­ing chapter.

The work types
  • In­vent­ory and clas­si­fic­a­tion, map­ping every sys­tem against the European, In­dian, Chinese and Amer­ican clas­si­fic­a­tion tests, and re­cord­ing each de­cision with its date.
  • Doc­u­ment­a­tion and evid­ence, build­ing the qual­ity man­age­ment set to EN 18286:2026 and the im­pact as­sess­ment and audit ca­dence that Rule 13 re­quires an­nu­ally.
  • Dis­clos­ure and proven­ance en­gin­eer­ing, im­ple­ment­ing metadata and vis­ible la­belling once, to the strict­est of the three la­belling re­gimes in the cli­ent's scope.
  • Pro­cure­ment and third-party gov­ernance, draft­ing di­li­gence, audit rights, sus­pen­sion and exit terms for the mod­els a cli­ent did not build and can­not in­spect.
  • Over­sight design, fix­ing ac­tion bound­ar­ies and fail-safes, with the over­ride tele­metry needed to demon­strate that an ap­proval gate is real rather than nom­inal.

The com­mon fail­ures are con­sist­ent. In­vent­or­ies are built once and never main­tained, so a clas­si­fic­a­tion that was ac­cur­ate on the day it was signed is wrong within a quarter.

AI governanceRules of the Game67 •
How an engagement runs
  • One ac­count­able part­ner, a de­liv­ery lead and a work­ing team drawn from Global De­liv­ery Centres, sized to the cli­ent's sys­tem count rather than to its head­count.
  • Fixed phases with named de­liv­er­ables and a dated exit, so the cli­ent ends the en­gage­ment own­ing a main­tained arte­fact rather than a re­port about one.
  • Weekly work­ing ses­sions with the sys­tem own­ers them­selves, be­cause clas­si­fic­a­tion can­not be done at a dis­tance from the people who ac­tu­ally run the sys­tems.
  • A stand­ing legal and en­gin­eer­ing pair­ing on every work­stream, since every duty in these rule­books has both a draft­ing com­pon­ent and a build com­pon­ent.
The evidence produced
  • A sys­tem in­vent­ory car­ry­ing own­ers, tiers, the test ap­plied and the date, held in the cli­ent's own sys­tems rather than in a con­sult­ant's work­ing file.
  • A clas­si­fic­a­tion memor­andum for each sys­tem, writ­ten so that a su­per­visor read­ing it two years later can fol­low the reas­on­ing back to its source.
  • The qual­ity man­age­ment and im­pact as­sess­ment set, mapped clause by clause to the in­stru­ment each clause an­swers, with the gaps named rather than smoothed over.
  • A con­tract an­nexe for AI pro­cure­ment cov­er­ing di­li­gence, doc­u­ment­a­tion, audit rights, sus­pen­sion and exit, draf­ted to sit on the cli­ent's own paper.
How it is measured
  • Share of sys­tems in­vent­or­ied and clas­si­fied, against a de­nom­in­ator the cli­ent agrees at the start and does not quietly re­vise down­ward later.
  • Days from a reg­u­lator's re­quest to a com­plete evid­ence pack, meas­ured on a re­hearsal rather than es­tim­ated from the con­tents of the file.
  • Com­mence­ment dates met, coun­ted against the dated list in the pre­ced­ing chapter, with each duty either met, sched­uled or form­ally ac­cep­ted as a risk.
  • Over­ride rate and re­viewer re­sponse time on every sys­tem car­ry­ing an ap­proval gate, since a gate nobody uses is not an over­sight con­trol at all.

Clas­si­fic­a­tion is del­eg­ated to pro­cure­ment, which sees the con­tract rather than the de­cision the out­put feeds. Evid­ence is writ­ten after the fact, which is the one thing no reg­u­lator ac­cepts. Over­sight is as­signed to some­body who has neither the time nor the au­thor­ity to re­fuse an ac­tion.

AI governanceRules of the Game68 •
The operator's read

A firm that builds to the Chinese metadata requirement satisfies the European and Indian duties on the way.

Sources

Every body cited in this read, with the instrument and the date carried in the text.

AI governanceRules of the Game69 •
Version · VRG-AIR-001 · 2026
volgeEnable Powerful
Transformations
About Volge

Volge is an engineering firm providing enterprise technology transformation services to clients across the Middle East, India, APAC, Europe, and North America.

Delivery stands on two pillars: Forward Deployed Engineering, with senior specialists embedded in the client’s own team, and Intelligent Command Centre Operations, run from our Global Delivery Centres, which keeps what they build running. Senior engineers stay with an engagement from the first diagnosis through to production.

Our services span Strategy, Data, Platforms, AI and Automation, Cyber Trust, Technology Operations, and Applied Functional Engineering. Engagements begin with a diagnosis. The firm is built on working autonomy: engineers decide, innovate, co-create, and own their work. Motivated engineers build the systems that sustain our clients’ growth.

Terms of use

This edition may be shared with organisations evaluating a Volge engagement. It may not be altered or republished without Volge’s permission, and figures carry their stated basis and date. Volge, Intelligent Command Centre, and Intelligent Command Centre Operations are marks of Volge; other names are the trademarks of their respective owners.