The Microsoft ecosystem audit
Contents
About Microsoft
Microsoft, founded in 1975 by Bill Gates and Paul Allen, headquartered in Redmond, Washington, and led by chairman and chief executive Satya Nadella since 2014: is one of the world's most valuable companies (NASDAQ: MSFT). The Microsoft Cloud spans Azure, Microsoft 365, Dynamics 365, Microsoft Security, Microsoft Fabric and GitHub. Volge delivers across that estate through six alliance offerings.
What Volge delivers on Microsoft: six offerings
This capstone spans all six offerings across the Microsoft Cloud.
| Volge offering | What it covers | Anchored on |
|---|---|---|
| Modern Workspace & E7 | Secure modern work on the E7 bundle; human-agent collaboration | Microsoft 365 · Copilot |
| Applied AI | Building, governing and running enterprise AI agents and apps | Azure AI Foundry |
| Zero Trust Security | Identity, threat protection and data governance | Microsoft Entra · Defender · Purview |
| Business Applications | ERP, CRM and low-code, with in-process agents | Dynamics 365 · Power Platform |
| Digital Product Engineering | Autonomous software engineering and CI/CD | GitHub · GitHub Copilot |
| Cloud Infrastructure & Operations | Cloud build, run and cost governance (FinOps) | Microsoft Azure |
Sources: Microsoft corporate and product documentation, 2025–26.
This book: the whole Microsoft Cloud
The Microsoft Cloud is Microsoft's integrated commercial cloud, uniting Azure, Microsoft 365, Dynamics 365, Microsoft Security (Entra, Defender, Purview), Fabric and GitHub on a shared identity, data and security fabric: the full estate this capstone audits across all six Volge offerings.
In the field in 2026
- Azure and Azure AI Foundry, cloud infrastructure plus the platform to build, deploy and govern AI agents
- Microsoft 365 and Copilot, modern work, with Agent 365 for enterprise agent governance
- Microsoft Fabric and OneLake, the unified data foundation, governed by Microsoft Purview
- Microsoft Security: Entra (incl. Entra Agent ID), Defender and Purview for Zero Trust and compliance
Executive summary
The enterprise is being pulled in two directions at once, and the pull is structural rather than a matter of attitude. On one side, agents are now embedded in software the enterprise already owns, and every team that can build one wants to ship it this quarter. On the other side, security, governance, finance, and compliance carry the standing obligation to keep the estate controlled, auditable, and within budget and law. Operational agility and strict enterprise control are not preferences to be reconciled by goodwill. They are forces in tension, and the position an organisation occupies between them is, for the most part, undocumented and unmeasured.
That tension cuts both ways, which is the point most assessments miss. Too much control and the agent programme dies of friction: governance committees stall every deployment, every credential takes a fortnight to provision, and the return that justified the investment never arrives. Gartner projects that more than 40 per cent of agentic AI projects will be cancelled by the end of 2027, attributing the failures to weak governance, integration complexity, and unclear return rather than to the technology (Gartner, 2025). A share of those cancellations is over-control, not under-control. Too little control and the blast radius detonates: an over-permissioned agent reaches data it should never see, a fragmented data estate feeds it contradictions, an unwatched GPU fleet bleeds budget, and a regulator asks for an audit trail that was never written. Non-human identities already outnumber human users by more than 80 to one in large enterprises, and 84 per cent of organisations report that they could not currently pass a compliance audit of agent behaviour or access (CyberArk, 2026; Cloud Security Alliance, 2026).
The central finding of this capstone is the following. The balance between agility and control is not a matter of judgement to be argued in a meeting; it is a measurable position that can be audited across three pillars, identity, data, and spend, under a single compliance overlay, and then remediated against a prioritised plan. An organisation that cannot say where it sits on that balance is not exercising judgement. It is guessing, and the guess is usually optimistic. This document supplies the instrument that replaces the guess: a reusable scorecard, in Appendix A, that locates the estate across the three pillars, scores each dimension from 0 to 4, and resolves to a composite band, exposed, controlled, balanced, or optimised, with a remediation roadmap attached.
The audit anchors in the assessment tooling the Microsoft estate already ships, because for most large enterprises the platform decision is closed and the instruments are paid for. Identity grounds in Microsoft Entra and Microsoft Secure Score. Data grounds in OneLake, Microsoft Fabric, and Microsoft Purview. Spend grounds in Microsoft Cost Management, Azure Advisor, and the Azure Well-Architected Framework. The compliance overlay grounds in Microsoft Purview Compliance Manager, the EU AI Act, and the India Digital Personal Data Protection Rules, 2025. The constraint in every case is configuration and discipline, not procurement.
This capstone synthesises four companion papers in the Voice of Volge series and does not repeat their mechanics. The identity pillar draws on An identity for every agent; the data pillar on The data estate is the accuracy budget; the spend pillar on The meter is always running; and the compliance overlay on Oversight is the enabler, not the brake. Where those papers build the controls, this one measures whether they are in place.
Five conclusions are advanced, each developed in the chapters that follow.
