>
volgeMicrosoft · 2027
APPLIED AI · VOICE OF VOLGE

From copilot to autopilot

Autonomous agents as a new class of managed digital labour

Voice of Volge is a series of books, and the firm's thought leadership on the solutions and the domains Volge specialises in. Each edition takes one of them and sets out what the technology makes possible, what it makes mandatory, and what an operator has to put in place before either is true of them.

Autonomous agents as a new class of managed digital labour

Figures carry their basis and their date. Where published estimates disagree, the range is given and the basis named rather than averaged away, and where a widely quoted number does not survive checking it is corrected here rather than repeated. Every chart plots a figure stated in the text, and nothing is projected beyond what a named source projects.

Contents

  1. 02Executive summary
  2. 03, Why agents die in pilot
  3. 05, The control surface that decides the outcome
  4. 08, The operating model as the binding constraint
  5. 11, Reference architecture and what actually ships
  6. 15, Agents in the Gulf and India
  7. 19, The leadership agenda
  8. 20Appendices
Voice of Volge is a series of books, and the firm’s thought leadership on the solutions and the domains Volge specialises in.
Microsoft · 2027
Applied AIVoice of Volge01 •

Executive summary

Most agent pro­grammes never reach pro­duc­tion, and they stall for reas­ons that have noth­ing to do with the model. Only a quarter of en­ter­prises have moved 40 per cent or more of their AI pi­lots into pro­duc­tion, and just 21 per cent of those plan­ning agent de­ploy­ments within two years re­port a ma­ture gov­ernance model for them (De­loitte, 21 Janu­ary 2026). Gart­ner pro­jects more than 40 per cent of agen­tic AI pro­jects can­celled by the end of 2027, in­ad­equate risk con­trols among the causes (Gart­ner, 25 June 2025). Ninety-eight per cent of senior AI ex­ec­ut­ives at lis­ted com­pan­ies above one bil­lion dol­lars in rev­enue re­port a formal AI gov­ernance policy, and 47 per cent say their or­gan­isa­tion has pre­vi­ously not fol­lowed it for ur­gent de­ploy­ments (EY, 15 Septem­ber 2026). Hold that pro­duc­tion rate in mind. It is the base rate every agent in­vest­ment is bet­ting against.

The sup­ply side is not the prob­lem. Agents now ar­rive em­bed­ded in soft­ware the en­ter­prise already owns rather than as a sep­ar­ate pur­chase. Mi­crosoft made agen­tic ac­tion gen­er­ally avail­able in­side Word, Excel, and PowerPoint on 22 April 2026, and brought Agent 365 to gen­eral avail­ab­il­ity as the en­ter­prise con­trol plane on 1 May 2026, priced at 15 United States dol­lars per user per month stan­dalone or in­cluded in Mi­crosoft 365 E7 at 99 dol­lars per user per month (Mi­crosoft, 22 April and 1 May 2026). Autonomy is now abund­ant, cheap, and already pro­vi­sioned. The con­straint sits en­tirely on the de­mand side: the ca­pa­city to ab­sorb autonomy without los­ing con­trol of it.

Cross that threshold and the pic­ture changes, though the meas­ured re­cord is nar­rower than the am­bi­tion. Among or­gan­isa­tions with rev­enue above USD 1 bil­lion, 40 per cent now re­port scal­ing AI agents in at least one func­tion, up from 27 per cent a year earlier (McKinsey, Au­gust 2026). Over the same period the share of re­spond­ents at­trib­ut­ing any pos­it­ive EBIT im­pact to AI held at 37 per cent, and the pro­por­tion qual­i­fy­ing as AI high per­formers held at about 6 per cent. Scal­ing rose and en­ter­prise-level fin­an­cial im­pact did not fol­low. The dis­tance between de­ploy­ment and de­clared profit is the whole prob­lem.

The ar­gu­ment in one line. En­ter­prises that cap­ture the re­turn treat agents as a new class of di­gital la­bour, given iden­tity and scoped au­thor­ity, su­per­vised and per­form­ance-man­aged, and owned by a named ac­count­able ex­ec­ut­ive. En­ter­prises that get can­celled treat an agent as a chat­bot with el­ev­ated per­mis­sions and live sys­tem ac­cess. They dis­cover the blast ra­dius after it has already det­on­ated.

This is a paper about ex­e­cu­tion, not se­lec­tion. For most large en­ter­prises the plat­form de­cision is already made by ex­ist­ing

Mi­crosoft com­mit­ments. The de­cision that re­mains open, the one that de­term­ines whether the pro­gramme reaches audited pro­duc­tion or joins the 40 per cent can­cel­la­tion pool, is op­er­a­tional: iden­tity, gov­ernance, op­er­at­ing-model design, and the dis­cip­line of mov­ing along the autonomy spec­trum no faster than the con­trols allow.

Five con­clu­sions, de­veloped in the chapters that fol­low.

First, autonomy is gradu­ated, not bin­ary. Work runs from as­sist­ive to agen­tic to autonom­ous, and each step car­ries a sharply higher con­trol re­quire­ment. En­ter­prises that jump straight to full autonomy, skip­ping the su­per­vised middle, sup­ply a dis­pro­por­tion­ate share of the can­celled pro­jects.

Second, value con­cen­trates in high-volume, rules-bound, di­git­ally nat­ive work. Ser­vice op­er­a­tions, IT op­er­a­tions, the fin­ance close, and soft­ware en­gin­eer­ing are the early win­ners. The re­turn comes from cycle-time com­pres­sion and la­bour re­deploy­ment, not head­count re­duc­tion taken alone.

Third, agents stall in a con­sist­ent order. Gov­ernance comes first, then in­teg­ra­tion debt and or­gan­isa­tional read­i­ness. The model is al­most never the bind­ing con­straint.

Fourth, the op­er­at­ing model is the dif­fer­en­ti­ator. Agents need an ac­count­able owner, an op­er­a­tions func­tion , AgentOps , and an ex­pli­cit human-team­ing model. De­loitte found that only 21 per cent of the com­pan­ies plan­ning agent de­ploy­ments within two years re­port a ma­ture model for agent gov­ernance (De­loitte, 21 Janu­ary 2026). Where that ac­count­ab­il­ity is miss­ing, nobody is an­swer­able for an agent's ac­tions, nobody is em­powered to au­thor­ise its ac­cess, and nobody owns its de­com­mis­sion­ing.

Fifth, iden­tity is the front­line. An agent without a first-class iden­tity can­not be in­vent­or­ied or audited, and it can­not be scoped or re­voked, so it can­not be gov­erned at all. Twenty-six per cent of senior AI ex­ec­ut­ives whose or­gan­isa­tions use agen­tic AI say they can­not de­tect un­au­thor­ised agents op­er­at­ing in­side their own sys­tems (EY, 15 Septem­ber 2026). Entra Agent ID and the Zero Trust pos­ture for agents are not se­cur­ity gar­nish. They are the pre­con­di­tion for everything else, and they now carry an en­ti­tle­ment. Since Mes­sage Cen­ter no­tice MC1395007 of 17 June 2026, the Con­di­tional Ac­cess and ID Pro­tec­tion con­trols for agents re­quire Agent 365 or Mi­crosoft 365 E7 rather than E5.

That asym­metry mat­ters more in this re­gion than any li­cens­ing de­cision in this paper. Saudi Ar­a­bia has no Azure re­gion and no Mi­crosoft 365 res­id­ency com­mit­ment, fall­ing into the cat­egory Mi­crosoft calls All Other Geo­graph­ies, where cus­tomer data "isn't com­mit­ted to reside in any par­tic­u­lar data­cen­ter" and the loc­a­tion "is sub­ject to change without no­tice" (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). Qatar has

Applied AIVoice of Volge02 •

an Azure re­gion and a res­id­ency com­mit­ment, yet it is ab­sent from the Co­pi­lot Stu­dio data loc­a­tions table, which by Mi­crosoft's own rule puts a Qatari ten­ant's agent data in the United States. Agent 365, the con­trol plane, has no re­gional dial at all. Chapter five sets out what an ar­chi­tect does about each.

Sixth, the re­gion de­cides what the ar­chi­tec­ture is al­lowed to do. Mi­crosoft com­mits to stor­ing data at rest in the ten­ant's geo­graphy. It does not com­mit to run­ning in­fer­ence there. Its own Co­pi­lot pri­vacy doc­u­ment­a­tion states that calls to the large lan­guage model are routed to the closest data­centres in the re­gion but can also call into other re­gions where ca­pa­city is avail­able dur­ing high util­isa­tion peri­ods, and that cus­tom­ers out­side the European Union may have their quer­ies pro­cessed in the United States, the EU or other re­gions (Mi­crosoft Learn, page up­dated 3 Novem­ber 2026). The European Union Data Bound­ary is the only pro­cessing com­mit­ment Mi­crosoft pub­lishes any­where. No Gulf or In­dian equi­val­ent ex­ists.

The Mi­crosoft es­tate is the primary lens here, re­flect­ing its po­s­i­tion as the de­fault pro­ductiv­ity and iden­tity layer of the large en­ter­prise. The com­pet­it­ors , Sales­force Agent­force, ServiceNow, Amazon Bed­rock AgentCore, Google Gem­ini En­ter­prise , ap­pear where they sharpen the ana­lysis. The open pro­to­cols that stitch these es­tates to­gether, prin­cip­ally Model Con­text Pro­tocol (MCP) and Agent-to-Agent (A2A), are treated as load-bear­ing, be­cause the pro­duc­tion en­ter­prise will be multi-vendor whether it plans for that or not.

, Why agents die in pilot

Agents stall for three reasons, and they bind in a fixed order: governance, integration debt, organisational readiness. None is a technology problem. The value waiting on the far side is real and large but conditional on production, and the conditionality , not the headline return , is what a board needs to underwrite. Read the failure anatomy first, because it is the cheapest way to avoid joining the 40 per cent.

The two numbers, and why they do not conflict

Forty per cent of large en­ter­prises scal­ing agents, sit­ting next to 37 per cent re­port­ing any EBIT im­pact from AI, looks like a con­tra­dic­tion. It is a se­lec­tion ef­fect, and it is the most im­port­ant ana­lyt­ical point in this doc­u­ment.

The board-level read­ing is pre­cise. The in­vest­ment thesis is not "agents gen­er­ate a re­turn". It is "a re­turn ap­pears only where an en­ter­prise crosses the pro­duc­tion threshold and re­designs the work be­hind it, and most en­ter­prises do neither". Nearly three-quar­ters of high per­formers have fun­da­ment­ally

re­designed work­flows, against one quarter of other re­spond­ents (McKinsey, Au­gust 2026). The en­tire ma­na­gerial prob­lem is the cross­ing. That re­frames the agenda away from plat­form se­lec­tion and onto ex­e­cu­tion dis­cip­line, which is where the rest of this paper sits.

De­ploy­ment is not de­liv­ery. Scal­ing counts pro­grammes that run, EBIT im­pact counts pro­grammes that pay, and the second pop­u­la­tion is much the smal­ler. Just 6 per cent of re­spond­ents at­trib­ute 5 per cent or more of EBIT to their use of AI (McKinsey, Au­gust 2026). A quarter of en­ter­prises have moved 40 per cent or more of their pi­lots into pro­duc­tion (De­loitte, 21 Janu­ary 2026), and the share of com­pan­ies abandon­ing most of their AI ini­ti­at­ives rose to 42 per cent from 17 per cent the year be­fore (S&P Global Mar­ket In­tel­li­gence, March 2025). The re­turn is con­cen­trated in the few that shipped and then re­built the work around what shipped.

PopulationShare of enterprisesMeasured outcome
Attributing 5% or more of EBIT to AI, with significant value6%AI high performers; workflows fundamentally redesigned
Attributing any positive EBIT impact to AI37%Impact reported, mostly below 5% of EBIT
Scaling agents in at least 1 function (revenue above USD 1bn)40%In production; profit impact not yet declared
Abandoned most AI initiatives42%Withdrawn, up from 17% the year before

Sources: McKinsey, Au­gust 2026; S&P Global Mar­ket In­tel­li­gence, March 2025. Pop­u­la­tions are drawn from dif­fer­ing sur­vey bases and are not strictly ad­dit­ive.

The fun­nel below traces the same pop­u­la­tion from in­terest to cap­tured re­turn. The re­turn con­cen­trates in the final, nar­row band , the pro­duc­tion minor­ity this paper is con­cerned with.

The failure anatomy: three causes, fixed order

Gart­ner at­trib­utes the pro­jec­ted can­cel­la­tions to un­der­in­vest­ment in gov­ernance, in­teg­ra­tion com­plex­ity, and or­gan­isa­tional read­i­ness re­l­at­ive to the tech­no­logy it­self (Gart­ner, 25 June 2025). Those three recur across every pro­duc­tion re­view, and they bind in se­quence. Each is pre­vent­able, which is the only reason a can­cel­la­tion base rate above 40 per cent is ac­cept­able to design against.

Second, in­teg­ra­tion debt. Agents cre­ate value by act­ing against live sys­tems, and most en­ter­prises carry years of ac­cu­mu­lated in­teg­ra­tion debt in brittle in­ter­faces and un­doc­u­mented data, with in­con­sist­ent ac­cess con­trols on top. Foundry's lib­rary of pre-built con­nect­ors and the MCP stand­ard exist pre­cisely to re­tire this debt. They do not erase the

Applied AIVoice of Volge03 •

un­der­ly­ing data and ac­cess-con­trol de­fi­cien­cies be­neath it. In­teg­ra­tion debt is the cause most often missed at the busi­ness-case stage, and most often re­spons­ible for sched­ule over­run once the agent meets the real es­tate.

First , ab­sent gov­ernance. Formal policy is not the gap. Among senior AI ex­ec­ut­ives at lis­ted com­pan­ies above one bil­lion dol­lars in rev­enue, 98 per cent re­port a formal AI gov­ernance policy, yet 49 per cent of those whose or­gan­isa­tions run agen­tic AI say that frame­work has not been up­dated for agen­tic re­quire­ments and risks at all (EY, 15 Septem­ber 2026). Gov­ernance binds first be­cause, without it, an agent can­not be safely gran­ted the au­thor­ity that makes it use­ful. The pro­gramme then forks into two fail­ure modes: an agent too con­strained to de­liver value, or an agent too per­mis­sioned to be safe. Both end in can­cel­la­tion.

Third , or­gan­isa­tional un­read­i­ness. The last con­straint is the op­er­at­ing-model gap: no owner, no op­er­a­tions func­tion, no team­ing model, and no plan to re­deploy the la­bour the agent frees. This is the cause that most cleanly sep­ar­ates the pro­duc­tion minor­ity from the pi­lot­ing ma­jor­ity, and it is the sub­ject of chapter three.

What each stall looks like when it happens

A stalling pro­gramme an­nounces it­self in a spe­cific way, and each of the three causes car­ries its own sig­na­ture. Nam­ing them early is cheaper than dia­gnos­ing them late.

