>
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.
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
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.
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.
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.
| Population | Share of enterprises | Measured outcome |
|---|---|---|
| Attributing 5% or more of EBIT to AI, with significant value | 6% | AI high performers; workflows fundamentally redesigned |
| Attributing any positive EBIT impact to AI | 37% | 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 initiatives | 42% | 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.
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
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.
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.
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.
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.
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.
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.
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 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.
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.
| Stage | What it does | Human role | Primary control requirement |
|---|---|---|---|
| Assistive (copilot) | Drafts, summarises, suggests within one app on request | Actor and approver of every output | Content review; data-access scoping |
| Agentic (supervised) | Plans and executes multi-step tasks; calls tools | Reviewer of outcomes; approver of consequential steps | Tool-permission scoping; human-in-the-loop gates; audit logging |
| Autonomous (autopilot) | Triggered by events; runs continuously within guardrails | Supervisor of a fleet; exception handler | Agent identity; least-privilege authority; continuous monitoring; kill-switch and revocation |
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.
| Condition | What it means | How it is tested |
|---|---|---|
| Identity | The agent holds its own first-class identity and appears in the registry | Query the registry for the agent, then confirm that its sign-in record resolves to that identity alone |
| Scope | Tool permissions match the task, and a sub-agent cannot exceed its parent | Attempt a call outside scope and confirm that it is refused and logged |
| Evidence | The agent's actions are logged in a form a reviewer reads | Reconstruct one week of the agent's work from logs alone, without asking the build team |
| Reversal | The agent's authority can be withdrawn across the landscape within minutes | Fire 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.
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.
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.
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.
| Capability | What is required |
|---|---|
| Creating and managing agent identities | Available to all Microsoft Entra customers |
| Operating agents across Microsoft 365 and enterprise workflows | A Microsoft Agent 365 licence per user |
| Conditional Access for agents | Microsoft 365 E7, or Agent 365 paired with Entra ID P1 or Microsoft 365 E3 |
| Risk-based Conditional Access for agents | As above with Entra ID P2 or Microsoft 365 E5; the agent risk condition remains in preview |
| Identity Protection for agents | Agent 365 or Microsoft 365 E7 |
| Identity governance for agents | Microsoft 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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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 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.
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.
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.
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.
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.
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.
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.
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
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.
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.
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.
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.
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
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 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.
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.
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
load, including re-evaluation every time the underlying model changes, and the model changes on the vendor's schedule rather than the enterprise's.
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.
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.
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 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.
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.
| Stage | Entry conditions | Exit evidence |
|---|---|---|
| One, assistive layer and governance foundation | A named agent owner; an agent access policy signed by security; a configured registry; licence entitlements confirmed against the controls the programme intends to use | A measured productivity change in a defined population, and a registry holding every existing agent, third-party agents included |
| Two, supervised agentic value | Two candidate processes screened on volume, rule density and verifiability; a baseline measured before the agent exists; one connector tested against production-shaped data | An audited return in production from one process, with a reviewer able to reconstruct a week of agent actions from logs |
| Three, supervised autonomy and scale | Evaluation evidence from the supervised stage; a dated revocation drill; a teaming model chosen per process | A portfolio supervised as a fleet, with monthly reporting to the agent owner and a dated drill per agent class |
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.
| Control | What it does | Test |
|---|---|---|
| Agent identity | Gives each agent an identity the platform recognises natively, so actions attribute to it alone | Trace one agent action end to end in the sign-in and audit logs, and confirm that no shared account appears |
| Registry entry | Records every agent, in-house and third-party, with its owner and its data reach | Reconcile the registry against a tenant-wide query and against the build teams' own lists, then investigate every difference |
| Tool-level least privilege | Grants each agent only the tool permissions its task needs | Invoke a tool outside the agent's scope and confirm refusal and a log entry |
| Delegation constraint | Stops a sub-agent exceeding its parent's authority | Chain a call through a sub-agent to a resource the parent cannot reach, and confirm that it fails |
| Human gate | Holds a consequential action until a person approves it | Submit a consequential action, confirm that it waits, and confirm that the approval is recorded with the approver's identity |
| Policy evaluation | Evaluates every token the agent acquires against approval state and risk | Run the policy in report-only, read the evaluation record, then enforce and read it again |
| Revocation | Withdraws an agent's authority across the landscape | Fire it against a live agent in a controlled window, then confirm lost access, invalid tokens and a log entry |
| Evaluation harness | Tests behaviour continuously against defined scenarios, because the agent is non-deterministic | Re-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.
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.
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.
The model below locates the enterprise and defines its next increment , the diagnostic counterpart to the roadmap.
| Level | Description | Identity and governance | Teaming | Typical outcome |
|---|---|---|---|---|
| 0 , Ad hoc | Unmanaged use of assistants; shadow AI | None; agents unknown | Undefined | Data-leak risk; no return captured |
| 1 , Assistive | Sanctioned copilots in productivity apps | Human identity only | Human as actor | Modest productivity gain |
| Level | Description | Identity and governance | Teaming | Typical outcome |
|---|---|---|---|---|
| 2 , Supervised agentic | Composed agents in 1-2 processes | Agent registry; Entra Agent ID; HITL gates | Human-in-the-loop | First audited production return |
| 3 , Scaled autonomy | Fleet of autonomous agents across processes | Full Agent 365 control plane; continuous monitoring | Human-on-the-loop | Declared EBIT impact; fleet supervision |
| 4 , Agentic enterprise | Multi-vendor agents coordinating via A2A | Estate-wide governance spanning vendors | Human-out-of-the-loop where reversible | Process-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.
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.
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 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
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.