First, the balance point is the unit of assessment, and most organisations cannot locate it. Agility and control trade against each other along a curve, and the productive position sits at neither extreme. An organisation that has never measured its position defaults to whichever force is louder in its culture, and pays for the imbalance in either cancelled projects or uncontrolled exposure.
Second, identity is the pillar that gates the other two, and it is the one most often skipped. An agent without a first-class identity cannot be inventoried, scoped, audited, or revoked, which means its data access and its spend cannot be attributed either. Microsoft Entra Agent ID reached general availability in April 2026 and Microsoft Agent 365 on 1 May 2026; the gap in most estates is adoption, not availability (Microsoft, 2026).
Third, data architecture sets the ceiling on agent value and a floor under compliance risk. A fragmented, ungoverned estate caps agent accuracy and, simultaneously, exposes the organisation to residency and lineage failures a regulator will test. Gartner projects organisations will abandon 60 per cent of AI projects unsupported by AI-ready data through 2026 (Gartner, 2025).
Fourth, cloud spend is the pillar where imbalance is most visible and least governed. The AI bill is volatile by construction, agentic workloads consume 5 to 30 times the tokens of a chatbot, and 73 per cent of enterprises report AI costs exceeding projection (Gartner, 2026; FinOps Foundation, 2026). The waste, between a quarter and a half of an unoptimised bill, is the cost of the control that was never applied.
Fifth, compliance is not a fourth pillar but an overlay that binds all three. The EU AI Act Article 14 human-oversight obligation, the India DPDP Rules, and Microsoft Purview Compliance Manager attach to identity, data, and spend at once. An audit that scores the three pillars without the overlay measures efficiency and misses lawfulness.
Volge's position is plain. The platform is settled, the controls are available, and the failure is one of measurement and discipline. The organisations that audit their position, remediate against a prioritised plan, and re-audit on a cadence are the ones that hold agility and control together. The rest oscillate between the two and pay for the oscillation. This guide is the instrument for the first path.
Chapter one: The agility-versus-control tension, and why an audit
The takeaway
Operational agility and strict enterprise control trade against each other along a curve, and the productive position sits at neither end. Too much control kills adoption and the return that justified it; too little control creates security blast radius, runaway cost, and compliance exposure. The position an organisation occupies on that curve is measurable, and measuring it is the precondition for moving it. An audit is the only honest way to locate the balance point, because culture and self-report both bias the answer toward the louder force.
2.1 The tension is real, and it is bidirectional
The reflex framing treats control as the enemy of agility, every gate a tax on speed, every policy a brake on the team that wants to ship. That framing is half right and dangerously incomplete, because it names only one of the two failure modes.
Run with too little control and the failures are familiar. An agent handed broad access reaches data it should never see: 80 per cent of organisations report their agents have already acted beyond their intended scope (industry survey data, 2026). A fragmented data estate feeds the agent three conflicting copies of one record, and it answers fluently and wrongly. A forgotten GPU fleet runs at 5 per cent utilisation while it bills the other 95 (Rack2Cloud, 2026). A regulator asks how a decision was reached and the audit trail does not exist. Each is a control gap with a measurable cost.
Run with too much control and the failures are quieter and no less expensive. A monthly governance committee becomes the rate limiter on every deployment. A weeks-long credential process trains teams to route around it through shadow AI. An approval gate surfacing ten thousand items a day trains its reviewers to rubber-stamp. The return that underwrote the programme arrives late or not at all, and the project joins the cancellation pool, not because the technology failed, but because the control killed the value before it could be captured. A share of Gartner's projected 40-per-cent-plus cancellation rate is exactly this: control without calibration (Gartner, 2025).
The tension is therefore a curve, not a slider with a single safe end. The productive position is where control is sufficient to bound the blast radius and no heavier, where the marginal control pays for itself in risk reduced rather than value destroyed. That position is specific to the organisation, its data, and its regulation. It cannot be assumed; it has to be found.
2.2 The balance, drawn as a quadrant
The two forces resolve into four positions, and naming them fixes the target the audit aims at.
The target is the balanced quadrant, and the three other quadrants reach it by different routes. The exposed organisation adds calibrated control to the three pillars without slowing adoption. The paralysed organisation cuts control back to what risk actually justifies, replacing committee approval with engineered guardrails. The dormant organisation adopts with control built in from the start, skipping the exposed and paralysed detours both. The audit's first job is to say which quadrant the organisation is in, by pillar, because the remediation differs by quadrant.
2.3 Why self-assessment fails, and why an instrument is needed
Ask a leadership team where it sits on the balance and the answer reflects culture, not evidence. A security-led organisation reports itself controlled and is often paralysed in places it cannot see. A product-led organisation reports itself agile and is often exposed in pillars it never looks at. Self-report is biased toward the louder force, and the bias is systematic rather than random.
An instrument removes the bias by scoring observable dimensions against defined anchors. The question is not "do we feel in control of agent identity" but "does every agent hold a first-class Entra Agent ID identity, recorded in the Agent 365 registry, and can the kill-switch be demonstrated against a live agent": a question with a verifiable answer and a 0-to-4 score. Across enough dimensions, the scores aggregate into a position that culture cannot flatter. That is what the scorecard in Appendix A supplies, and it is why this document is an assessment guide rather than an explainer.