  • Gov­ernance stall. The agent is built and then waits. Ac­cess re­quests sit with a se­cur­ity team that has no policy for a non-human re­quester, and the spon­sor es­cal­ates rather than scopes. The tell is an agent still in a sand­box eight weeks after it worked.
  • Per­mis­sion stall, the same cause in­ver­ted. The agent ships with a stand­ing cre­den­tial and broad rights, be­cause that was the only way to un­block it. The tell is an agent whose reach nobody can de­scribe, and a se­cur­ity re­view de­ferred to a later phase.
  • In­teg­ra­tion stall. The agent works against test data and fails against the live land­scape. Con­nect­ors re­turn par­tial re­cords, iden­ti­fi­ers do not re­con­cile across sys­tems, and the error rate rises with volume. The tell is a sched­ule that slips in fort­nightly in­cre­ments with no single block­ing de­fect.
  • Read­i­ness stall. The agent works, the re­turn is meas­ured, and noth­ing changes. The freed hours are ab­sorbed rather than re­deployed, and no owner is named for the next pro­cess. The tell is a suc­cess­ful pilot with no second pro­cess be­hind it.

Each tell is vis­ible in a weekly re­view, and each has a cheap rem­edy if it is ap­plied early. A gov­ernance stall is pre­ven­ted by writ­ing the agent ac­cess policy be­fore the first agent rather than after the first re­quest. An in­teg­ra­tion stall is pre­ven­ted by test­ing one con­nector against pro­duc­tion-shaped data in the first fort­night. A read­i­ness stall is pre­ven­ted by nam­ing the second pro­cess be­fore the first one ships.

Where the value concentrates

Autonomy does not pay evenly. It pays where work is high in volume, bound by ex­pli­cit rules, di­git­ally nat­ive, and tol­er­ant of a meas­ur­able error rate. The early pro­duc­tion do­mains are con­sist­ent.

Ser­vice op­er­a­tions. Cus­tomer ser­vice and sup­port, where agents re­solve routine tick­ets end-to-end and es­cal­ate the ex­cep­tions.

IT op­er­a­tions. In­cid­ent triage, routine re­medi­ation, and ser­vice-desk auto­ma­tion, where work already flows through sys­tems such as ServiceNow and the ac­tions are well-defined.

Fin­ance op­er­a­tions. Re­con­cili­ation and the period-end close, ac­counts pay­able and ex­cep­tion hand­ling. The rules are ex­pli­cit and the audit trail is man­dat­ory re­gard­less.

Soft­ware en­gin­eer­ing. Code gen­er­a­tion and test­ing, then mi­gra­tion. The agent works in a di­git­ally nat­ive en­vir­on­ment, and its out­put is veri­fi­able on com­mit.

The com­mon thread is not the func­tion. It is the pro­cess sig­na­ture. That means volume and rule dens­ity, di­gital nat­ive­ness and veri­fi­ab­il­ity. A low-volume, judge­ment-heavy, or poorly in­stru­mented pro­cess is a weak early can­did­ate no mat­ter which func­tion it sits in. Screen the early port­fo­lio on those char­ac­ter­ist­ics, not on the org chart.

What the early domains actually involve

The pro­cess sig­na­ture picks the can­did­ate, and the work be­hind each early do­main is spe­cific enough to plan. Four pic­tures fol­low, each with the pre­requis­ite that most often blocks it.

Ser­vice op­er­a­tions. The agent reads the ticket, re­trieves the en­ti­tle­ment and the his­tory, ap­plies the res­ol­u­tion path, and writes the re­cord back. The hard part is not the lan­guage. It is that the res­ol­u­tion path is usu­ally un­doc­u­mented and lives with ex­per­i­enced staff, so the first fort­night goes on turn­ing tacit prac­tice into ex­pli­cit rules. The pre­requis­ite is a know­ledge base good enough to ground an an­swer, and the meas­ure is res­ol­u­tion at first con­tact without human cor­rec­tion.

Applied AIVoice of Volge04 •

IT op­er­a­tions. Triage, en­rich­ment, routine re­medi­ation and clos­ure in­side the ser­vice-man­age­ment tool. The work already flows through sys­tems, which is why this do­main reaches pro­duc­tion quick­est. The pre­requis­ite is a con­fig­ur­a­tion re­cord ac­cur­ate enough that the agent acts on the right asset. The char­ac­ter­istic fail­ure is a cor­rect ac­tion taken against the wrong host.

Fin­ance op­er­a­tions. Re­con­cili­ation, ex­cep­tion hand­ling in ac­counts pay­able, and the mech­an­ical parts of the period-end close. The rules are ex­pli­cit and the audit trail is man­dat­ory in any case, which makes the con­trol bur­den lower than it first ap­pears. The pre­requis­ite is a clean ledger-to-subledger map­ping, and the meas­ure is ex­cep­tions cleared per day with the error rate held flat.

Soft­ware en­gin­eer­ing. Gen­er­a­tion, test au­thor­ing, then mi­gra­tion at volume. Out­put is veri­fi­able on com­mit, which is the strongest veri­fi­ab­il­ity sig­nal avail­able any­where in the en­ter­prise. The pre­requis­ite is a test suite worth trust­ing, be­cause an agent that raises through­put against a weak suite raises de­fect es­cape with it.

Two things recur across all four. The work is already in­stru­mented, and the cor­rect an­swer is check­able without con­ven­ing a meet­ing. Where the second is miss­ing, the pro­cess is not an early can­did­ate how­ever at­tract­ive its volume, be­cause there is no cheap way to know whether the agent is right.

What this fixes downstream

The am­bi­tion is large, the pro­duc­tion rate is low, and the bind­ing con­straint is con­trol rather than cap­ab­il­ity. In­tent is close to uni­ver­sal: 91 per cent of senior AI ex­ec­ut­ives at lis­ted com­pan­ies above USD 1 bil­lion in rev­enue re­port using agen­tic AI through pi­lots or full de­ploy­ment, and 85 per cent of those say at least a hand­ful of these sys­tems already ex­ecute ac­tions without real-time human in­volve­ment (EY, 15 Septem­ber 2026). Over­sight has not fol­lowed. Forty-nine per cent of the same group say their gov­ernance frame­work has not yet been up­dated for agen­tic re­quire­ments and risks, and 26 per cent can­not de­tect un­au­thor­ised agents op­er­at­ing in­side their own es­tate. Thirty-six per cent have already had an AI in­cid­ent or fail­ure with a ma­ter­i­ally neg­at­ive im­pact. The con­straints are gov­ernance first, then in­teg­ra­tion and or­gan­isa­tional read­i­ness, and none is primar­ily tech­nical. The next two chapters re­tire them. Chapters two and three cover the con­trol sur­face and op­er­at­ing model, chapter four the ar­chi­tec­ture that pays down in­teg­ra­tion debt.

An as­sist­ant, or co­pi­lot, works in­side a single ap­plic­a­tion, re­sponds to an ex­pli­cit human prompt, re­turns a draft or

sug­ges­tion, and hands con­trol back. The human stays the actor. The ac­count­ab­il­ity bound­ary is un­am­bigu­ous.

, The control surface that decides the outcome

The control requirement rises far faster than the autonomy it permits, and that asymmetry is the single most consequential planning fact in this document. Moving a process up the spectrum means replacing human approval with engineered guardrails , a trade that is sound only when the guardrails, the identity controls, and the monitoring already exist. Failed projects are, overwhelmingly, processes pushed up the spectrum before the controls underneath them were ready.

Three points on a delegation continuum

Loose ter­min­o­logy drives loose gov­ernance, so the vocab­u­lary is fixed here. The three terms are not com­pet­ing products; they are points on a con­tinuum of del­eg­ated au­thor­ity, and an en­ter­prise will run all three at once.

An agent is handed an ob­ject­ive rather than a prompt. It plans a se­quence of steps, se­lects and calls tools, and ex­ecutes against sys­tems with lim­ited human in­ter­ven­tion. A per­son re­views the out­come, and some­times ap­proves con­sequen­tial steps, but does not dir­ect each ac­tion.

An autonom­ous agent does not wait for a prompt. An event or con­di­tion trig­gers it. It then runs con­tinu­ously in the back­ground, mon­it­or­ing data and act­ing in­side pre­defined guard­rails. Mi­crosoft's own guid­ance de­scribes autonom­ous agents as ex­tend­ing gen­er­at­ive or­ches­tra­tion through trig­gers and in­struc­tions set in­side guard­rails, op­er­at­ing con­tinu­ously to mon­itor data, react to con­di­tions and run work­flows at scale (Mi­crosoft Learn, 2026).

The move from the first to the third is the move from co­pi­lot to auto­pi­lot, and every step of it trans­fers au­thor­ity from a human to a con­trol set.

Read the right-hand column top to bot­tom. The con­trol re­quire­ment does not scale lin­early with autonomy; it steps up sharply at each trans­ition. Pro­mot­ing a pro­cess from agen­tic to autonom­ous is a de­cision to swap human-in-the-loop ap­proval for en­gin­eered guard­rails, and it holds only where those guard­rails, the iden­tity, and the mon­it­or­ing are demon­strably in place. Every pro­ject that skips this veri­fic­a­tion step is a can­did­ate for the can­cel­la­tion pool.

The same spec­trum, read as a lad­der, shows au­thor­ity trans­fer­ring from the human to the con­trol set at each step. The label on each rung names the con­trol that has to be in place be­fore the climb is safe.

Applied AIVoice of Volge05 •

The taxonomy, and the asymmetry that matters

StageWhat it doesHuman rolePrimary control requirement
Assistive (copilot)Drafts, summarises, suggests within one app on requestActor and approver of every outputContent review; data-access scoping
Agentic (supervised)Plans and executes multi-step tasks; calls toolsReviewer of outcomes; approver of consequential stepsTool-permission scoping; human-in-the-loop gates; audit logging
Autonomous (autopilot)Triggered by events; runs continuously within guardrailsSupervisor of a fleet; exception handlerAgent identity; least-privilege authority; continuous monitoring; kill-switch and revocation

The test before each promotion

A pro­cess should move up the spec­trum only when four con­di­tions hold, and each is veri­fi­able by in­spec­tion rather than by as­ser­tion. Treat the list as a gate.

ConditionWhat it meansHow it is tested
IdentityThe agent holds its own first-class identity and appears in the registryQuery the registry for the agent, then confirm that its sign-in record resolves to that identity alone
ScopeTool permissions match the task, and a sub-agent cannot exceed its parentAttempt a call outside scope and confirm that it is refused and logged
EvidenceThe agent's actions are logged in a form a reviewer readsReconstruct one week of the agent's work from logs alone, without asking the build team
ReversalThe agent's authority can be withdrawn across the landscape within minutesFire the revocation in a controlled window and confirm lost access, invalid tokens and a log entry

Con­di­tional pro­mo­tion is the dis­cip­line the can­celled pro­jects skip. Where one con­di­tion fails, the pro­cess stays on the rung below, and the fail­ing con­di­tion be­comes the next piece of work. That rule is un­pop­u­lar in a pro­gramme under de­liv­ery pres­sure, which is why it be­longs in the gov­ernance forum rather than in the build team's own judge­ment.

The Microsoft stack mapped to the spectrum

Mi­crosoft has as­sembled, across 2025 and 2026, a layered es­tate that sup­plies tool­ing at each point on the spec­trum. The de­lib­er­ate sep­ar­a­tion of lay­ers is the stra­tegic­ally use­ful part: each layer can be ad­op­ted, gov­erned, and re­placed on its own.

Mi­crosoft 365 Co­pi­lot , the as­sist­ive layer. Agen­tic ac­tion reached gen­eral avail­ab­il­ity in­side Word, Excel, and

PowerPoint on 22 April 2026, let­ting Co­pi­lot take multi-step ac­tions dir­ectly within doc­u­ments without step-by-step in­struc­tion (Mi­crosoft, 22 April 2026). PowerPoint shipped first on the web and was still rolling out across plat­forms at that date. This is the entry point for most en­ter­prises and the layer with the widest in­stalled base.

Co­pi­lot Stu­dio , the build and or­ches­tra­tion layer. Co­pi­lot Stu­dio has grown from a low-code con­ver­sa­tional-agent builder into a full agent plat­form for en­ter­prise auto­ma­tion, with model-driven or­ches­tra­tion, re­usable tools, and autonom­ous event-driven ex­e­cu­tion (Mi­crosoft, 2026). Two 2026 gen­eral-avail­ab­il­ity cap­ab­il­it­ies mat­ter. Com­puter-using agents drive web­sites and desktop ap­plic­a­tions dir­ectly through the in­ter­face, auto­mat­ing pro­cesses that pre­vi­ously de­pended on brittle scripts. Agent-to-agent com­mu­nic­a­tion lets agents ex­change in­form­a­tion, del­eg­ate tasks, and co­ordin­ate across sys­tems.

Mi­crosoft Agent Frame­work , the de­veloper layer. Ver­sion 1.0 shipped on 3 April 2026 for .NET and Py­thon, with Go in pre­view, mer­ging the AutoGen re­search line and the en­ter­prise-ready Se­mantic Ker­nel into one com­mer­cial-grade SDK and runtime. AutoGen has moved to main­ten­ance mode and Se­mantic Ker­nel con­tin­ues as a 1.x pack­age with crit­ical fixes only. It sup­ports multi-agent or­ches­tra­tion, the A2A and MCP pro­to­cols, and op­er­a­tion across .NET and Py­thon with any model pro­vider (Mi­crosoft, 2026).

Foundry , the model and tool layer. Foundry agents call MCP-com­pat­ible tools dir­ectly and im­ple­ment A2A through Se­mantic Ker­nel, en­abling col­lab­or­a­tion across runtimes. Foundry ships with thou­sands of con­nect­ors into SaaS and en­ter­prise sys­tems, cut­ting the in­teg­ra­tion re­build that oth­er­wise con­sumes an agent pro­gramme (Mi­crosoft, 2026).

Agent 365 , the con­trol plane. An­nounced at Ig­nite in Novem­ber 2025 and gen­er­ally avail­able since 1 May 2026, Agent 365 gov­erns AI agents across the en­ter­prise re­gard­less of where they were built, what model they run, or which plat­form hosts them (Mi­crosoft, 1 May 2026). It is li­censed per user, not per agent, which is the com­mer­cial fact a reader siz­ing an agent es­tate most needs. Stan­dalone it costs 15 United States dol­lars per user per month and re­quires a pre­requis­ite suite of Mi­crosoft 365 E5, A5, or Busi­ness Premium; in­side Mi­crosoft 365 E7 it is in­cluded, at 99 dol­lars per user per month. Since 1 July 2026 the De­fender agent-pro­tec­tion cap­ab­il­it­ies, free in pre­view, also re­quire Agent 365 li­cens­ing. Its com­pon­ents are the gov­ernance es­tate, dis­sec­ted in chapter three.

Applied AIVoice of Volge06 •

The sep­ar­a­tion is it­self a hedge against lock-in. An en­ter­prise can adopt the as­sist­ive layer without com­mit­ting to the build layer, and can gov­ern third-party and in-house agents through Agent 365 without those agents touch­ing a Mi­crosoft model. The ad­op­tion se­quence in chapter four ex­ploits ex­actly that.

Prerequisites, licences and what is still in preview

The con­trol set car­ries li­cence pre­requis­ites, and a pro­gramme should con­firm them be­fore com­mit­ting to a se­quence rather than at con­fig­ur­a­tion. Mi­crosoft sets them out dir­ectly.