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.
| Jurisdiction | Azure region | Microsoft 365 Local Region Geography | Foundry project | Copilot Studio in region |
|---|---|---|---|---|
| Saudi Arabia | None | No | No | No |
| United Arab Emirates | Two, one of them restricted | Yes | Yes, UAE North | Yes, UAE North |
| Qatar | One, unpaired | Yes | Yes, Qatar Central | No; data stored in the United States |
| India | Four, two with availability zones | Yes | Yes, two regions | Yes, two regions |
| Kuwait, Oman, Bahrain | None | No | No | No |
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.
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
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.
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 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
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.
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.
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
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.
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.
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 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.
| Platform | General availability | Greatest strength | Principal constraint |
|---|---|---|---|
| Microsoft (Copilot, Copilot Studio, Agent 365, Foundry) | Layered; M365 agentic action GA 22 Apr 2026; Agent 365 GA 1 May 2026 | Identity-led control plane spanning in-house and third-party agents; widest productivity installed base | Strongest value realised within the Microsoft estate |
| Salesforce Agentforce | 2024 | Depth inside the Salesforce CRM ecosystem | Agents outside the CRM boundary require additional integration decisions |
| ServiceNow AI Agents | 2024 | Governed autonomous work where flow runs through ServiceNow (ITSM, security, HR) | Value concentrated in ServiceNow-resident workflows |
| Amazon Bedrock AgentCore | 2025 | Bring-your-own-framework; any framework, model, or protocol; AWS-grade isolation | Higher engineering burden; less out-of-the-box governance |
| Google Gemini Enterprise | 2025 | Integrated partner agents (Box, Workday, and others); full-stack model access | Newer 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.
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
Oversight and accountability
Observability and evaluation
Data protection
Region and residency
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
advice; Volge is not a law firm, and qualified counsel should be engaged for compliance decisions.
| Regime | Provision | Date | Status |
|---|---|---|---|
| EU AI Act | Prohibited-practice provisions | In force since 2 Feb 2025 | Active |
| EU AI Act | Article 50(2) , machine-readable marking and disclosure of synthetic content | Applies from 2 Dec 2026 | Moved from 2 Aug 2026 by Regulation (EU) 2026/1744 |
| EU AI Act | High-risk obligations, Annex III (use-based) | Postponed to 2 Dec 2027 | Deferred by Regulation (EU) 2026/1744, in force 27 Jul 2026 |
| EU AI Act | High-risk obligations, Annex I (product-regulated) | Postponed to 2 Aug 2028 | Deferred by Regulation (EU) 2026/1744, in force 27 Jul 2026 |
| EU AI Act | Article 14 , human oversight of high-risk systems | Tied to high-risk timelines | Substantive standard unchanged by deferral |
| India DPDP | DPDP Rules, 2025, notified by MeitY | Notified 2025; phased | Active; phased commencement |
| India DPDP | Breach notification within 72 hours | Per Rules | Active |
| India DPDP | Significant Data Fiduciary obligations (DPIA, audit, algorithmic-fairness assessment, DPO) | Per Rules | Active for designated SDFs |
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 mode | Selected for | Status, November 2026 |
|---|---|---|
| Auto | Everyday work; Copilot routes the request to the model it judges best suited | The default in the selector |
| GPT-5.6 series | Microsoft's preferred models across Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork since 9 July 2026 | Available, and where Auto generally lands |
| Claude Opus 5 (Anthropic) | Writing, reasoning, and data and analysis; long multi-step work | Added to the selector on 24 July 2026, across Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork and Copilot Studio |
| Think deeper | Slower, deeper reasoning on complex tasks | Available 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.
| Jurisdiction | Instrument | What it does | Status |
|---|---|---|---|
| Saudi Arabia | CST Cloud Computing Services Provisioning Regulations v4, Decision 506/1445, effective 10 October 2023 | Prohibits transfer of Saudi government agency data outside the Kingdom; gates providers by Class A, B or C registration | In force |
| Saudi Arabia | NCA draft AI Cybersecurity Guidelines AICG-1:2026, consulted 5 July to 5 August 2026 | Brings agentic AI into scope and defines AI Agent, Agentic AI and Tool or Plugin | Proposed; no final instrument published |
| Saudi Arabia | SDAIA Generative AI for Government Guidelines, index date 15 May 2025 | Directs government entities to refrain from generative AI in critical decisions affecting individuals | Non-binding guideline |
| United Arab Emirates | DIFC Data Protection Regulations, Regulation 10, 1 September 2023 | Allocates accountability for an autonomous system between Deployer and Operator, in DIFC only | In force |
| United Arab Emirates | Cabinet decision of 23 April 2026 | Commits half of federal government sectors, services and operations to agentic AI within two years | Announced framework; no published instrument text |
| United Arab Emirates | DESC Cloud Service Provider Security Standard | Mandatory certification for providers offering cloud to Dubai government and semi-government entities | In force |
| India | DPDP Rules, 2025, G.S.R. 846(E), notified 13 November 2025 | Notice, security safeguards, breach intimation and Significant Data Fiduciary duties | Phasing; consent managers 13 November 2026, the remainder 13 May 2027 |
| India | SEBI (Intermediaries) (Amendment) Regulations, 2025, regulation 16C | Places sole responsibility for an AI tool's output on the regulated firm, built or bought | In force since early February 2025 |
| India | SEBI CSCRF of 20 August 2024, as amended 28 August 2025 | Required default-deny Zero Trust for critical systems; amended to an IT-committee-approved menu | In force as amended |
| India | IT Amendment Rules, 2026, G.S.R. 120(E) | Obligations on synthetically generated information, binding on intermediaries | In force 20 February 2026 |
| Qatar, Kuwait, Oman, Bahrain | No AI-specific instrument of general application | Nothing regulates agents or machine identity as a category | Verified negative |
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.
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.
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.
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.