2.4 The three pillars, and the overlay that binds them
The estate is audited across three pillars because those are the three surfaces on which the agility-control tension actually plays out, and a compliance overlay sits across all three because lawfulness is not separable from any one of them.
| Pillar | What it governs | Agility risk (too much) | Control risk (too little) | Microsoft anchor |
|---|---|---|---|---|
| Identity and access | Who and what reaches the estate | Provisioning friction; shadow agents route around gates | Over-permissioned, untracked, immortal agents; rogue exposure | Entra; Secure Score; Agent 365 |
| Data architecture | What the agent grounds on | Ungoverned access slows nothing but exposes everything | Fragmented, stale data caps accuracy; residency and lineage failures | OneLake; Fabric; Purview |
| Cloud spend | What the estate costs | Cost gates throttle experimentation | Volatile, attributed-to-nobody spend; 27%+ waste | Cost Management; Azure Advisor; Well-Architected |
| Compliance overlay | Whether all three are lawful | Compliance theatre slows every deployment | Undeployable workloads; failed audits; enforcement exposure | Purview Compliance Manager; EU AI Act; DPDP |
The pillars are not independent. Identity gates the other two: spend cannot be attributed and data access cannot be scoped without a first-class identity to attach them to. Data sets the ceiling on agent value and a floor under compliance risk. Spend is where imbalance shows first and governs last. The overlay binds all three, because the same EU AI Act and DPDP obligations attach to who acts, what they touch, and what it costs to run. The chapters that follow audit each pillar in turn, weaving the overlay through each, and chapter five composes the four scores into one position.
2.5 The implication for the audit
The balance between agility and control is the unit of assessment, and it is measurable across three pillars under one overlay. An organisation that has not measured its position is not balanced by default; it is exposed or paralysed in ways its culture hides from it. The instrument that locates the position is the scorecard, and the chapters that follow build the assessment dimension by dimension. The audit does not begin with technology. It begins with an honest measurement of where control is too light, too heavy, or correctly sized, and that measurement is the subject of the rest of this guide.
Chapter two: Pillar one: the identity and access security audit
The takeaway
Identity is the pillar that gates the other two, and it is the one most often skipped. An agent without a first-class identity cannot be inventoried, scoped, audited, or revoked, which means its data access and its spend cannot be attributed either. The identity audit measures whether every agent is counted, scoped to least privilege, evaluated by policy, logged, and revocable, and it anchors in Microsoft Entra and Microsoft Secure Score. The companion paper An identity for every agent builds these controls in full; this audit measures whether they are in place.
3.1 Why identity is audited first
The agility-control tension is sharpest at the identity boundary, because identity is where agility most tempts the shortcut and control most depends on the discipline. A team shipping an agent grants it broad access to move fast; the security function needs that access scoped to bound the blast radius. That gap is the most common and consequential imbalance in the estate.
Identity also makes the other two pillars auditable: a data-access policy and a spend attribution both attach to an identity. An agent with no first-class identity is invisible to the data audit and the spend audit alike, because there is nothing to attach the policy or the cost to. Audit identity first, or the other two pillars score against a population the estate cannot see.
The scale of the exposure is established in the companion paper and not re-argued here: non-human identities outnumber human users by more than 80 to one, 68 per cent of identity-related incidents now involve machine identities, and 84 per cent of organisations doubt they could pass an audit of agent behaviour (CyberArk, 2026; Cloud Security Alliance, 2026; industry research, 2026). The audit's job is to convert that population into a counted, scoped, governed inventory.
By the numbers: the identity audit gap
- 80: 1: non-human to human identity ratio in large enterprises, the population the identity pillar must count (CyberArk, 2026).
- 84% of organisations could not currently pass a compliance audit of agent behaviour or access (Cloud Security Alliance, 2026).
- April 2026: Entra Agent ID general availability; 1 May 2026: Agent 365 general availability. The control plane is shipped; the gap is adoption (Microsoft, 2026).
3.2 What the identity audit measures
The identity audit scores six dimensions, each a step the breach pipeline of the companion paper either breaks or leaves open. Each maps to a Microsoft capability and a 0-to-4 maturity anchor in Appendix A.
Inventory and identity. Does every agent, in-house, third-party, and previously shadow, hold a first-class Entra Agent ID identity and appear in the Agent 365 registry? An estate that registers only Microsoft-built agents leaves its largest exposure unmeasured. Good looks like a complete, attributable inventory; the failing anchor is shared keys and human credentials with no registry at all.
Least privilege. Is access scoped at the tool call and the credential, evaluated per task, or granted once at deployment and forgotten? The least-privilege practice binds permission to each tool the agent invokes and constrains inheritance down delegation chains so a sub-agent cannot exceed its parent.
Conditional access. Does Conditional Access evaluate every token an agent identity acquires, scoping policy by object identifier, custom security attribute, or blueprint, and blocking agents that Identity Protection rates high-risk? The two canonical policies, block every agent except those carrying an approved attribute, and block any high-risk agent, are the baseline.