CapabilityWhat is required
Creating and managing agent identitiesAvailable to all Microsoft Entra customers
Operating agents across Microsoft 365 and enterprise workflowsA Microsoft Agent 365 licence per user
Conditional Access for agentsMicrosoft 365 E7, or Agent 365 paired with Entra ID P1 or Microsoft 365 E3
Risk-based Conditional Access for agentsAs above with Entra ID P2 or Microsoft 365 E5; the agent risk condition remains in preview
Identity Protection for agentsAgent 365 or Microsoft 365 E7
Identity governance for agentsMicrosoft Entra ID P1

Source: Mi­crosoft Learn, Con­di­tional Ac­cess for agents and the Agent 365 li­cens­ing guid­ance, 2026. Two con­sequences fol­low. An en­ter­prise already li­cens­ing Agent 365, or sit­ting on E7, owns the en­ti­tle­ment, and its gap is con­fig­ur­a­tion rather than pro­cure­ment. An E5 ten­ant without Agent 365 has a com­mer­cial line to close first, and the li­cence is coun­ted per user who man­ages or spon­sors agents, or whose work the agents do, rather than per agent, so the cost scales with the gov­ern­ing pop­u­la­tion rather than with the fleet.

One con­trol rests on a pre­view cap­ab­il­ity, and a design doc­u­ment should say so. The agent risk con­di­tion be­hind risk-based block­ing was still in pre­view as at Novem­ber 2026. A pos­ture that de­pends on it needs a stated fall­back, usu­ally the ap­proval-at­trib­ute policy, so the land­scape is not left un­guarded if the pre­view be­ha­viour changes.

The competitors, and the interoperability fact

A board should know which al­tern­at­ives it is im­pli­citly de­clin­ing. The plat­forms are com­pared in full in Ap­pendix A; in brief:

Sales­force Agent­force has been gen­er­ally avail­able since 2024 and is strongest in­side the Sales­force eco­sys­tem.

En­ter­prises with sig­ni­fic­ant non-Sales­force work­flows still have to de­cide how agents reach across the CRM bound­ary.

ServiceNow AI Agents have been gen­er­ally avail­able since 2024. They are strongest where work already flows through the plat­form, in IT ser­vice man­age­ment and se­cur­ity op­er­a­tions, and in human re­sources and en­ter­prise work­flow.

Amazon Bed­rock AgentCore shipped in 2025 as the bring-your-own-frame­work op­tion. It ac­cepts any frame­work, model or pro­tocol, and wraps them in AWS isol­a­tion, vir­tual-private-cloud net­work­ing and in­fra­struc­ture con­trols.

Google Gem­ini En­ter­prise shipped in 2025 with part­ner agents from Box, Work­day, Sales­force, ServiceNow, and oth­ers in­teg­rated into the plat­form.

The de­cis­ive fact is in­ter­op­er­ab­il­ity. The A2A pro­tocol, con­trib­uted by Google and hos­ted by the Linux Found­a­tion since 2025, passed 150 sup­port­ing or­gan­isa­tions at its one-year mark, among them Amazon Web Ser­vices, Cisco, Google, IBM, Mi­crosoft, Sales­force, SAP and ServiceNow, with en­ter­prise pro­duc­tion use re­por­ted in sup­ply chain and fin­an­cial ser­vices, in in­sur­ance and in IT op­er­a­tions (Linux Found­a­tion, 9 April 2026). The prac­tical con­sequence: the agen­tic en­ter­prise is multi-vendor by de­fault. An agent built in Co­pi­lot Stu­dio will del­eg­ate to an agent run­ning in ServiceNow, and the gov­ernance es­tate has to span both. That ar­chi­tec­tural real­ity drives chapters three and four.

Who answers when one agent delegates to another

Multi-vendor del­eg­a­tion is the ar­chi­tec­tural de­fault, and it breaks the ac­count­ab­il­ity model un­less the design ad­dresses it. When an agent built in Co­pi­lot Stu­dio hands work to an agent run­ning on an­other vendor's plat­form, two things have to sur­vive the han­doff. The iden­tity of the ori­ginal re­quester is the first, and the limit on what may be done on that per­son's be­half is the second.

Neither sur­vives auto­mat­ic­ally. The A2A pro­tocol stand­ard­ises the com­mu­nic­a­tion rather than the au­thor­isa­tion, so the en­ter­prise de­cides what crosses the bound­ary. Three de­cisions be­long in the design. Whether a del­eg­ated call car­ries the user's con­text or the agent's own iden­tity, since the reach dif­fers sharply between the two. Whether the re­ceiv­ing agent may del­eg­ate on­ward, which is where a nar­row grant be­comes a broad one. And which side holds the audit re­cord, be­cause an ac­tion taken on a part­ner plat­form still has to be re­con­struct­able from the en­ter­prise's own logs.

Applied AIVoice of Volge07 •

Where the an­swer to the third ques­tion is the other party, the ar­range­ment is a third-party de­pend­ency and should be con­trac­ted as one. SEBI's frame­work makes reg­u­lated en­tit­ies solely ac­count­able for the se­cur­ity of their data and logs in third-party ser­vices, and the same prin­ciple runs through every out­sourcing re­gime in the re­gion. Del­eg­at­ing the work does not del­eg­ate the ac­count­ab­il­ity.

What this constrains

Sup­ply is no longer the con­straint. Tool­ing ex­ists at every point on the spec­trum, from the lead­ing plat­form and its com­pet­it­ors, con­nec­ted by pro­to­cols already in pro­duc­tion. The lead­er­ship ques­tion is no longer "can we" but "how far, how fast, and under what con­trols", and the an­swer is gated by the con­trol sur­face above, not by the model. The next chapter takes up the or­gan­isa­tional half of that con­trol sur­face, which is where most pro­grammes ac­tu­ally break.

, The operating model as the binding constraint

Agents have to be run as a workforce, not deployed as a feature. That needs four things: an accountable owner, an operations function , AgentOps , a human-agent teaming model, and an identity-led governance estate. Naming a dedicated owner is where the accountability an autonomous agent generates has to land; assigning a first-class identity to every agent is the precondition for governing the estate at all. This is the organisational readiness whose absence chapter one identified as the third cause of stall.

Identity first, because nothing governs without it

An agent that lacks a first-class iden­tity can­not be in­vent­or­ied or scoped, and it can­not be audited or re­voked. Iden­tity is not one con­trol among sev­eral. It is the floor the rest stand on. The scale of the prob­lem is set by pro­lif­er­a­tion, and the vis­ib­il­ity gap it cre­ates is already meas­ured. Twenty-six per cent of senior AI ex­ec­ut­ives whose or­gan­isa­tions use agen­tic AI say they can­not de­tect un­au­thor­ised agents op­er­at­ing in­tern­ally (EY, 15 Septem­ber 2026). An un­gov­erned es­tate of that kind is not a se­cur­ity gap. It is a struc­tural in­ab­il­ity to an­swer who did what, on whose au­thor­ity.

The prin­ciple now co­di­fied across the in­dustry is that no agent is trus­ted by de­fault, re­gard­less of pur­pose or claimed cap­ab­il­ity; trust is earned through demon­strated be­ha­viour and con­tinu­ously re-veri­fied through mon­it­or­ing (Cloud Se­cur­ity Al­li­ance, 2 Feb­ru­ary 2026; stew­ard­ship of the frame­work passed to MassiveScale.AI in May 2026, and the ca­non­ical spe­cific­a­tion now sits out­side CSA). This is Zero Trust

ex­ten­ded to autonom­ous sys­tems. It de­mands a cryp­to­graphic iden­tity per agent, con­tinu­ous au­then­tic­a­tion across the li­fe­cycle, least-priv­ilege ac­cess on every tool call, and out­put val­id­a­tion be­fore any­thing propag­ates.

Mi­crosoft has built this con­trol set in, and it is gated. Entra Agent ID as­signs each agent a first-class iden­tity, ex­tend­ing Zero Trust to AI work­loads (Mi­crosoft, 2025). Mes­sage Cen­ter no­tice MC1395007 of 17 June 2026 added the Entra Con­di­tional Ac­cess for Agents and Entra ID Pro­tec­tion for Agents ser­vice plans to Mi­crosoft Agent 365 and Mi­crosoft 365 E7, with world­wide rol­lout from early July to early Au­gust 2026, and the no­tice is ex­pli­cit that a cus­tomer using either con­trol without Agent 365 or E7 li­cens­ing needs the ap­pro­pri­ate li­cence to con­tinue. Mi­crosoft 365 E5 does not carry them. Agent 365 provides an agent re­gistry, an in­vent­ory of every agent, in-house or third-party, to­gether with ac­cess con­trol. It maps the re­la­tion­ships between agents, people and data, and it in­teg­rates with Mi­crosoft De­fender and Pur­view. The Se­cur­ity Dash­board for AI ag­greg­ates sig­nals from De­fender, Entra and Pur­view into a single pos­ture view across the AI es­tate (Mi­crosoft, 2026).

The es­tate also has to line up with ex­ternal frame­works, be­cause aud­it­ors and reg­u­lat­ors in­creas­ingly ex­pect it: the OWASP Top 10 for Agen­tic Ap­plic­a­tions (Decem­ber 2025), the Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work (Feb­ru­ary 2026, now stew­ar­ded out­side CSA), and the Co­ali­tion for Se­cure AI (in­dustry sources, 2025-2026). These are cata­logued in Ap­pendix B.

Agents as digital labour

The gov­ern­ing meta­phor is not rhet­or­ical. An autonom­ous agent is a unit of di­gital la­bour, and that fram­ing is the most re­li­able guide to the con­trols it needs, be­cause the dis­cip­lines a ma­ture en­ter­prise already ap­plies to its human work­force map dir­ectly onto the con­trols an agent es­tate re­quires.

A human em­ployee is hired against a defined role, gran­ted an iden­tity, given scoped sys­tem ac­cess, su­per­vised, per­form­ance-man­aged, and off­boarded with ac­cess re­voked. An agent needs each in turn: a defined task scope, a first-class iden­tity, least-priv­ilege ac­cess, con­tinu­ous mon­it­or­ing, eval­u­ation against per­form­ance met­rics, and clean de­com­mis­sion­ing with cre­den­tial re­voc­a­tion. The en­ter­prises that suc­ceed reuse their work­force-gov­ernance dis­cip­lines rather than in­vent­ing a par­al­lel, weaker con­trol set be­side them.

Applied AIVoice of Volge08 •

The meta­phor also fixes the eco­nom­ics. The re­turn on an agent is not real­ised by the agent. It is real­ised when the la­bour the agent frees is re­deployed to higher-value work. A pro­gramme that cap­tures the cost sav­ing but ig­nores re­deploy­ment for­goes the lar­ger pro­ductiv­ity gain and breeds the or­gan­isa­tional res­ist­ance that sinks the pro­gramme.

The owner, and why accountability has to be named

Only 21 per cent of the com­pan­ies plan­ning agent de­ploy­ments within two years re­port a ma­ture model for agent gov­ernance (De­loitte, 21 Janu­ary 2026). The reason a named owner mat­ters is struc­tural. An agent act­ing autonom­ously against live sys­tems gen­er­ates ac­count­ab­il­ity that has to land on a named per­son. Without one, nobody is an­swer­able for the agent's ac­tions, nobody is em­powered to au­thor­ise its ac­cess, and nobody owns its de­com­mis­sion­ing. A miss­ing owner is not a hole in the org chart; it is the ab­sence of the ac­count­ab­il­ity that gov­ernance is made of.

The owner sits at the in­ter­sec­tion of the busi­ness func­tion the agent serves, the tech­no­logy func­tion that builds it, and the se­cur­ity func­tion that gov­erns it. Volge's view is that the owner be­longs in the busi­ness func­tion, with dot­ted-line ac­count­ab­il­ity to tech­no­logy and se­cur­ity, be­cause the busi­ness owns the out­come and car­ries the risk. Lodging own­er­ship in tech­no­logy alone re­pro­duces the fail­ure pat­tern in which agents are gov­erned as soft­ware rather than as la­bour.

What the owner actually decides

Nam­ing an owner is worth noth­ing until the de­cisions the role holds are writ­ten down. Five de­cisions be­long to the agent owner and to nobody else, and each is the point at which an un­gov­erned pro­gramme im­pro­vises.

  • Au­thor­isa­tion. Whether an agent may hold a given tool per­mis­sion, on the re­cord, with the busi­ness ra­tionale. Se­cur­ity ad­vises and the owner de­cides, which is one reason the role can­not sit in­side se­cur­ity.
  • Pro­mo­tion. Whether a pro­cess moves up the autonomy spec­trum, judged against the four con­di­tions in chapter two rather than against de­liv­ery pres­sure.
  • Team­ing. Where the human stays in the loop for this pro­cess, and the trig­ger that would move the pat­tern back to­wards ap­proval.
  • Tol­er­ance. The error rate the pro­cess will ac­cept, stated as a num­ber be­fore the agent runs, and the level at which the agent is sus­pen­ded.
  • De­com­mis­sion­ing. When the agent stops, who con­firms that the cre­den­tial was with­drawn, and what be­comes of the work it was doing.

Write those five into the role de­scrip­tion and into the gov­ernance forum's terms of ref­er­ence. A pro­gramme that can­not say who holds the tol­er­ance de­cision has not named an owner. It has named a spon­sor.

AgentOps , running the fleet

AgentOps is the op­er­a­tions func­tion for agents. It in­teg­rates ob­serv­ab­il­ity and eval­u­ation, gov­ernance and feed­back, as well as se­cur­ity, re­si­li­ence and ver­sion­ing across the agent li­fe­cycle (in­dustry defin­i­tion, 2026). It turns an autonom­ous sys­tem from a re­search ex­per­i­ment into a re­li­able and ac­count­able en­ter­prise layer that scales. Four cap­ab­il­it­ies define a cred­ible one.

Ob­serv­ab­il­ity. Every au­then­tic­a­tion event and tool in­voc­a­tion is cap­tured, as is every del­eg­a­tion han­doff and policy de­cision, in a form that sup­ports real-time mon­it­or­ing and audit. The re­cord shows not only what happened, but why, on whose be­half and under what policy (in­dustry guid­ance, 2026). The emer­ging stand­ard is OpenTelemetry ex­ten­ded for AI, track­ing agent per­form­ance and sys­tem health, as well as risk across com­plex en­vir­on­ments.

Eval­u­ation. Test­ing runs in every phase of de­ploy­ment, against pre­defined scen­arios and met­rics for re­li­ab­il­ity and fair­ness as well as se­cur­ity. An agent, un­like con­ven­tional soft­ware, be­haves non-de­term­in­ist­ic­ally. It needs con­tinu­ous eval­u­ation rather than one-time ac­cept­ance test­ing.

Policy en­force­ment. A con­trol plane en­forces policy on data scope and high-risk ac­tions in real time, hold­ing agents in­side defined ser­vice-level ob­ject­ives. This is where the gov­ernance es­tate meets the op­er­a­tions func­tion.

Re­si­li­ence and ver­sion­ing. Agents need defined roll­back, ver­sion con­trol, and a kill-switch. The ca­pa­city to re­voke an agent's au­thor­ity im­me­di­ately, across the whole es­tate, is the op­er­a­tional coun­ter­part to the iden­tity con­trols in 4.1.

What AgentOps measures

An agent fleet needs its own in­stru­ment panel, be­cause the meas­ures that gov­ern soft­ware do not de­scribe di­gital la­bour. Six carry most of the weight, and a pro­gramme that can­not pro­duce them is not yet run­ning a fleet.