Credential lifecycle. Are credentials just-in-time, purpose-bound, and auto-expiring, or standing secrets that outlive the task? The standing secret is the durable exposure; auto-expiry closes it without relying on anyone to remember.
Audit trail. Is every authentication, token acquisition, tool invocation, and policy decision logged with a unique agent identifier, in a form that answers what was done, on whose behalf, under what policy, and why? This is the dimension 84 per cent of organisations cannot currently clear (Cloud Security Alliance, 2026).
Revocation. Has the kill-switch been fired against a live agent in a controlled window, confirming loss of access, token invalidation, and logging, or is revocation an untested assumption? An untested kill-switch is a hope, and the breach is the one test no one wants to run first.
3.3 Grounding the audit in Microsoft Secure Score
The identity audit does not start from a blank page, because Microsoft already publishes a quantified posture measure. The Identity Secure Score is the identity portion of Microsoft Secure Score, expressed as a percentage that indicates alignment with Microsoft's recommended identity controls; it is reached through the Microsoft Entra admin centre under recommendations, filtered to the Security set (Microsoft Learn, 2026). In 2026 Microsoft recategorised several recommendations previously grouped under cloud apps as identity-related, tightening the identity score's coverage while leaving the total Secure Score unchanged (Microsoft Learn, 2026).
The Secure Score is a useful starting baseline and an insufficient finishing one, and the distinction matters for the audit. Secure Score measures the configuration of the human and tenant identity estate well. It does not, on its own, measure the agent-specific dimensions of section 3.2, delegated-versus-own-access classification, tool-level least privilege, tested revocation, which sit in the Entra Agent ID and Agent 365 surfaces. Read the Secure Score as the foundation layer of the identity pillar, then score the agent-specific dimensions on top of it. The audit's identity score is the composite, not the Secure Score alone.
3.4 The control-gap-to-risk mapping
Each unmet identity dimension maps to a specific, named risk. The mapping is what converts a low score into a remediation priority.
| Control gap | Risk created | Blast radius | Primary remediation |
|---|---|---|---|
| No first-class identity / no registry | Agent uninventoried; cannot be governed at all | Whole estate invisible; rogue exposure | Entra Agent ID; Agent 365 registry |
| No least privilege at tool call | Over-permissioned agent reaches excess data | Union of all tool permissions, often vast | Tool-scoped permissions; delegation constraint |
| No conditional access on tokens | Compromised or unapproved agent reaches resources | Every resource the agent can request | Two canonical CA policies; report-only first |
| Standing credentials | Immortal credential outlives the task | Indefinite; persists after task ends | Just-in-time, auto-expiring credentials |
| No agent-specific audit trail | Cannot attribute or reconstruct agent action | Failed compliance audit; uninvestigable breach | Unique agent ID logging; Purview integration |
| Untested revocation | Cannot stop a misbehaving agent on demand | Continuous until manual containment | Test kill-switch against a live agent |
3.5 The identity audit flow
The diagram traces how an estate is assessed against the six dimensions, producing a pillar score and a remediation list rather than a pass/fail.
3.6 The implication for the audit
The identity pillar gates the other two, and it is scored on whether every agent is counted, scoped, evaluated, logged, and revocable. Microsoft Secure Score supplies the foundation baseline; the agent-specific dimensions in Entra Agent ID and Agent 365 complete it. A low identity score is the highest-priority remediation in the composite, because an uncounted agent is invisible to the data and spend audits that follow. The data pillar, which the identity-scoped agent grounds on, is chapter three.
Chapter three: Pillar two: the data architecture audit
The takeaway
Data architecture sets the ceiling on agent value and a floor under compliance risk at the same time. A fragmented, ungoverned estate caps agent accuracy and exposes the organisation to residency and lineage failures a regulator will test. The data audit measures whether the estate is unified, curated, governed, current, and residency-controlled, and it anchors in OneLake, Microsoft Fabric, and Microsoft Purview. The companion paper The data estate is the accuracy budget builds the estate in full; this audit measures whether the agent grounds on curated truth or on whatever is lying around.
4.1 Why the data pillar carries two risks at once
The data pillar is where the agility-control tension is least visible and most expensive, because ungoverned data slows nothing. An agent grounded on a fragmented estate runs exactly as fast as one grounded on a curated estate; it simply answers wrongly, and the wrongness arrives at machine speed and scale. The absence of control here does not show up as friction. It shows up as confident, untraceable error and, separately, as a compliance exposure no one priced.
That dual character is what the data audit has to capture. A fragmented estate caps accuracy: Gartner projects organisations will abandon 60 per cent of AI projects unsupported by AI-ready data through 2026, and 63 per cent of organisations either lack the data-management practices AI requires or cannot say whether they have them (Gartner, 2025). The same fragmentation creates compliance risk: an agent that retrieves across ungoverned boundaries cannot prove residency, cannot show lineage, and cannot demonstrate that the requesting user was entitled to the data it surfaced. Accuracy and compliance fail together, from the same root, and the audit scores both.
By the numbers: the data audit gap
- 60% of AI projects unsupported by AI-ready data are projected for abandonment through 2026 (Gartner, 2025).