  • Task com­ple­tion without human cor­rec­tion, by pro­cess. The head­line pro­ductiv­ity meas­ure, and the first to fall when an up­stream sys­tem changes.
Applied AIVoice of Volge09 •
  • Es­cal­a­tion rate and es­cal­a­tion ac­cur­acy. A fleet that es­cal­ates too little is un­safe, and one that es­cal­ates too much has moved the work rather than re­moved it.
  • Scope re­fus­als per thou­sand tool calls. A re­fusal is a con­trol work­ing. A sud­den fall usu­ally means a per­mis­sion was widened, not that be­ha­viour im­proved.
  • At­tri­bu­tion com­plete­ness. The share of agent ac­tions re­con­struct­able from logs alone, meas­ured by sampling rather than in­ferred from con­fig­ur­a­tion.
  • Time to re­voke, meas­ured in a drill. The in­ter­val between the de­cision to cut an agent's au­thor­ity and the con­firmed loss of ac­cess.
  • Re­deploy­ment of freed ca­pa­city, by func­tion. The meas­ure that de­cides whether the sav­ing reaches the in­come state­ment or dis­sip­ates.

Re­port these monthly to the agent owner and quarterly to the risk com­mit­tee. Two of them, at­tri­bu­tion com­plete­ness and time to re­voke, are what an aud­itor asks for, and both de­grade quietly as the fleet grows.

Staffing AgentOps, and the operating rhythm

AgentOps is a small func­tion with a fixed rhythm, not a new de­part­ment. In most en­ter­prises it starts at three or four people drawn from ex­ist­ing teams, and it grows with the fleet rather than ahead of it.

  • An op­er­a­tions lead, ac­count­able for the fleet's re­li­ab­il­ity and for the monthly re­port to the agent owner.
  • A plat­form en­gin­eer, hold­ing the re­gistry, the policy con­fig­ur­a­tion and the con­nector cata­logue.
  • An eval­u­ation en­gin­eer, hold­ing the scen­ario set and re-run­ning it on every model or prompt change.
  • Named pro­cess con­tacts in each busi­ness func­tion, part-time, own­ing the ex­cep­tion queue and the error tol­er­ance for their own pro­cess.

The rhythm is weekly, monthly and quarterly. Weekly brings the ex­cep­tion queue and the error rate per pro­cess, with any agent out­side tol­er­ance sus­pen­ded rather than de­bated. Monthly brings the six meas­ures above to the agent owner, to­gether with the re­gister of agents dis­covered out­side the re­gistry. Quarterly brings a re­voc­a­tion drill, a per­mis­sion re­view per agent, and a re-run of the eval­u­ation set against the cur­rent model. A fleet without the quarterly ca­dence de­grades in­vis­ibly, be­cause noth­ing fails loudly when an agent's per­mis­sions have quietly widened.

What done looks like

A pro­cess is fin­ished, rather than pi­loted, when five state­ments hold without qual­i­fic­a­tion. Use them as the ac­cept­ance cri­teria for a stage, in place of a read­i­ness opin­ion.

  • A named per­son in the busi­ness func­tion is ac­count­able for the agent, and the name sits in the re­gistry rather than in a slide.
  • The agent's au­thor­ity is doc­u­mented as a list of tool per­mis­sions, and the list was re­viewed in the last quarter.
  • A re­viewer has re­con­struc­ted a week of the agent's work from logs alone.
  • Re­voc­a­tion was ex­er­cised against the live agent, and the drill is dated.
  • The pro­cess owner has re­por­ted a meas­ured out­come against a baseline taken be­fore the agent ex­is­ted.

The teaming model, where the human stays

From the worker's seat, the move from co­pi­lot to auto­pi­lot is a move from doing the work to su­per­vising it. The team­ing model defines where the human stays in the loop, and it has to be de­signed on pur­pose rather than left to emerge.

Three pat­terns ap­pear in pro­duc­tion. Human-in-the-loop means the agent pro­poses and the human ap­proves each con­sequen­tial ac­tion, which is right for high-risk, low-volume work. Human-on-the-loop means the agent acts within guard­rails while the human su­per­vises a fleet and handles ex­cep­tions, the tar­get state for high-volume autonom­ous work. Human-out-of-the-loop means the agent acts without routine human in­volve­ment, de­fens­ible only for low-risk, re­vers­ible and well-in­stru­mented work.

The error to avoid is one pat­tern across all pro­cesses. Se­lect per pro­cess, on risk and volume and re­vers­ib­il­ity, then mi­grate the pat­tern as con­fid­ence and con­trols ma­ture. Reg­u­la­tion con­strains the choice dir­ectly. The European Union AI Act, Art­icle 14(1), re­quires high-risk AI sys­tems to be de­signed for ef­fect­ive over­sight by nat­ural per­sons (European Union, 2024), which fore­closes the human-out-of-the-loop pat­tern for any­thing in its scope.

Designing the teaming model, rather than inheriting it

The team­ing pat­tern is a design choice per pro­cess, and three in­puts settle it. Re­vers­ib­il­ity comes first, since an ac­tion that can be un­done cheaply tol­er­ates less su­per­vi­sion than one that can­not. Volume comes next, be­cause a gate re­ceiv­ing

Applied AIVoice of Volge10 •

more de­cisions than a per­son can weigh is not a con­trol. The cost of being wrong comes last, and it is rarely sym­met­rical, since a false re­fusal and a false ac­tion usu­ally fall on dif­fer­ent parties.

Two rules fol­low in prac­tice. Put the human where the judge­ment is rather than where the work­flow is con­veni­ent, which often means ap­prov­ing the ex­cep­tion and not the routine case. And write the mi­gra­tion trig­ger at the same time as the pat­tern, so that mov­ing from ap­proval to su­per­vi­sion is a planned event with evid­ence be­hind it rather than a quiet drift.

The regulatory perimeter

Two re­gimes are first-order for en­ter­prises op­er­at­ing in or serving the European Union and India. Both are sum­mar­ised here and de­tailed in Ap­pendix C.

The European Union AI Act. The de­fer­ral is now law. Reg­u­la­tion (EU) 2026/1744, the Di­gital Om­ni­bus on AI, was ad­op­ted by the Coun­cil on 29 June 2026, pub­lished in the Of­fi­cial Journal on 24 July 2026, and entered into force on 27 July 2026, six days be­fore the dead­line it moved. The high-risk ob­lig­a­tions under Annex III now fall due on 2 Decem­ber 2027 rather than 2 Au­gust 2026, and the Annex I ob­lig­a­tions on 2 Au­gust 2028 rather than 2 Au­gust 2027 (European Union, 2026). One ob­lig­a­tion moved the other way on the cal­en­dar and falls due first. Art­icle 50(2), on ma­chine-read­able mark­ing and the dis­clos­ure of syn­thetic con­tent, ap­plies from 2 Decem­ber 2026 rather than 2 Au­gust 2026. The post­pone­ment buys plan­ning time; it does not soften the sub­stant­ive ob­lig­a­tions, prin­cip­ally the Art­icle 14 re­quire­ment for ef­fect­ive human over­sight of high-risk sys­tems. Read the de­fer­ral as time to pre­pare, not as a lowered bar.

India's Di­gital Per­sonal Data Pro­tec­tion re­gime. The Min­istry of Elec­tron­ics and In­form­a­tion Tech­no­logy no­ti­fied the Di­gital Per­sonal Data Pro­tec­tion Rules, 2025, op­er­a­tion­al­ising the Act of 2023 (MeitY, 2025). The Rules re­quire itemised no­tice, pur­pose-based re­ten­tion, reas­on­able se­cur­ity safe­guards, breach no­ti­fic­a­tion within 72 hours, and veri­fi­able par­ental con­sent for chil­dren. Sig­ni­fic­ant Data Fi­du­ciar­ies carry en­hanced ob­lig­a­tions: an­nual data-pro­tec­tion im­pact as­sess­ments, audits, and al­gorithmic-fair­ness as­sess­ments. The de­cis­ive point for agen­tic sys­tems is that the or­gan­isa­tion stays ac­count­able as a data fi­du­ciary even when a de­cision is auto­mated, which makes data-pro­tec­tion com­pli­ance a design re­quire­ment, not a post-de­ploy­ment patch.

What this resolves

The op­er­at­ing model is the dif­fer­en­ti­ator between the pro­duc­tion minor­ity and the pi­lot­ing ma­jor­ity. The four dis­cip­lines , an ac­count­able owner, an AgentOps func­tion, a de­lib­er­ate team­ing model, and an iden­tity-led gov­ernance es­tate , to­gether con­sti­tute the or­gan­isa­tional read­i­ness whose ab­sence chapter one named as the third cause of stall. The ar­chi­tec­ture that re­tires the second cause, in­teg­ra­tion debt, is chapter four.

, Reference architecture and what actually ships

A durable agent architecture is layered and protocol-based, never monolithic. Separate the model layer, the tool and data layer, the orchestration layer, the governance layer, and the experience layer, and connect them through open standards , principally MCP and A2A , so the estate spans vendors without re-engineering. What ships from that architecture is staged: start at the assistive layer, prove value in one or two characteristic-matched processes, and climb the autonomy spectrum only as the governance estate earns it. Autonomy never outruns control.

Five layers, each sourced and governed independently

The ar­chi­tec­ture below is anchored to Mi­crosoft, con­sist­ent with the paper's lens, but the lay­er­ing is port­able. The point of stat­ing it as lay­ers is that each can be sourced, gov­erned, and re­placed on its own , the struc­tural hedge against lock-in.

The model layer. Found­a­tion and reas­on­ing mod­els, reached through Foundry, which sup­ports mul­tiple pro­viders rather than a single house model. The prin­ciple is model in­de­pend­ence: or­ches­tra­tion and gov­ernance stay de­coupled from any spe­cific model, so mod­els can be swapped as price and cap­ab­il­ity move.

The tool and data layer. The con­nect­ors, APIs, and data sources the agent acts upon, ex­posed through the Model Con­text Pro­tocol (MCP). MCP is the op­er­a­tional bridge between high-level or­ches­tra­tion and low-level tool ex­e­cu­tion, con­vert­ing planned ob­ject­ives into struc­tured, policy-aligned in­voc­a­tions and feed­ing ex­e­cu­tion data back into or­ches­tra­tion memory (in­dustry sources, 2026). Foundry's con­nector lib­rary is the primary lever against the in­teg­ra­tion debt chapter one named as the second cause of stall.

The or­ches­tra­tion layer. The runtime that plans, se­quences, and co­ordin­ates the work is the Mi­crosoft Agent Frame­work and Co­pi­lot Stu­dio. It de­com­poses a goal into steps and se­lects tools. Through the Agent-to-Agent (A2A) pro­tocol it

Applied AIVoice of Volge11 •

del­eg­ates to and co­ordin­ates with other agents, in­clud­ing agents on other vendors' plat­forms.

The gov­ernance layer. The iden­tity, policy, and ob­serv­ab­il­ity con­trol plane , Entra Agent ID and Agent 365, per chapter three. The prin­ciple is that gov­ernance spans the es­tate as a layer, not as a fea­ture baked into each agent. That is what lets one con­trol plane gov­ern agents built in-house, by third parties, and on com­pet­ing plat­forms.

The ex­per­i­ence layer. The sur­faces through which people in­ter­act with and su­per­vise agents: Mi­crosoft 365 Co­pi­lot, Teams, and line-of-busi­ness ap­plic­a­tions. This is where the team­ing model of chapter three is op­er­a­tion­al­ised.

The five lay­ers stack as below. The gov­ernance layer is drawn as a con­trol sur­face span­ning the oth­ers, be­cause it gov­erns the es­tate as a layer rather than as a fea­ture in­side each agent.

The protocols are the architecture

The two pro­to­cols de­cide whether the agen­tic en­ter­prise is an open es­tate or a set of vendor silos, which makes them worth a dir­ect state­ment.

Model Con­text Pro­tocol (MCP) stand­ard­ises how an agent con­nects to tools and data. It lets an agent act against any com­pli­ant sys­tem without be­spoke in­teg­ra­tion, and it is the prin­cipal lever against in­teg­ra­tion debt.

Agent-to-Agent (A2A) stand­ard­ises how agents com­mu­nic­ate , ne­go­ti­ation, del­eg­a­tion, and co­ordin­a­tion across dis­trib­uted eco­sys­tems while pre­serving in­ter­op­er­ab­il­ity, trace­ab­il­ity, and se­cur­ity (in­dustry sources, 2026). A2A is a Linux Found­a­tion pro­ject, con­trib­uted by Google in 2025, and it passed 150 sup­port­ing or­gan­isa­tions and its first stable 1.0 spe­cific­a­tion at the one-year mark, with en­ter­prise pro­duc­tion use re­por­ted in sup­ply chain and fin­an­cial ser­vices, in in­sur­ance and in IT op­er­a­tions (Linux Found­a­tion, 9 April 2026).

Build to these pro­to­cols, not to a single vendor's pro­pri­et­ary in­ter­faces. Doing so pre­serves the op­tion of a multi-vendor es­tate , the de­fault end-state es­tab­lished in chapter two , and pro­tects the in­teg­ra­tion in­vest­ment against plat­form change.

Governing the tool layer

The tool layer is where an agent's au­thor­ity ac­tu­ally lives, and it is the layer most often left un­gov­erned. An agent's ef­fect­ive power is the union of the tools it can in­voke, so a tool cata­logue nobody owns is a per­mis­sion sys­tem nobody con­trols.

That last rule is the one en­gin­eer­ing teams un­der­es­tim­ate. OWASP's Top 10 for Agen­tic Ap­plic­a­tions, pub­lished 9 Decem­ber 2025, names in­sec­ure tool ex­e­cu­tion and ex­cess­ive agency among its risks, and the mech­an­ism be­hind both

is the same. The plan­ner acts on text, and text ar­riv­ing from a doc­u­ment, a web page or an­other sys­tem is in­dis­tin­guish­able from an in­struc­tion un­less the ar­chi­tec­ture makes it oth­er­wise. Val­id­ate tool out­put be­fore it propag­ates, con­strain what a tool may trig­ger without a fur­ther check, and log the con­tent that caused a con­sequen­tial ac­tion rather than only the ac­tion.

Four rules make the layer gov­ern­able. Every tool car­ries a named owner and a stated pur­pose, re­cor­ded in a cata­logue rather than in­side an agent's con­fig­ur­a­tion. Every tool server is ap­proved be­fore use, be­cause a tool reached through the Model Con­text Pro­tocol is code the en­ter­prise now trusts, whether it was writ­ten in-house or not. Every tool de­clares the data it can read and the ac­tions it can take, in terms a re­viewer un­der­stands rather than in an in­ter­face spe­cific­a­tion. And tool out­put is treated as un­trus­ted input, be­cause con­tent re­turned by a tool can carry in­struc­tions the agent will fol­low.

Op­er­a­tion­ally, keep the cata­logue short. A dozen tools serve the first sev­eral agents in most en­ter­prises, and the gov­ernance cost rises with the cata­logue rather than with the fleet. Adding a tool should be a de­cision with an owner be­hind it, not a side ef­fect of a sprint.

Compose by default, buy at the edges, build rarely

The ar­chi­tec­ture turns make-or-buy into a build-buy-com­pose de­cision.