- 63% of organisations either lack the data-management practices AI requires or cannot say whether they have them (Gartner, 2025).
- The data pillar scores accuracy and compliance against the same six dimensions, unification, curation, quality, lineage, governance, and residency, because in this pillar they are one estate seen from two angles.
4.2 What the data audit measures
The data audit scores six dimensions, mirroring the AI-readiness dimensions of the companion paper and adding the compliance-overlay tests the regulated estate requires. Each maps to a Microsoft capability and a 0-to-4 anchor in Appendix A.
Unification. Is the estate consolidated into one logical lake, or do agents retrieve from conflicting siloed copies? OneLake collapses the copy sprawl that manufactures contradiction. Good looks like a single authoritative representation per core entity; the failing anchor is the same customer recorded inconsistently across a dozen systems.
Curation. Are agents grounded on a curated gold layer, or on raw ingest? The medallion architecture, bronze to silver to gold, produces AI-ready data as a property of how it is built. Grounding on bronze inherits every upstream defect.
Measured quality. Is data quality scored by domain, by system, and by business rule, or merely asserted in a policy binder? Readiness is demonstrated by scorecards, not documented by policy, and Microsoft Purview supplies the scorecards.
Lineage. Can every agent answer be traced through the gold model, through the silver transformations, to the bronze source? An answer that cannot be traced cannot be defended, regardless of whether it is correct. Purview tracks lineage continuously across the Fabric estate.
Governance and access. Does classification and access policy follow the data through its lifecycle, so an agent cannot surface records the requesting user is not entitled to see? Where policy does not follow the data, the agent becomes a tunnel around the access controls: among the most serious risks an ungoverned deployment carries.
Recency and residency. Is current data available to ground on where staleness produces wrong answers, through Fabric Real-Time Intelligence, and does regulated data stay inside its permitted residency boundary across storage, inference, and logging? Residency is the compliance-overlay test on the data pillar, and the workload inherits the sovereignty level of the most sensitive data it touches.
4.3 The data pillar as the compliance fault line
The compliance overlay bites hardest on the data pillar, because data is what the regulation is mostly about. The EU AI Act's traceability requirements and the India DPDP Rules' obligations on accountable fiduciaries both attach to how data is handled, not merely to how a decision is made. Three data-pillar tests are therefore compliance tests as much as accuracy tests.
Lineage is a compliance control: the EU AI Act requires institutions deploying high-risk systems to maintain reconstructible decision logs, and an agent answer that cannot be traced to its source cannot satisfy that. Access policy is a compliance control: the DPDP Rules hold the organisation accountable as a data fiduciary even when a decision is automated, which means the agent's retrieval must respect the same purpose and consent boundary that would constrain the user. Residency is a compliance control: regulated personal data processed outside its permitted boundary is unlawful, and the boundary covers the data that transits the model and the data written to the audit log, not just storage at rest.
The data audit therefore scores accuracy and compliance against the same six dimensions, because in this pillar they are the same dimensions seen from two angles. An estate that scores well on unification, curation, quality, lineage, governance, and residency is both accurate and defensible. An estate that scores poorly is neither.
4.4 The data-pillar control-gap-to-risk mapping
4.5 Grounding the audit in Microsoft Purview
The data audit grounds in Microsoft Purview, which supplies the measured quality, continuous lineage, classification, and policy that the six dimensions test for. Purview tracks data automatically as it moves through lakehouses, pipelines, and AI workloads, so lineage is continuous rather than reconstructed after an incident, and its data-observability capabilities surface root-cause issues that support regulatory reporting (Microsoft, 2025). The quality scorecards are the evidence the audit scores against: a domain whose gold layer carries a measured, visible quality score clears the readiness threshold; a domain whose quality is asserted in policy does not.
The residency dimension grounds in the Microsoft Sovereign Cloud mechanics established in the compliance companion: data residency, customer-managed keys, in-country inference, and disconnected operation, assembled to the level the data classification demands. The audit scores whether the residency level matches the most sensitive data the workload touches, over-provisioned residency wastes budget, under-provisioned residency is unlawful.
4.6 The implication for the audit
The data pillar is scored on unification, curation, measured quality, lineage, governance, and residency: the same six dimensions seen as accuracy controls and as compliance controls at once. A fragmented estate caps agent value and fails the regulator simultaneously, from one root. The remediation is the modernisation roadmap of the companion paper, sequenced so the estate is reformed before agents ground on it. The spend pillar, which prices the estate and the agents that run on it, is chapter four.
Chapter four: Pillar three: the cloud spend audit, with the compliance overlay
The takeaway
Cloud spend is the pillar where imbalance is most visible and least governed. The AI bill is volatile by construction, agentic workloads burn 5 to 30 times the tokens of a chatbot, and reserved GPUs bill whether they compute or idle, and the waste, between a quarter and a half of an unoptimised bill, is the cost of the control that was never applied. The spend audit measures whether spend is visible, attributed, optimised, and committed on evidence, and it anchors in Microsoft Cost Management, Azure Advisor, and the Azure Well-Architected Framework. This chapter also closes the compliance overlay across all three pillars.