Buy a pre-built agent where a vendor of­fers one matched to a stand­ard pro­cess , a ser­vice agent from Sales­force, an IT-op­er­a­tions agent from ServiceNow , and where the pro­cess does not dif­fer­en­ti­ate the en­ter­prise. Fast­est re­turn, low­est gov­ernance bur­den.

Com­pose an agent from ex­ist­ing tools and con­nect­ors in Co­pi­lot Stu­dio or the Agent Frame­work where the pro­cess is spe­cific to the en­ter­prise but built on stand­ard sys­tems. This is the modal choice for most en­ter­prise pro­cesses and the one the Mi­crosoft es­tate is op­tim­ised for.

Build a be­spoke agent only where the pro­cess is both dif­fer­en­ti­at­ing and un­sup­por­ted by avail­able tool­ing. Set the bar high; be­spoke agents carry the full gov­ernance and main­ten­ance load.

De­fault to com­pose, re­serve buy for com­mod­it­ised pro­cesses, and build only for genu­ine dif­fer­en­ti­ation. That se­quen­cing min­im­ises both time-to-value and the gov­ernance sur­face.

The last ques­tion dis­qual­i­fies most be­spoke pro­pos­als. A be­spoke agent car­ries the full gov­ernance and main­ten­ance

Applied AIVoice of Volge12 •

load, in­clud­ing re-eval­u­ation every time the un­der­ly­ing model changes, and the model changes on the vendor's sched­ule rather than the en­ter­prise's.

Testing the build, buy and compose decision

The de­cision is usu­ally taken on in­stinct and de­fen­ded af­ter­wards. Four ques­tions settle it in one ses­sion, and the an­swers are mat­ters of fact rather than of opin­ion.

  • Does the pro­cess dif­fer­en­ti­ate the en­ter­prise in the eyes of a cus­tomer or a reg­u­lator. If not, buy.
  • Does a vendor already ship an agent for it, in­side a plat­form the en­ter­prise already gov­erns. If so, buy, and spend the saved ef­fort on the gov­ernance wir­ing.
  • Do the sys­tems the pro­cess touches ex­pose us­able in­ter­faces, through MCP or a sup­por­ted con­nector. If so, com­pose.
  • Is the en­ter­prise will­ing to own eval­u­ation, ver­sion­ing and model change for the life of the pro­cess. If not, do not build.

Paying down integration debt, specifically

In­teg­ra­tion debt is the cause most often missed at the busi­ness-case stage, and it is also the most tract­able, be­cause the work is enu­mer­able. Four passes re­tire most of it, and each leaves an arte­fact the next agent re­uses.

  • In­vent­ory the in­ter­faces the can­did­ate pro­cess touches, with the owner and the fail­ure his­tory of each. Brittle in­ter­faces an­nounce them­selves in the in­cid­ent re­cord.
  • Re­con­cile iden­ti­fi­ers across the sys­tems in­volved. Where the same cus­tomer or asset car­ries dif­fer­ent keys, the agent will act on the wrong re­cord at volume, and no amount of prompt design re­pairs it.
  • Ex­pose the tools through a stand­ard in­ter­face rather than through be­spoke code, so the next agent in­her­its them. Foundry's con­nector lib­rary and MCP exist for this, and the con­nector is where the reuse ac­crues.
  • Test against pro­duc­tion-shaped data be­fore the build is ac­cep­ted, mal­formed re­cords in­cluded. Most sched­ule slip­page is this pass being skipped and then re­peated under pres­sure.

The arte­facts from those passes are dur­able as­sets. They are the in­ter­face re­gister, the iden­ti­fier map, the tool cata­logue and the test cor­pus. A second agent against the same sys­tems should cost a frac­tion of the first, and where it does not, the passes were not fin­ished.

The roadmap , governance ahead of autonomy

The ca­dence below is de­lib­er­ately con­ser­vat­ive on autonomy and de­lib­er­ately ag­gress­ive on gov­ernance , the in­verse of the failed-pro­ject pat­tern, in which gov­ernance lags and autonomy races ahead. The stages map to the autonomy spec­trum of chapter two and the ma­tur­ity model in 5.5.

Stage one, first 90 days , stand up the as­sist­ive layer and the gov­ernance found­a­tion. De­ploy Mi­crosoft 365 Co­pi­lot to a defined pop­u­la­tion. Con­cur­rently, and this is the non-ob­vi­ous dis­cip­line, stand up the gov­ernance found­a­tion: name an agent owner, es­tab­lish the agent re­gistry, con­fig­ure Entra Agent ID. The gov­ernance es­tate pre­cedes the first autonom­ous agent; it does not fol­low it. The ob­ject­ive is a meas­ured pro­ductiv­ity gain in as­sist­ance and an es­tate ready to ab­sorb autonomy.

Stage two, months three to nine, proves agen­tic value in one or two pro­cesses. Pick one or two pro­cesses that match the char­ac­ter­istic screen of chapter one. They should be high-volume and rules-bound, di­git­ally nat­ive and veri­fi­able. Com­pose su­per­vised agents under a human-in-the-loop team­ing model. In­stru­ment them fully through AgentOps. The ob­ject­ive is a demon­strated, audited re­turn in pro­duc­tion, not a pro­lif­er­a­tion of pi­lots.

Stage three, months nine to eight­een , mi­grate to su­per­vised autonomy and scale. For the proven pro­cesses, mi­grate from human-in-the-loop to human-on-the-loop as the eval­u­ation evid­ence and the con­trols jus­tify it. Ex­tend to ad­ja­cent pro­cesses. In­tro­duce agent-to-agent del­eg­a­tion where it is gov­erned. The ob­ject­ive is a port­fo­lio of autonom­ous agents in­side a ma­ture gov­ernance es­tate, su­per­vised as a fleet.

Prerequisites, by stage

Each stage has entry con­di­tions, and most slipped roadmaps are stages begun without them. The con­di­tions are de­lib­er­ately mundane, be­cause the work they de­scribe is.

Noth­ing in that table needs a cap­ab­il­ity the Mi­crosoft es­tate does not already ship. Every entry con­di­tion is an or­gan­isa­tional act, which is this chapter's ar­gu­ment re­stated as a plan.

Applied AIVoice of Volge13 •
StageEntry conditionsExit evidence
One, assistive layer and governance foundationA named agent owner; an agent access policy signed by security; a configured registry; licence entitlements confirmed against the controls the programme intends to useA measured productivity change in a defined population, and a registry holding every existing agent, third-party agents included
Two, supervised agentic valueTwo candidate processes screened on volume, rule density and verifiability; a baseline measured before the agent exists; one connector tested against production-shaped dataAn audited return in production from one process, with a reviewer able to reconstruct a week of agent actions from logs
Three, supervised autonomy and scaleEvaluation evidence from the supervised stage; a dated revocation drill; a teaming model chosen per processA portfolio supervised as a fleet, with monthly reporting to the agent owner and a dated drill per agent class

The controls, and how each is tested

A con­trol nobody has ex­er­cised is a claim. The mat­rix below names the con­trol, what it does, and the test that demon­strates it, so that a pro­gramme can evid­ence its pos­ture rather than de­scribe it.

ControlWhat it doesTest
Agent identityGives each agent an identity the platform recognises natively, so actions attribute to it aloneTrace one agent action end to end in the sign-in and audit logs, and confirm that no shared account appears
Registry entryRecords every agent, in-house and third-party, with its owner and its data reachReconcile the registry against a tenant-wide query and against the build teams' own lists, then investigate every difference
Tool-level least privilegeGrants each agent only the tool permissions its task needsInvoke a tool outside the agent's scope and confirm refusal and a log entry
Delegation constraintStops a sub-agent exceeding its parent's authorityChain a call through a sub-agent to a resource the parent cannot reach, and confirm that it fails
Human gateHolds a consequential action until a person approves itSubmit a consequential action, confirm that it waits, and confirm that the approval is recorded with the approver's identity
Policy evaluationEvaluates every token the agent acquires against approval state and riskRun the policy in report-only, read the evaluation record, then enforce and read it again
RevocationWithdraws an agent's authority across the landscapeFire it against a live agent in a controlled window, then confirm lost access, invalid tokens and a log entry
Evaluation harnessTests behaviour continuously against defined scenarios, because the agent is non-deterministicRe-run the scenario set after a model change and compare against the previous result

Sources: Mi­crosoft Entra and Agent 365 guid­ance, 2026; OWASP Top 10 for Agen­tic Ap­plic­a­tions, 9 Decem­ber 2025; Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work, 2 Feb­ru­ary 2026. The test column is Volge's own for­mu­la­tion of how each con­trol is evid­enced.

Failure modes, and what each looks like

Agent fail­ures in pro­duc­tion are re­cog­nis­able, and most are vari­ants of five pat­terns. Each has an early sig­nal, and the sig­nal usu­ally sits in a log nobody reads.

  • Si­lent scope creep. Per­mis­sions widen through a suc­ces­sion of un­block­ing de­cisions, none of them re­viewed. The sig­nal is a fall­ing re­fusal rate against a rising call volume.
  • Del­eg­a­tion blow-out. A nar­row agent calls a broad one, and the ef­fect­ive reach is the union of the two. The sig­nal is a re­source touched by an agent that holds no per­mis­sion to it in its own grant.
  • Con­fid­ent error at volume. The agent is wrong in a con­sist­ent way, and the volume that jus­ti­fied it turns a small error rate into a large ex­pos­ure. The sig­nal is a cluster of down­stream cor­rec­tions in­side one pro­cess.
  • Ap­proval fa­tigue. A human-in-the-loop gate re­ceives more ap­provals than a per­son can con­sider, and the gate be­comes a click. The sig­nal is ap­proval latency fall­ing to­wards zero while volume rises.
  • Orphaned agent. The pro­cess changes, the owner moves, and the agent runs on. The sig­nal is an agent with re­cent activ­ity and no sign-in from its owner in a quarter.

Four of the five are caught by in­stru­ment­a­tion the land­scape already car­ries. The fifth, ap­proval fa­tigue, is caught only where someone meas­ures the time a human spends on each ap­proval, which is why that meas­ure be­longs in the AgentOps panel rather than in a sat­is­fac­tion sur­vey.

An agent maturity model

The model below loc­ates the en­ter­prise and defines its next in­cre­ment , the dia­gnostic coun­ter­part to the roadmap.

LevelDescriptionIdentity and governanceTeamingTypical outcome
0 , Ad hocUnmanaged use of assistants; shadow AINone; agents unknownUndefinedData-leak risk; no return captured
1 , AssistiveSanctioned copilots in productivity appsHuman identity onlyHuman as actorModest productivity gain
Applied AIVoice of Volge14 •
LevelDescriptionIdentity and governanceTeamingTypical outcome
2 , Supervised agenticComposed agents in 1-2 processesAgent registry; Entra Agent ID; HITL gatesHuman-in-the-loopFirst audited production return
3 , Scaled autonomyFleet of autonomous agents across processesFull Agent 365 control plane; continuous monitoringHuman-on-the-loopDeclared EBIT impact; fleet supervision
4 , Agentic enterpriseMulti-vendor agents coordinating via A2AEstate-wide governance spanning vendorsHuman-out-of-the-loop where reversibleProcess-level transformation

On Volge's own as­sess­ment, rather than on any pub­lished sur­vey, most large en­ter­prises sit between Level 0 and Level 1 as at late 2026. The 6 per cent of re­spond­ents who at­trib­ute 5 per cent or more of EBIT to their use of AI sit at Level 3 (McKinsey, Au­gust 2026). Level 4 is emer­gent.

What this delivers

The ar­chi­tec­ture and the roadmap are in­sep­ar­able from the gov­ernance es­tate of chapter three. The layered, pro­tocol-based ar­chi­tec­ture re­tires in­teg­ra­tion debt; the staged roadmap keeps autonomy from out­run­ning con­trol. The en­ter­prise that builds the gov­ernance found­a­tion first, com­poses rather than builds by de­fault, and climbs the autonomy spec­trum on the evid­ence of audited re­turns is the en­ter­prise that crosses the pro­duc­tion threshold, and de­clares an EBIT im­pact rather than join­ing the 40 per cent can­cel­la­tion pool.

, Agents in the Gulf and India

Microsoft commits to keeping data in region at rest. It does not commit to inferring in region, and says so in its own documentation. That asymmetry governs what an agent landscape may lawfully do in the Gulf and India, and no residency add-on repairs it. The regional picture is uneven in ways the policy rhetoric obscures. The state with the loudest localisation apparatus has the thinnest vendor footprint, and the control plane for agents has no regional dial at all.

The re­gional ques­tion for an agent pro­gramme is not whether the data stays in coun­try. It is whether the in­fer­ence does, and the pub­lished an­swer is no. Mi­crosoft's Co­pi­lot pri­vacy doc­u­ment­a­tion states that calls to the large lan­guage model are routed to the closest data­centres in the re­gion but can also call into other re­gions where ca­pa­city is avail­able dur­ing high util­isa­tion peri­ods, and that cus­tom­ers out­side the European Union may have their quer­ies pro­cessed in the United States, the EU or other re­gions (Mi­crosoft Learn, page up­dated 3 Novem­ber 2026). Co­pi­lot

be­came a covered work­load in the Mi­crosoft Product Terms data res­id­ency com­mit­ments on 1 March 2024, and that com­mit­ment is about stor­age. The European Union Data Bound­ary is the only pro­cessing com­mit­ment Mi­crosoft pub­lishes, and eli­gib­il­ity for it rests on a sign-up loc­a­tion in the EU or EFTA.

Read that against the way the re­gional com­mit­ments are usu­ally sold. India, Qatar and the United Arab Emir­ates are Local Re­gion Geo­graph­ies, so a ten­ant in any of the three holds dur­able com­mit­ments on data loc­a­tion for Ex­change On­line, SharePoint, OneDrive, Teams, Co­pi­lot and Co­pi­lot Chat, through Product Terms, Multi-Geo and the Ad­vanced Data Res­id­ency add-on (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). Every one of those com­mit­ments is about where data rests. None says where a prompt is pro­cessed. An en­ter­prise that bought Ad­vanced Data Res­id­ency to sat­isfy a reg­u­lator has bought a stor­age com­mit­ment and a se­cur­ity-tier com­mit­ment, and it should not rep­res­ent it as more.

Saudi Arabia, and the sharpest collision in the region

Saudi Ar­a­bia com­bines the re­gion's most elab­or­ate loc­al­isa­tion ap­par­atus with its thin­nest vendor foot­print, and the two do not meet. Mi­crosoft's pub­lished Azure re­gions list car­ries no Saudi entry, live, re­stric­ted or in pre­view, the Middle East entries being Qatar Cent­ral, UAE Cent­ral, UAE North and Is­rael Cent­ral (Mi­crosoft Learn, page up­dated 13 Oc­to­ber 2026). Saudi Ar­a­bia is ab­sent from Mi­crosoft 365 data res­id­ency as well. It is not a Local Re­gion Geo­graphy, is not eli­gible for the Ad­vanced Data Res­id­ency add-on, and does not ap­pear in Mi­crosoft's table of dur­able com­mit­ments on data loc­a­tion. A Saudi ten­ant falls into the cat­egory Mi­crosoft calls All Other Geo­graph­ies, where, in Mi­crosoft's words, cus­tomer data "isn't com­mit­ted to reside in any par­tic­u­lar data­cen­ter" and the loc­a­tion "is sub­ject to change without no­tice" (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026).

Two fur­ther Saudi facts be­long along­side it. A pro­vider serving any Saudi gov­ern­ment cus­tomer must hold CST re­gis­tra­tion at Class A or above, and Class A cov­ers only data clas­si­fied pub­lic, with con­fid­en­tial re­quir­ing Class B and secret or top secret re­quir­ing Class C, per CST's own re­gister. And the Na­tional Cy­ber­se­cur­ity Au­thor­ity's Cloud Cy­ber­se­cur­ity Con­trols, CCC-2:2024, de­leted the sub-con­trols of the 2020 edi­tion that had re­quired cloud ser­vices to be provided from in­side the King­dom, stat­ing that loc­al­isa­tion con­trols had been trans­ferred to the Na­tional Data Man­age­ment Of­fice. The re­ceiv­ing in­stru­ment has not been iden­ti­fied. Any

Applied AIVoice of Volge15 •

claim that NCA con­trols man­date in-King­dom host­ing is cit­ing a su­per­seded edi­tion.

Now set that against Saudi law. Art­icle 3-3-5 of the Cloud Com­put­ing Ser­vices Pro­vi­sion­ing Reg­u­la­tions ver­sion 4, is­sued by the Com­mu­nic­a­tions, Space and Tech­no­logy Com­mis­sion under De­cision No. 506/1445 and ef­fect­ive 10 Oc­to­ber 2023, pro­hib­its a pro­vider or a sub­scriber from trans­fer­ring con­tent com­pris­ing data of Saudi gov­ern­ment agen­cies out­side the King­dom for any pur­pose and in any form, per­man­ently or tem­por­ar­ily, in­clud­ing tem­por­ary stor­age and backup, un­less an­other Saudi in­stru­ment ex­pressly per­mits it. The art­icle num­ber and word­ing rest on named law-firm com­ment­ary rather than the primary text, Ald­habaan Law Firm writ­ing in the Cham­bers Cloud Com­put­ing 2026 guide, be­cause the CST file could not be read. The col­li­sion is com­plete. The law for­bids ex­port of that data, and the vendor of­fers no com­mit­ment to avoid it.

What an ar­chi­tect does about this has three hon­est an­swers, and a sov­er­eign Azure re­gion is not among them. Keep the reg­u­lated work­load off the pub­lic es­tate and run Mi­crosoft soft­ware on in­fra­struc­ture the cus­tomer or a local part­ner op­er­ates, which is what Mi­crosoft's own Sov­er­eign Private Cloud branch de­scribes through Azure Local and Mi­crosoft 365 Local. Or con­fine the agent land­scape to data that is not gov­ern­ment agency data, and doc­u­ment the bound­ary. Or ac­cept a cross-bor­der pro­cessing po­s­i­tion and ob­tain the ex­press per­mis­sion the reg­u­la­tion con­tem­plates. Mi­crosoft's dif­fer­en­ti­at­ing sov­er­eign fea­tures, the EU Data Bound­ary, Data Guard­ian and op­er­a­tions con­trolled by European res­id­ents, are doc­u­mented as EU and EFTA fea­tures, and the only two Na­tional Part­ner Clouds Mi­crosoft names are Bleu in France and Delos in Ger­many (Mi­crosoft Learn, pages up­dated 31 March 2026). Out­side Europe, Mi­crosoft's sov­er­eignty story is en­cryp­tion and policy rather than jur­is­dic­tion.

One vendor, four regional footprints, one country

Qatar shows why the ques­tion of in-re­gion avail­ab­il­ity has no single an­swer. Qatar Cent­ral, at Doha, is a live Azure re­gion with three avail­ab­il­ity zones, and a Foundry pro­ject can be cre­ated in it (Mi­crosoft Learn, page up­dated 25 Au­gust 2026). Qatar is a Mi­crosoft 365 Local Re­gion Geo­graphy car­ry­ing the full set of dur­able com­mit­ments on the core work­loads. And Qatar is ab­sent from the Co­pi­lot Stu­dio data loc­a­tions table, where Mi­crosoft states plainly that if a ten­ant's loc­a­tion is not lis­ted, data is stored in the United States (Mi­crosoft Learn, page up­dated 30 April 2026). A Qatari ten­ant can there­fore build an agent against a Foundry pro­ject in Doha and have

the agent's own con­fig­ur­a­tion and con­ver­sa­tion data held in Amer­ica.

JurisdictionAzure regionMicrosoft 365 Local Region GeographyFoundry projectCopilot Studio in region
Saudi ArabiaNoneNoNoNo
United Arab EmiratesTwo, one of them restrictedYesYes, UAE NorthYes, UAE North
QatarOne, unpairedYesYes, Qatar CentralNo; data stored in the United States
IndiaFour, two with availability zonesYesYes, two regionsYes, two regions
Kuwait, Oman, BahrainNoneNoNoNo

Sources: Mi­crosoft Learn re­gion and data-loc­a­tion doc­u­ment­a­tion, pages up­dated between 25 Au­gust and 27 Oc­to­ber 2026, veri­fied 8 Novem­ber 2026. Com­pil­a­tion by Volge.

Two re­fine­ments mat­ter to an ar­chi­tect read­ing that table. Foundry Agent Ser­vice, the man­aged runtime rather than the pro­ject, lists South India and UAE North among its sup­por­ted re­gions and does not list Qatar Cent­ral, so Qatar can host a Foundry pro­ject but not, on the pub­lished evid­ence, the man­aged agent ser­vice (Mi­crosoft Learn, veri­fied 8 Novem­ber 2026). And Mi­crosoft de­clines to pub­lish a single mat­rix of model avail­ab­il­ity by re­gion, stat­ing that avail­ab­il­ity de­pends on the pro­vider and the de­ploy­ment type. A pro­ject re­gion is not a model re­gion. A pro­gramme that as­sumes a named fron­tier model is de­ploy­able in Doha or Dubai is mak­ing a claim the vendor does not sup­port.

Agent 365 has no regional dial

The con­trol plane takes the ten­ant's geo­graphy and of­fers noth­ing else. Mi­crosoft's Dy­nam­ics 365 and Power Plat­form res­id­ency doc­u­ment­a­tion states that Agent 365 pro­cesses and stores the traces for agents built across Mi­crosoft ser­vices, that it "cur­rently stores and pro­cesses the data at the ten­ant geo home loc­a­tion", and that macro re­gion geo­graphy and en­vir­on­ment-level re­gion se­lec­tion do not apply to it (Mi­crosoft Learn, page up­dated 7 Oc­to­ber 2026). Where a Power Plat­form cus­tomer can place an en­vir­on­ment in a chosen re­gion, an Agent 365 cus­tomer has no such choice.

Three re­lated ex­cep­tions are doc­u­mented and easy to miss. His­tor­ical agent activ­ity data in Co­pi­lot Stu­dio, con­ver­sa­tion logs in­cluded, is stored in the re­gion of the end user's Ex­change mail­box, which moves agent logs out of the Power Plat­form res­id­ency bound­ary and into the Mi­crosoft 365 one. Pre­view and pre-re­lease ser­vices typ­ic­ally store cus­tomer

Applied AIVoice of Volge16 •

data in the United States and may store it glob­ally, which is the clause to read be­fore re­com­mend­ing an early-ac­cess agent cap­ab­il­ity to a res­id­ency-con­strained cus­tomer. And every Gulf and In­dian ten­ant's Mi­crosoft Forms data sits in the United States, with Viva In­sights, Viva Learn­ing and Plan­ner de­fault­ing to Europe for all six Gulf states (Mi­crosoft Learn, page up­dated 8 Oc­to­ber 2026).

The con­sequence lands on the audit re­cord. Agent traces are the evid­ence of what the agents did, and they fol­low the ten­ant rather than the work­load. For a ten­ant whose home geo­graphy is a Local Re­gion Geo­graphy, the po­s­i­tion is tol­er­able. For a Saudi, Kuwaiti, Omani or Bahraini ten­ant it means the gov­ernance evid­ence for the agent land­scape sits wherever Mi­crosoft places the ten­ant, which is the po­s­i­tion Mi­crosoft it­self de­scribes as no com­mit­ment. A pro­gramme that has told its reg­u­lator that agent over­sight hap­pens in coun­try should check this be­fore it is asked.

What the regulator requires, and what it does not

No in­stru­ment in Saudi Ar­a­bia, the UAE, Qatar, Kuwait, Oman, Bahrain or India im­poses an AI-spe­cific ob­lig­a­tion of gen­eral ap­plic­a­tion on an en­ter­prise de­ploy­ing agents. None reg­u­lates a ma­chine iden­tity as a cat­egory either. That is a veri­fied neg­at­ive from ex­haust­ive text search of the primary in­stru­ments, not an ab­sence of re­search. Four bind­ing in­stru­ments do bite on an op­er­ator, and each is nar­row. DIFC Data Pro­tec­tion Reg­u­la­tions Reg­u­la­tion 10, in force 1 Septem­ber 2023, al­loc­ates ac­count­ab­il­ity for an autonom­ous or semi-autonom­ous sys­tem in­side the Dubai In­ter­na­tional Fin­an­cial Centre. India's In­form­a­tion Tech­no­logy Amend­ment Rules, 2026, G.S.R. 120(E), in force 20 Feb­ru­ary 2026, bind in­ter­me­di­ar­ies on syn­thet­ic­ally gen­er­ated in­form­a­tion. Reg­u­la­tion 16C of the SEBI (In­ter­me­di­ar­ies) (Amend­ment) Reg­u­la­tions, 2025, has been in force since early Feb­ru­ary 2025. And Art­icle 15 of Bahrain's Per­sonal Data Pro­tec­tion Law, with Min­is­terial Or­ders 43 and 44 of 2022, gov­erns auto­mated pro­fil­ing and bio­met­ric pro­cessing.

Reg­u­la­tion 16C is the one every agent pro­gramme in the re­gion should read. It places sole re­spons­ib­il­ity for the out­put of an ar­ti­fi­cial-in­tel­li­gence or ma­chine-learn­ing tool on the reg­u­lated per­son using it, in-house or pro­cured, ir­re­spect­ive of the scale and scen­ario of ad­op­tion. It binds In­dian se­cur­it­ies-mar­ket in­ter­me­di­ar­ies only. It is non­ethe­less the only pro­vi­sion in the seven jur­is­dic­tions that al­loc­ates li­ab­il­ity for an AI sys­tem's out­put, and it an­swers the ques­tion a board asks first. Build­ing the agent does not move the li­ab­il­ity. Buy­ing it does not either.

The Saudi Na­tional Cy­ber­se­cur­ity Au­thor­ity has gone fur­thest in re­cog­nising the ob­ject. Its draft AI Cy­ber­se­cur­ity Guidelines, AICG-1:2026, con­sul­ted from 5 July to 5 Au­gust 2026, bring agen­tic AI ex­pressly into scope, define AI Agent, Agen­tic AI and Tool or Plu­gin, and carry Model Con­text Pro­tocol in the ab­bre­vi­ation list. It is the first reg­u­lat­ory text in the Gulf to treat an agent as a thing to be se­cured. It is also a draft, it is non-bind­ing, and no final in­stru­ment had been pub­lished as at 8 Novem­ber 2026. Saudi guid­ance that is op­er­at­ive runs the other way on sub­stance. SDAIA's Gen­er­at­ive Ar­ti­fi­cial In­tel­li­gence for Gov­ern­ment Guidelines, index date 15 May 2025, state that gov­ern­ment en­tit­ies must re­frain from using gen­er­at­ive AI in crit­ical de­cision-mak­ing pro­cesses af­fect­ing in­di­vidu­als or the vital in­terests of the King­dom, with out­puts human-re­viewed. That is ad­min­is­trat­ive guid­ance writ­ten in a man­dat­ory re­gister, car­ry­ing no stat­utory in­stru­ment num­ber and no pen­alty re­gime.

The ab­sence of a rule is not the ab­sence of an ob­lig­a­tion, and an op­er­ator should not read it as per­mis­sion. Three ob­lig­a­tions bind whether or not the in­stru­ment men­tions agents. Data pro­tec­tion law ap­plies to the pro­cessing an agent per­forms, and India's Di­gital Per­sonal Data Pro­tec­tion Rules, 2025, no­ti­fied on 13 Novem­ber 2025 as G.S.R. 846(E), bring no­tice, se­cur­ity safe­guards, breach in­tim­a­tion and Sig­ni­fic­ant Data Fi­du­ciary du­ties into force on 13 May 2027, with con­sent man­agers from 13 Novem­ber 2026. Sec­toral cyber frame­works apply to the sys­tems the agent touches. And third-party rules apply to the plat­form, be­cause the firm stays ac­count­able for a ser­vice it has con­trac­ted out. SEBI's Cy­ber­se­cur­ity and Cyber Re­si­li­ence Frame­work says so in terms, mak­ing reg­u­lated en­tit­ies solely ac­count­able for all as­pects of third-party ser­vices, in­clud­ing the se­cur­ity of their data and logs.

India, and the only dated Zero Trust mandate in the region

India pro­duced the re­gion's clearest Zero Trust re­quire­ment and then softened it, and the arc is worth know­ing be­cause it shows what a reg­u­lator does when a prin­ciple meets an in­stalled es­tate. SEBI's Cy­ber­se­cur­ity and Cyber Re­si­li­ence Frame­work, cir­cu­lar SEBI/HO/ITD-1/ITD_CSC_EXT/P/CIR/2024/113 of 20 Au­gust 2024, re­quired under stand­ards PR.AA.S4 and PR.AA.S5 a zero-trust se­cur­ity model in which ac­cess to crit­ical sys­tems, from in­side or out­side the en­tity's net­work, is denied by de­fault and al­lowed only after proper au­then­tic­a­tion and au­thor­isa­tion. Tech­nical Cla­ri­fic­a­tions cir­cu­lar SEBI/HO/ITD-1/ITD_CSC_EXT/P/CIR/2025/119 of 28 Au­gust 2025 re­placed

Applied AIVoice of Volge17 •

that with a re­quire­ment to im­ple­ment sug­ges­ted strategies and meth­od­o­lo­gies such as zero-trust net­works, seg­ment­a­tion, no single point of fail­ure and high avail­ab­il­ity, ap­proved by the en­tity's IT com­mit­tee. A de­fault-deny man­date be­came a com­mit­tee-ap­proved menu in twelve months.

The read­ing for an agent pro­gramme is not that Zero Trust re­treated. A de­fault-deny rule writ­ten against an es­tate of human users and leg­acy ap­plic­a­tions proved un­en­force­able at the pace the cir­cu­lar set. An agent pop­u­la­tion is the one place where de­fault-deny is straight­for­ward, be­cause every agent is new, every agent has a defin­able task, and noth­ing leg­acy de­pends on its stand­ing ac­cess. Apply to agents the pos­ture the reg­u­lator could not im­pose on everything.