5.1 Why spend is the most legible imbalance
Spend is the pillar where the cost of imbalance is denominated in currency, which makes it the easiest to measure and, paradoxically, the most often left ungoverned. An over-permissioned agent's blast radius is hard to price until it detonates; a fragmented data estate's accuracy cost is hard to attribute. An idle GPU fleet's cost arrives on an invoice every month, itemised and ignored.
The volatility is established in the companion paper and not re-argued: Gartner forecasts worldwide AI spending of USD 2.59 trillion in 2026, a 47 per cent year-on-year increase, with agentic workloads consuming 5 to 30 times the tokens of a standard chatbot, and 73 per cent of enterprises reporting AI costs exceeding projection (Gartner, 2026; FinOps Foundation, 2026). The waste is measured, not estimated: GPU fleets observed at 5 per cent utilisation, roughly 27 per cent of cloud spend lost to waste generally (Rack2Cloud, 2026; Spendark, 2026). The spend audit converts that volatility into a scored position and a remediation list.
By the numbers: the spend audit gap
- USD 2.59 trillion projected worldwide AI spend in 2026, up 47% year on year (Gartner, 2026).
- 5–30× more tokens per task for agentic workloads versus a standard chatbot, the multiplier the audit must make visible (Gartner, 2026).
- 73% of enterprises report AI costs exceeding original projection; 20–30% compute reduction is available from Azure Advisor right-sizing alone (FinOps Foundation, 2026; Microsoft Learn, 2026).
5.2 What the spend audit measures
The spend audit scores six dimensions, mirroring the FinOps-for-AI maturity of the companion paper. Each maps to a Microsoft capability and a 0-to-4 anchor in Appendix A.
Visibility and attribution. Is AI spend tagged to a team, a project, a model, and a use case, or is it one aggregate bill with no owner? An untagged cost cannot be attributed, and an unattributed cost cannot be governed. The five-dimension tagging standard is the baseline.
Token telemetry. Does telemetry meter at the step level, so the agent multiplier is visible rather than buried in a request-level total? A cost report that logs only the user request understates the bill by the full agent multiplier.
Model right-sizing. Are models routed so simple work reaches cheaper models, or does every request default to the most expensive model on hand? Right-sizing is the largest single lever on the token axis.
Discount discipline. Are prompt caching and the Batch API applied to every qualifying workload, given that the discounts stack? Caching discounts the repeated context an agent racks up most; the Batch API hands over a flat discount for tolerating delay.
Commitment discipline. Are provisioned-throughput and GPU commitments validated against 30 to 60 days of real telemetry and reviewed quarterly, or set on forecast and left untouched? A reservation placed below its utilisation threshold costs more than the pay-as-you-go it replaced.
Unit economics. Is spend governed on cost per unit of business value, or merely watched against an absolute budget? Unit economics is the measure that distinguishes a profitable workload from a costly one.
5.3 Grounding the audit in Azure Advisor, Cost Management, and Well-Architected
The spend audit grounds in three Microsoft instruments, each measuring a different part of the position.
Azure Advisor supplies the optimisation baseline. It analyses historical usage to identify idle and underutilised resources and recommends right-sizing, and implementing its recommendations can reduce Azure compute costs by 20 to 30 per cent (Microsoft Learn, 2026; EPC Group, 2026). Advisor's recommendations are the spend audit's ready-made evidence for the right-sizing and commitment dimensions.
Microsoft Cost Management supplies the visibility and attribution spine. It carries cost analysis, budgets, and exports, and its FinOps lifecycle maps onto the audit dimensions directly: visibility, allocation, optimisation, forecasting, and governance (Microsoft Azure, 2026; nOps, 2026). The audit scores the tagging and showback dimensions against what Cost Management can actually attribute.
The Azure Well-Architected Framework supplies the structured review. Its cost-optimisation pillar, one of five, alongside reliability, security, operational excellence, and performance efficiency, is assessed through the Well-Architected Review, a self-assessment of roughly 60 questions that returns prioritised recommendations with links to supporting guidance (Microsoft Learn, 2026; Microsoft Community Hub, 2026). The Well-Architected Review is the closest Microsoft analogue to this guide's own instrument, and the spend audit borrows its cost-pillar questions as a starting set.
5.4 The spend optimisation decision, as an audit lens
The companion paper's optimisation flow doubles as an audit lens: each branch the estate fails to take is a scored gap.
5.5 The compliance overlay, closed across all three pillars
Compliance is not a fourth pillar audited separately; it is an overlay that attaches to identity, data, and spend at once, and it closes here because the spend pillar is the last of the three. Three regimes drive it, and each binds more than one pillar.
The EU AI Act binds identity and data. Its Article 14 human-oversight obligation requires high-risk systems to be designed so a natural person can effectively oversee them, an identity-and-oversight control, and its traceability requirements bind the data pillar's lineage and audit-log dimensions. The Digital Omnibus provisional agreement of 7 May 2026 deferred the high-risk obligations: stand-alone Annex III systems move from 2 August 2026 to 2 December 2027, and Annex I product-embedded systems to 2 August 2028 (Gibson Dunn, 2026; Hogan Lovells, 2026). The deferral buys planning time and does not soften the substantive standard, and the enforcement cycle for provisions already in force is running through 2026. The audit treats the deferral as runway, not relief.