Two fur­ther In­dian facts con­strain an agent ar­chi­tec­ture dir­ectly, and both come from the vendor rather than the state. In­dian tax law pre­vents a ten­ant whose home loc­a­tion is out­side India from cre­at­ing a Data­verse en­vir­on­ment in India, and Mi­crosoft doc­u­ments it with no ex­cep­tion avail­able, where an ex­cep­tion is offered for Aus­tralia (Mi­crosoft Learn, page up­dated 29 May 2026). A Gulf-headquartered group there­fore can­not loc­al­ise its In­dian sub­si­di­ary's agent and ap­plic­a­tion data in India. And on-premises data gate­ways are not avail­able in the India re­gion, which re­moves the stand­ard route from a cloud work­flow to an on-premises sys­tem of re­cord, in the mar­ket with the largest on-premises in­stalled base in this set.

India's data-pro­tec­tion ob­lig­a­tions trans­late into agent design in four con­crete ways, and all four fall due on 13 May 2027 under the 2025 Rules. No­tice must be itemised, so an agent act­ing on per­sonal data has to be trace­able to a stated pur­pose rather than to a gen­eral con­sent. Re­ten­tion is pur­pose-bound, and rules 6(1)(e) and 8(3) im­pose a min­imum one-year re­ten­tion of per­sonal data and of logs of pro­cessing, bind­ing on a pro­cessor en­gaged by a fi­du­ciary as well, which cuts against a de­le­tion-by-de­fault log­ging policy. Breach in­tim­a­tion runs to the Data Pro­tec­tion Board and to af­fected prin­cipals, so an agent's ac­tions have to be re­con­struct­able fast enough to scope a breach. And rule 6(1) re­quires en­cryp­tion or ob­fus­ca­tion or mask­ing, ac­cess con­trol, and vis­ib­il­ity of ac­cess through logs, mon­it­or­ing and re­view. The or­gan­isa­tion stays ac­count­able as a data fi­du­ciary even where the de­cision is auto­mated, which makes all of this a design re­quire­ment rather than a later patch.

The demand side, and why the Gulf timetable is short

Gulf pub­lic-sec­tor de­mand for agents is a stated com­mit­ment with named ac­count­ab­il­ity, which is un­usual any­where.

The UAE Cab­inet, meet­ing on 23 April 2026, ap­proved a frame­work to trans­form half of fed­eral gov­ern­ment sec­tors, ser­vices and op­er­a­tions to agen­tic AI for autonom­ous ex­e­cu­tion and de­cision-mak­ing within two years, phased across min­is­tries and fed­eral en­tit­ies, with min­is­ters, dir­ect­ors gen­eral and en­tit­ies as­sessed on ad­op­tion speed. Over­sight sits with Sheikh Man­sour bin Zayed, and the taskforce is chaired by Mo­hammad Al Ger­gawi, Min­is­ter of Cab­inet Af­fairs. No in­stru­ment text has been pub­lished, so the status is an an­nounced frame­work with ap­proved Cab­inet res­ol­u­tions. Noth­ing com­par­able ex­ists in the other six jur­is­dic­tions.

The same meet­ing ap­proved four in­stru­ments a sup­plier will meet in­side a fed­eral busi­ness case. They are a Code for Gov­ern­ment Ser­vices and Zero Bur­eau­cracy, a Gov­ern­ment Ser­vices Di­gital Re­cords Policy, a Gov­ern­ment Ser­vices Data Shar­ing Policy framed as col­lect once and use se­curely across en­tit­ies, and a Fed­eral Gov­ern­ment Guide to align­ing di­gital gov­ern­ment pro­jects with na­tional pri­or­it­ies, which en­cour­ages reuse of cap­ab­il­it­ies such as di­gital iden­tity and di­gital wal­lets and re­quires feas­ib­il­ity stud­ies. None has pub­lished text. Abu Dhabi re­ports fur­ther along. Its Gov­ern­ment Di­gital Strategy 2025 to 2027, an­nounced by the De­part­ment of Gov­ern­ment En­able­ment on 30 Septem­ber 2025, com­mits AED 13 bil­lion, names full sov­er­eign cloud ad­op­tion as a pil­lar, and re­ports more than 100 AI use cases across more than 40 en­tit­ies. Those fig­ures are com­mit­ments and pro­gramme self-re­ports, not meas­ured ad­op­tion.

India's de­mand is fun­ded rather than man­dated. The Union Cab­inet ap­proved the IndiaAI Mis­sion on 7 March 2024 with an out­lay of Rs 10,371.92 crore over five years, and a MeitY reply laid in the Lok Sabha on 5 Au­gust 2026 re­por­ted 15 com­pute ser­vice pro­viders em­pan­elled, 237 pro­jects ap­proved for sub­sid­ised com­pute and 93.18 lakh GPU hours sanc­tioned. Ac­cred­it­a­tion is the other timetable, and it gates the plan rather than clos­ing it. A pro­vider of­fer­ing cloud ser­vices to Dubai gov­ern­ment and semi-gov­ern­ment en­tit­ies must hold the Dubai Elec­tronic Se­cur­ity Cen­ter's Cloud Ser­vice Pro­vider Se­cur­ity Stand­ard cer­ti­fic­a­tion, which DESC de­scribes as man­dat­ory in its own words, with an­nual sur­veil­lance audits and re­cer­ti­fic­a­tion every three years. Kuwait re­quires a CITRA cloud li­cence with an in-Kuwait data centre to serve the pub­lic sec­tor, and Oman re­quires a Tele­com­mu­nic­a­tions Reg­u­lat­ory Au­thor­ity au­thor­isa­tion be­fore con­tract­ing.

What this constrains

The re­gional ground changes the se­quence rather than the des­tin­a­tion. Build the in­vent­ory of what the agent land­scape touches be­fore choos­ing where it runs, be­cause the

Applied AIVoice of Volge18 •

res­id­ency ques­tion is answered per work­load and not per ten­ant. Sep­ar­ate the stor­age com­mit­ment from the pro­cessing com­mit­ment in every doc­u­ment that reaches a reg­u­lator, be­cause the vendor does. Treat the ten­ant-bound Agent 365 traces as a design con­straint on where the audit evid­ence lives, not as a set­ting to be changed later. And where no in­stru­ment speaks, write the con­trol, have it ap­proved, and re­cord the de­cision. The ab­sence of a rule on ma­chine iden­tity will not sur­vive the next su­per­vis­ory cycle, and the de­ploy­ing firm car­ries the li­ab­il­ity for the out­put in the mean­time.

What to write down, and where

A re­gional po­s­i­tion is only use­ful once it sits on paper in three places, and the same facts serve all three. Put them in the ar­chi­tec­ture de­cision re­cord, in the vendor con­tract and in the su­per­vis­ory sub­mis­sion, and keep the word­ing identical, be­cause a dif­fer­ence between them is what an ex­am­iner finds.

  • In the ar­chi­tec­ture de­cision re­cord. The data loc­a­tion for each layer, named per work­load, with the pro­cessing po­s­i­tion stated sep­ar­ately from the stor­age po­s­i­tion, and the Agent 365 trace loc­a­tion stated ex­pli­citly.
  • In the con­tract. The res­id­ency com­mit­ment re­lied upon, by name, with the add-on it de­pends on and the ser­vices it ex­cludes. Mi­crosoft's own doc­u­ment­a­tion is the right ref­er­ence, be­cause it is spe­cific where a mar­ket­ing claim is not.
  • In the su­per­vis­ory sub­mis­sion. The clas­si­fic­a­tion of the data the agents touch, the legal basis for any cross-bor­der pro­cessing, and the con­trol that would de­tect a breach of the stated po­s­i­tion rather than the policy that for­bids it.

One fur­ther line be­longs in all three, and it is the one most often miss­ing. Name what is not in re­gion, the iden­tity layer and any pre­view cap­ab­il­ity in use in­cluded. A dis­clos­ure made in ad­vance is a design con­straint. The same fact found by a su­per­visor is a find­ing.

, The leadership agenda

The trans­ition from co­pi­lot to auto­pi­lot is the de­fin­ing op­er­at­ing-model change of the present cycle, and the evid­ence in this paper re­solves to a short list of con­clu­sions for the board and the ex­ec­ut­ive team.

Six ques­tions for the lead­er­ship agenda. First, who is the named owner of the agent land­scape, and is that owner loc­ated in the busi­ness rather than in tech­no­logy alone. Second, does the gov­ernance found­a­tion of agent iden­tity, re­gistry and policy pre­cede the first autonom­ous agent or lag it. Third, are the early pro­cesses screened on volume and rule dens­ity, on di­gital nat­ive­ness and veri­fi­ab­il­ity, or on en­thu­si­asm.

Fourth, is the team­ing model de­signed per pro­cess, and does it sat­isfy the human-over­sight re­quire­ments of the ap­plic­able reg­u­la­tion. Fifth, is the ar­chi­tec­ture built to open pro­to­cols, pre­serving the op­tion of a multi-vendor land­scape. Sixth, does the res­id­ency po­s­i­tion dis­tin­guish stor­age from pro­cessing, and has the loc­a­tion of the agent con­trol plane's own data been checked against the ob­lig­a­tion the en­ter­prise has ac­cep­ted.

Sup­ply is settled and abund­ant. The con­straint is the en­ter­prise's ca­pa­city to ab­sorb autonomy safely, and that ca­pa­city is gov­ern­able rather than tech­no­lo­gical. The re­turns are large but con­di­tional on reach­ing pro­duc­tion, a threshold three-quar­ters of en­ter­prises have yet to cross. Only a quarter have moved 40 per cent or more of their pi­lots into pro­duc­tion (De­loitte, 21 Janu­ary 2026), and just 6 per cent at­trib­ute 5 per cent or more of EBIT to their use of AI (McKinsey, Au­gust 2026). The cross­ing is an act of ex­e­cu­tion dis­cip­line, not of tech­no­logy se­lec­tion, and it is where the 40 per cent can­cel­la­tion rate is either earned or avoided.

What the board should see, and how often. Four items, quarterly, and they are de­lib­er­ately few. The agent pop­u­la­tion, by ac­cess pat­tern, with the count dis­covered out­side the re­gistry. The share of agent ac­tions a re­viewer could re­con­struct from logs alone, sampled rather than as­ser­ted. The date of the last re­voc­a­tion drill, per agent class. And the meas­ured out­come of each pro­duc­tion pro­cess against the baseline taken be­fore its agent ex­is­ted. A board re­ceiv­ing those four can tell whether the pro­gramme is gov­erned. A board re­ceiv­ing a ma­tur­ity score can­not.

The re­gional item be­longs on the same page. Where the en­ter­prise op­er­ates in the Gulf or India, the board should see the res­id­ency po­s­i­tion layer by layer, with the parts that are not in re­gion named, and with any re­li­ance on a pre­view cap­ab­il­ity dis­closed. That po­s­i­tion moves when the vendor changes its doc­u­ment­a­tion, which hap­pens without no­tice, so it is a stand­ing item rather than a one-off as­sur­ance.

The cross­ing is bounded by con­trol, and the con­trols are already bought. For most large en­ter­prises the plat­form de­cision was settled by ex­ist­ing Mi­crosoft com­mit­ments, and the en­ti­tle­ment to the agent con­trols ar­rived with Agent 365 or Mi­crosoft 365 E7. What re­mains is con­fig­ur­a­tion, se­quence and ac­count­ab­il­ity, in that order. A pro­gramme that names an owner in the busi­ness, stands the gov­ernance es­tate up be­fore the first autonom­ous agent, screens its pro­cesses on sig­na­ture rather than on en­thu­si­asm, and veri­fies each con­trol by ex­er­cising it, will cross the pro­duc­tion threshold. A pro­gramme that does the re­verse will sup­ply an­other entry to the can­cel­la­tion pool, and will do so on sched­ule.

Applied AIVoice of Volge19 •

Appendices

Appendix A , Enterprise agent platform comparison

PlatformGeneral availabilityGreatest strengthPrincipal constraint
Microsoft (Copilot, Copilot Studio, Agent 365, Foundry)Layered; M365 agentic action GA 22 Apr 2026; Agent 365 GA 1 May 2026Identity-led control plane spanning in-house and third-party agents; widest productivity installed baseStrongest value realised within the Microsoft estate
Salesforce Agentforce2024Depth inside the Salesforce CRM ecosystemAgents outside the CRM boundary require additional integration decisions
ServiceNow AI Agents2024Governed autonomous work where flow runs through ServiceNow (ITSM, security, HR)Value concentrated in ServiceNow-resident workflows
Amazon Bedrock AgentCore2025Bring-your-own-framework; any framework, model, or protocol; AWS-grade isolationHigher engineering burden; less out-of-the-box governance
Google Gemini Enterprise2025Integrated partner agents (Box, Workday, and others); full-stack model accessNewer enterprise governance estate

In­ter­op­er­ab­il­ity note. The Agent-to-Agent (A2A) pro­tocol, a Linux Found­a­tion pro­ject con­trib­uted by Google in 2025, passed 150 sup­port­ing or­gan­isa­tions at its first an­niversary, with en­ter­prise pro­duc­tion use re­por­ted in sup­ply chain and fin­an­cial ser­vices, in in­sur­ance and in IT op­er­a­tions (Linux Found­a­tion, 9 April 2026). The agen­tic en­ter­prise should be as­sumed multi-vendor by de­fault. Sources are primary vendor and found­a­tion dis­clos­ures, 2025-2026.

Appendix B , Agent governance checklist

The fol­low­ing is a min­imum con­trol set, drawn from the OWASP Top 10 for Agen­tic Ap­plic­a­tions (pub­lished 9 Decem­ber 2025, and styled by OWASP as the 2026 edi­tion), the Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work (Feb­ru­ary 2026, now stew­ar­ded out­side CSA), and Mi­crosoft's Entra and Agent 365 guid­ance (2025-2026).

Iden­tity and ac­cess

  • Every agent is as­signed a first-class, cryp­to­graphic iden­tity (for ex­ample, Entra Agent ID).
  • Every agent is re­cor­ded in a cent­ral re­gistry, in­clud­ing third-party and in-house agents.
  • Least-priv­ilege ac­cess is ap­plied to every tool call; sub-agent per­mis­sion in­her­it­ance is con­trolled.
  • Cre­den­tial re­voc­a­tion and an es­tate-wide kill-switch are tested, not as­sumed.

Over­sight and ac­count­ab­il­ity

  • A named owner, loc­ated in the busi­ness, is ac­count­able for each agent.
  • Human-over­sight gates are defined per pro­cess and sat­isfy ap­plic­able reg­u­la­tion (for ex­ample, EU AI Act Art­icle 14).
  • Con­sequen­tial ac­tions re­quire human sign-off until eval­u­ation evid­ence jus­ti­fies oth­er­wise.

Ob­serv­ab­il­ity and eval­u­ation

  • Every au­then­tic­a­tion, tool in­voc­a­tion, del­eg­a­tion, and policy de­cision is logged with what, why, on whose be­half, and under what policy.
  • Con­tinu­ous eval­u­ation runs against pre­defined re­li­ab­il­ity, fair­ness, and se­cur­ity scen­arios.
  • Ser­vice-level ob­ject­ives are defined and mon­itored in real time.

Data pro­tec­tion

  • Data-pro­tec­tion-by-design sat­is­fies the ap­plic­able re­gime (for ex­ample, India DPDP Rules 2025; EU AI Act).
  • Pur­pose-based re­ten­tion and breach-no­ti­fic­a­tion pro­cesses (72 hours under DPDP) are in place.
  • Ac­count­ab­il­ity is re­tained by the or­gan­isa­tion even where de­cisions are auto­mated.

Re­gion and res­id­ency

  • The res­id­ency po­s­i­tion is re­cor­ded per work­load, dis­tin­guish­ing stor­age at rest from the loc­a­tion of in­fer­ence.
  • The Agent 365 trace loc­a­tion and the Co­pi­lot Stu­dio data loc­a­tion for the ten­ant are con­firmed against cur­rent vendor doc­u­ment­a­tion rather than as­sumed from the ten­ant's coun­try.
  • Entra, and any pre­view cap­ab­il­ity in use, are named as data that may sit out­side the chosen geo­graphy.
  • Where no in­stru­ment reg­u­lates ma­chine iden­tity, the ad­op­ted con­trol is doc­u­mented, mapped to a re­cog­nised frame­work, and ap­proved through the en­ter­prise's own gov­ernance.

The EU AI Act de­fer­ral provides plan­ning time; it does not re­duce the sub­stant­ive stand­ard, and the Art­icle 50(2) trans­par­ency ob­lig­a­tion falls due on 2 Decem­ber 2026. En­ter­prises serving EU or In­dian data sub­jects should treat com­pli­ance as a design re­quire­ment. This cal­en­dar is in­dic­at­ive and not legal

Applied AIVoice of Volge20 •

ad­vice; Volge is not a law firm, and qual­i­fied coun­sel should be en­gaged for com­pli­ance de­cisions.

Appendix C , Regulatory calendar (selected, as at 8 November 2026)

RegimeProvisionDateStatus
EU AI ActProhibited-practice provisionsIn force since 2 Feb 2025Active
EU AI ActArticle 50(2) , machine-readable marking and disclosure of synthetic contentApplies from 2 Dec 2026Moved from 2 Aug 2026 by Regulation (EU) 2026/1744
EU AI ActHigh-risk obligations, Annex III (use-based)Postponed to 2 Dec 2027Deferred by Regulation (EU) 2026/1744, in force 27 Jul 2026
EU AI ActHigh-risk obligations, Annex I (product-regulated)Postponed to 2 Aug 2028Deferred by Regulation (EU) 2026/1744, in force 27 Jul 2026
EU AI ActArticle 14 , human oversight of high-risk systemsTied to high-risk timelinesSubstantive standard unchanged by deferral
India DPDPDPDP Rules, 2025, notified by MeitYNotified 2025; phasedActive; phased commencement
India DPDPBreach notification within 72 hoursPer RulesActive
India DPDPSignificant Data Fiduciary obligations (DPIA, audit, algorithmic-fairness assessment, DPO)Per RulesActive for designated SDFs

Appendix F , The platform in motion: the Copilot model layer (as at 8 November 2026)

This ex­hibit re­cords what shipped in Mi­crosoft 365 Co­pi­lot by Novem­ber 2026, be­cause the product moved in the dir­ec­tion this paper ar­gues: model choice at the point of work, and autonomy de­livered be­hind a human ap­proval step.

The model se­lector:

Model or modeSelected forStatus, November 2026
AutoEveryday work; Copilot routes the request to the model it judges best suitedThe default in the selector
GPT-5.6 seriesMicrosoft's preferred models across Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork since 9 July 2026Available, and where Auto generally lands
Claude Opus 5 (Anthropic)Writing, reasoning, and data and analysis; long multi-step workAdded to the selector on 24 July 2026, across Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork and Copilot Studio
Think deeperSlower, deeper reasoning on complex tasksAvailable as a mode

Agen­tic ac­tions across the apps: Word , choose Claude or a GPT model to re­write, re­fine and sum­mar­ise in place; catch

up on what changed across ver­sions (model choice at the point of au­thor­ing). Excel , Co­pi­lot pro­poses a plan, asks cla­ri­fy­ing ques­tions, then ap­plies only the changes you ap­prove (the ap­proval gate, shipped). Mi­crosoft re­tired the Agent Mode label at the 22 April 2026 gen­eral avail­ab­il­ity, on the reas­on­ing that the be­ha­viour is core to Co­pi­lot rather than a sep­ar­ate mode. PowerPoint , Co­pi­lot grills the deck, sur­fa­cing likely audi­ence ques­tions (the agent as ad­versarial re­viewer). Teams , man­age and re­move meet­ing re­caps (gov­ernance of the gen­er­ated re­cord).

The de­fault is Auto mode, which routes each re­quest to the model it judges best suited rather than to a fixed house model. Named mod­els are se­lect­able where the or­gan­isa­tion per­mits , the multi-vendor es­tate of chapter two, made con­crete.

The pat­tern is this paper's ar­gu­ment made product. Excel's plan-then-apply is the human-in-the-loop con­trol sur­face of chapter three, shipped as a de­fault in­ter­ac­tion. Cap­ab­il­it­ies re­flect Mi­crosoft 365 Co­pi­lot re­leases through Novem­ber 2026. Neither Claude nor any single GPT model is the de­fault. Auto mode is, and it se­lects per re­quest. Re-verify at re­view.

Pre­pared by Volge · June 2026, re­vised Novem­ber 2026. This doc­u­ment is a point of view in­ten­ded to in­form ex­ec­ut­ive de­cision-mak­ing. It is not legal, fin­an­cial, or in­vest­ment ad­vice. Stat­ist­ics are at­trib­uted to their sources and were ac­cur­ate to the best of avail­able know­ledge as at 8 Novem­ber 2026; read­ers are ad­vised that the agen­tic-AI field is mov­ing rap­idly and fig­ures should be re­con­firmed be­fore they are re­lied upon for in­vest­ment de­cisions.

Com­piled by Volge from primary in­stru­ments and, where the primary text could not be read, from named law-firm com­ment­ary iden­ti­fied at the point of use. This table is in­dic­at­ive and not legal ad­vice; Volge is not a law firm, and qual­i­fied coun­sel should be en­gaged for com­pli­ance de­cisions.

Applied AIVoice of Volge21 •

Appendix G: regional instruments bearing on an agent programme, as at 8 November 2026

JurisdictionInstrumentWhat it doesStatus
Saudi ArabiaCST Cloud Computing Services Provisioning Regulations v4, Decision 506/1445, effective 10 October 2023Prohibits transfer of Saudi government agency data outside the Kingdom; gates providers by Class A, B or C registrationIn force
Saudi ArabiaNCA draft AI Cybersecurity Guidelines AICG-1:2026, consulted 5 July to 5 August 2026Brings agentic AI into scope and defines AI Agent, Agentic AI and Tool or PluginProposed; no final instrument published
Saudi ArabiaSDAIA Generative AI for Government Guidelines, index date 15 May 2025Directs government entities to refrain from generative AI in critical decisions affecting individualsNon-binding guideline
United Arab EmiratesDIFC Data Protection Regulations, Regulation 10, 1 September 2023Allocates accountability for an autonomous system between Deployer and Operator, in DIFC onlyIn force
United Arab EmiratesCabinet decision of 23 April 2026Commits half of federal government sectors, services and operations to agentic AI within two yearsAnnounced framework; no published instrument text
United Arab EmiratesDESC Cloud Service Provider Security StandardMandatory certification for providers offering cloud to Dubai government and semi-government entitiesIn force
IndiaDPDP Rules, 2025, G.S.R. 846(E), notified 13 November 2025Notice, security safeguards, breach intimation and Significant Data Fiduciary dutiesPhasing; consent managers 13 November 2026, the remainder 13 May 2027
IndiaSEBI (Intermediaries) (Amendment) Regulations, 2025, regulation 16CPlaces sole responsibility for an AI tool's output on the regulated firm, built or boughtIn force since early February 2025
IndiaSEBI CSCRF of 20 August 2024, as amended 28 August 2025Required default-deny Zero Trust for critical systems; amended to an IT-committee-approved menuIn force as amended
IndiaIT Amendment Rules, 2026, G.S.R. 120(E)Obligations on synthetically generated information, binding on intermediariesIn force 20 February 2026
Qatar, Kuwait, Oman, BahrainNo AI-specific instrument of general applicationNothing regulates agents or machine identity as a categoryVerified negative
Applied AIVoice of Volge22 •

Reference

Glossary
  • Agent. A system given an objective that plans and executes multi-step tasks, calling tools with limited human intervention.
  • AgentOps. The operations discipline integrating observability, evaluation, governance, resilience, and versioning across the agent lifecycle.
  • Agent 365. Microsoft's unified control plane for AI agents across the enterprise, announced at Ignite in November 2025 and generally available from 1 May 2026, licensed per user rather than per agent.
  • A2A (Agent-to-Agent). An open protocol standardising communication, negotiation, and delegation between agents across vendors, contributed by Google and hosted by the Linux Foundation.
  • Autonomous agent. An agent triggered by events or conditions that runs continuously within guardrails, without per-action prompting.
  • Copilot. An assistant operating within an application that responds to explicit prompts and returns control to the user.
  • DPDP. India's Digital Personal Data Protection Act, 2023, and the Rules, 2025, that operationalise it.
  • Entra Agent ID. Microsoft's mechanism assigning a first-class identity to each agent, extending Zero Trust to AI workloads.
  • Foundry. Microsoft's model and tool layer, supporting multiple model providers and MCP-compatible tools.
  • MCP (Model Context Protocol). An open protocol standardising how agents connect to tools and data.
  • Non-human identity (NHI). A machine, service, or agent credential, held by a workload rather than by a person.
  • Zero Trust. A security posture in which no actor, human or agent, is trusted by default and trust is continuously verified.
Applied AIVoice of Volge23 •
Sources

This paper draws on the fol­low­ing, ac­cessed June 2026 and re-veri­fied 8 Novem­ber 2026. Fig­ures are at­trib­uted in-text to their primary source. Where a stat­istic ap­pears across mul­tiple sec­ond­ary out­lets, the most au­thor­it­at­ive primary at­tri­bu­tion is given.

  • McKinsey & Company , The state of AI in 2026: On the road to ROI (August 2026), on agent scaling at large enterprises, reported EBIT impact, and the practices of AI high performers; survey fielded 4 May to 8 June 2026, 1,719 respondents.
  • Deloitte , State of AI in the Enterprise: The Untapped Edge (21 January 2026), on movement of pilots into production and the maturity of agent-governance models; 3,235 leaders across 24 countries.
  • EY (Ernst & Young LLP) , AI Risk and Governance Survey (15 September 2026), on agentic adoption, governance coverage, agent visibility, departures from an organisation's own governance process, and realised AI incidents; 202 senior AI executives, fielded 28 May to 15 June 2026.
  • Gartner , projection that more than 40 per cent of agentic AI projects will be cancelled by the end of 2027 (25 June 2025); Magic Quadrant for AI Agent Management Platforms, inaugural edition (3 November 2026), in which Microsoft is named a Leader.
  • S&P Global Market Intelligence , Voice of the Enterprise research on abandonment of AI initiatives and the loss of proofs of concept before production (March 2025).
  • Microsoft , Copilot and Copilot Studio updates (2026); agentic capabilities in Word, Excel and PowerPoint generally available 22 April 2026; Agent 365 generally available 1 May 2026 at USD 15 per user per month standalone or inside Microsoft 365 E7 at USD 99; Agent Framework 1.0 (3 April 2026); Foundry; Entra Agent ID; Message Center notice MC1395007 (17 June 2026) on Entra Conditional Access for Agents and Entra ID Protection for Agents.
  • Cloud Security Alliance , Agentic Trust Framework (2 February 2026); stewardship transferred outside CSA in May 2026.
  • OWASP , Top 10 for Agentic Applications (9 December 2025, styled the 2026 edition).
  • European Union , AI Act; the Digital Omnibus on AI, Regulation (EU) 2026/1744, adopted 29 June 2026, published in the Official Journal 24 July 2026, in force 27 July 2026; Article 14 human-oversight provisions; Article 50(2) applying from 2 December 2026.
  • Ministry of Electronics and Information Technology (India) , Digital Personal Data Protection Rules, 2025.
  • Linux Foundation , A2A Protocol one-year milestone (9 April 2026), on more than 150 supporting organisations, the 1.0 specification, and enterprise production use.
  • Microsoft 365 Copilot blog and Microsoft Support , the model selector, Auto mode, and the agentic capabilities in Word, Excel and PowerPoint (2026).
Applied AIVoice of Volge24 •

Volge and our expertise on applied AI

Volge builds agent land­scapes that reach audited pro­duc­tion, and the work is en­gin­eer­ing rather than ad­vice. The prac­tice sits where this paper loc­ates the bind­ing con­straint, in the gov­erned data found­a­tion, the iden­tity and policy layer, and the op­er­at­ing model that su­per­vises a fleet. De­liv­ery is by senior en­gin­eers em­bed­ded in the cli­ent's own ten­ant, re­pos­it­or­ies and data es­tate, work­ing along­side the cli­ent's people through to the pro­duc­tion cutover rather than hand­ing a design across.

The work has a usual shape. A read­i­ness pass places the land­scape against the con­trol set in chapter three and the stage pre­requis­ites in chapter four, and re­turns the two or three pro­cesses that match the screen in chapter one. The build then fol­lows the de­pend­ency order. The gov­erned data layer and the ground­ing pipeline come first, then the agent iden­tity and re­gistry, then the con­di­tional-ac­cess and least-priv­ilege con­fig­ur­a­tion, then the audit in­teg­ra­tion and the re­voc­a­tion drill, and only then the first su­per­vised agent. Each in­cre­ment ships against an out­come meas­ured from a baseline taken be­fore the agent ex­is­ted.

The re­gional po­s­i­tion is the part a Gulf or In­dian cli­ent usu­ally needs most. Volge is headquartered in the Gulf, and the dis­tinc­tion between stor­age and pro­cessing set out in chapter five is handled per work­load rather than per ten­ant. The Agent 365 trace loc­a­tion, the Co­pi­lot Stu­dio data loc­a­tion and the model re­gion are es­tab­lished in writ­ing be­fore the ar­chi­tec­ture is fixed. Where no in­stru­ment ad­dresses ma­chine iden­tity, the prac­tice writes the con­trol, maps it to the frame­works an aud­itor will re­cog­nise, and takes it through the cli­ent's own gov­ernance so the de­cision is re­cor­ded.

Who does the work. A de­ploy­ment strategist holds the busi­ness out­come and the gov­ernance ca­dence. A trans­form­a­tion ar­chi­tect holds the tar­get design and the se­quence. Senior en­gin­eers own the work­streams and the build in­side the cli­ent's en­vir­on­ment. The same people stay from the first work­shop to op­er­a­tional han­dover, which is what makes the con­trol tests in chapter four drills the team has run rather than pro­ced­ures it has re­com­men­ded.

An en­gage­ment starts with a scoped as­sess­ment, usu­ally four to six weeks. It pro­duces an in­vent­ory of the agents already run­ning, un­re­gistered and third-party agents in­cluded, a res­id­ency and pro­cessing po­s­i­tion per work­load, a con­trol gap list with the test that closes each one, and a se­quenced plan with the first out­come quan­ti­fied. Volge is a Mi­crosoft part­ner and be­ne­fits com­mer­cially where a cli­ent ex­pands its Mi­crosoft es­tate, which a reader should weigh when read­ing the re­com­mend­a­tions in this paper. The con­ver­sa­tion be­gins wherever the land­scape stands today.

Applied AIVoice of Volge25 •
Version · VVM-AAI-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.