The India DPDP Rules, 2025 bind data and identity. They hold the organisation accountable as a data fiduciary even when a decision is automated, and impose enhanced obligations on Significant Data Fiduciaries, annual data-protection impact assessment, independent audit, algorithmic-fairness assessment, and a data protection officer (MeitY, 2025). Those obligations attach to how the agent's identity is scoped and how the data estate is governed, scored in pillars one and two.
Microsoft Purview Compliance Manager binds all three and supplies the overlay's quantified measure. It scores compliance posture from 0 to 900, calculated as points achieved from completed improvement actions over total achievable points, and ships over 320 ready-to-use regulatory assessment templates (Microsoft Learn, 2026; Microsoft Security, 2026). The baseline score reflects Microsoft-managed controls already in place, typically 30 to 40 per cent before any customer action, with the remainder requiring customer remediation (EPC Group, 2026). The Compliance Manager score is the overlay's headline number, and the audit reads it alongside the three pillar scores rather than instead of them.
| Regime / instrument | Identity pillar | Data pillar | Spend pillar | Audit test |
|---|---|---|---|---|
| EU AI Act, Art. 14 | Human oversight; agent identity | : | : | Oversight gate by risk class; override built |
| EU AI Act, traceability | Audit trail | Lineage; reconstructible logs | : | Reconstructible decision logs to WORM standard |
| India DPDP Rules 2025 | Scoped identity; accountability | Purpose-bound access; residency | : | Fiduciary accountability under automation; SDF assessments |
| Purview Compliance Manager | Identity controls scored | Data controls scored | Governance controls scored | 0–900 improvement score; customer actions remediated |
5.6 The implication for the audit
The spend pillar is scored on visibility, telemetry, right-sizing, discount discipline, commitment discipline, and unit economics, grounded in Azure Advisor, Cost Management, and the Well-Architected Review. The compliance overlay attaches to all three pillars at once and is measured by Purview Compliance Manager's 0-to-900 score read alongside them. The three pillar scores and the overlay now compose into a single position, which chapter five resolves into a composite band and a remediation roadmap.
Chapter five: The composite assessment, the remediation roadmap, and the maturity model
The takeaway
The three pillar scores and the compliance overlay compose into a single position on the agility-control balance, expressed as a band, exposed, controlled, balanced, or optimised. The composite is not an average; the lowest pillar caps the position, because identity, data, and spend gate each other. The remediation roadmap sequences the fixes by blast radius and dependency, identity first, and the maturity model lets leadership locate the estate and name the next defensible increment. This chapter assembles the instrument the appendices then make runnable.
6.1 How the four scores compose
The audit produces four scores, identity, data, spend, and the compliance overlay, and they do not simply average, because the pillars are not independent. The lowest pillar caps the composite, for the reason chapter one established: identity gates data and spend, data sets the ceiling on value and a floor under compliance, and an organisation that scores well on two pillars and poorly on a third is exposed at the weak point regardless of the strong ones.
The composition rule is therefore floor-weighted, not mean-weighted. An estate scoring 4 on spend and data but 1 on identity is not a 3; it is closer to a 1, because its uncounted agents make the data and spend governance unattributable. The composite band reads from the lowest pillar up, with the overlay applied as a gate: a strong three-pillar position that fails the compliance overlay is not balanced, because an efficient estate that cannot pass an audit is undeployable in regulated scope.
6.2 The composite bands
The composite resolves into four bands, defined in full in Appendix B and summarised here. The bands map onto the quadrants of chapter one, with the exposed and paralysed quadrants both resolving to low bands by different routes.
Exposed. One or more pillars at 0–1; agents operate without first-class identity, on fragmented data, with untracked spend, and the compliance overlay fails. The blast radius is uncontrolled and accumulating. This is the priority-one remediation state.
Controlled. Pillars at 2; the foundations are in place, agents counted, data unifying, spend visible, but control is uneven and not yet calibrated. The estate is audit-approaching but not audit-ready, and the compliance overlay partially clears.
Balanced. Pillars at 3; control is calibrated to risk across all three pillars, the overlay clears, and agility is preserved rather than throttled. This is the target band, where agents ship under bounded risk and value is captured.
Optimised. Pillars at 4; the estate is governed end-to-end, control is continuously re-tuned on evidence, and the agility-control balance is maintained as a managed variable rather than rediscovered each quarter.
6.3 The remediation roadmap
The remediation roadmap sequences the fixes by dependency and blast radius, and the sequence is forced: identity first, because it gates the other two; data second, because it sets the ceiling on value and the floor under compliance; spend third, because it cannot be attributed until identity is in place; and the compliance overlay woven through all three rather than bolted on at the end.
Stage one, the first 90 days: the identity foundation. Stand up Entra Agent ID, build the Agent 365 registry, onboard every agent including third-party and shadow agents, and deploy the two canonical conditional-access policies in report-only mode before enforcement. The objective is a counted, scoped estate, because nothing downstream can be attributed without it. No autonomous agent operates outside the registry past this stage.
Stage two, months three to nine: data and the compliance overlay. Unify sources into OneLake, build the medallion layers, stand up Purview for lineage, measured quality, classification, and policy, pin residency to the level the data classification demands, and complete the first data-protection impact assessment where Significant Data Fiduciary obligations apply. The objective is an estate the agent can ground on accurately and the organisation can defend to a regulator.
Stage three, months nine to eighteen: spend and optimisation. Apply the five-dimension tagging standard, stand up step-level token telemetry, route models for right-sizing, claim the caching and Batch discounts, validate every commitment against 30 to 60 days of telemetry, and govern on unit economics. The objective is spend converted from a volatile surprise into a managed, attributed variable.
Steady state: re-audit on cadence. Recompose the four scores quarterly, because consumption, data, and the threat surface all shift continuously, and a control calibrated once decays as the estate moves under it. The audit is a cycle, not a project.
6.4 The composite maturity model
The maturity model lets leadership locate the estate across all three pillars and the overlay at once, and name the next increment. It is the diagnostic counterpart to the roadmap, and it composes the pillar-specific maturity models of the four companion papers into one.
| Level | Identity | Data | Spend | Compliance overlay | Composite band |
|---|---|---|---|---|---|
| 0: Ungoverned | Shared keys; no registry | Fragmented; ungoverned | Untagged; discovered after the fact | Cannot pass an audit | Exposed |
| 1: Foundational | First-class identity; registry | Unifying into OneLake | Tagging and showback | Baseline Microsoft-managed controls only | Exposed → Controlled |
| 2: Controlled | Conditional access; least privilege | Medallion gold; Purview lineage | Telemetry; levers applied | Partial customer remediation; DPIA begun | Controlled |
| 3: Balanced | Risk-adaptive; kill-switch tested | Governed; residency-matched; current | Commitments on telemetry; unit economics | Overlay clears; assessments current | Balanced |
| 4: Optimised | Estate-wide, spanning vendors | AI-ready end-to-end | Continuously re-optimised | Continuous assurance; automated assessment inputs | Optimised |
Most large enterprises sit between Level 0 and Level 1 across all three pillars as of mid-2026, consistent with the identity, data, and spend companion findings. The audit-ready and deployable-in-regulated-scope threshold is Level 2; the target balanced position is Level 3.
6.5 The implication for the audit
The composite is floor-weighted and overlay-gated: the lowest pillar caps the position, and a failed compliance overlay caps the band at controlled regardless of pillar strength. The remediation roadmap is sequenced identity-first by dependency, not by preference. The maturity model composes the four companion models into one diagnostic. What remains is the instrument itself, the scorecard, the scoring rubric, and the remediation-priority matrix, which the appendices supply in runnable form.
Chapter six: The leadership agenda
Operational agility and strict enterprise control are not opponents to be chosen between; they are forces to be balanced, and the balance is measurable. The evidence assembled in this guide resolves into a short list of conclusions for the chief information officer, the chief information security officer, the enterprise architect, and the governance lead.
The balance point is a position, not an attitude, and most organisations cannot locate it because they have never measured it. Self-report is biased toward whichever force the culture favours: security-led estates report themselves controlled and are often paralysed in places they cannot see; product-led estates report themselves agile and are often exposed in pillars they never examine. The audit replaces the guess with a scored position across identity, data, and spend, under a compliance overlay, and that position is the only honest basis for action.
The three pillars gate each other, which fixes the remediation sequence. Identity is the foundation, because an uncounted agent is invisible to the data and spend audits and ungovernable by either. Data sets the ceiling on agent value and the floor under compliance risk, failing both from one root when it is left fragmented. Spend is where imbalance is most legible and least governed, denominated in currency on a monthly invoice and ignored anyway. The compliance overlay binds all three, and a strong three-pillar position that fails the overlay is undeployable in regulated scope regardless of its efficiency.
Six questions belong on the leadership agenda. First, can the organisation produce a single, evidenced position on the agility-control balance, by pillar, or does it rely on the self-report its culture flatters? Second, does every agent hold a first-class identity in the Agent 365 registry, so its data access and its spend can be attributed at all? Third, does the agent ground on a curated, governed, residency-controlled estate, or on whatever the retrieval layer surfaces? Fourth, is AI spend tagged, metered at the step level, and governed on unit economics, or watched as an aggregate bill with no owner? Fifth, does the compliance overlay clear, does the Purview Compliance Manager score, the EU AI Act oversight posture, and the DPDP fiduciary accountability all hold across the three pillars? Sixth, is the audit run on a cadence, recomposing the score quarterly, or was it a one-time exercise that has already decayed?
Volge's work is in the execution. The platform decision is, for most enterprises, already closed by existing Microsoft commitments, and the assessment instruments, Secure Score, Purview Compliance Manager, Cost Management, Azure Advisor, the Well-Architected Review, are already paid for. The disciplines that convert those instruments into a measured, balanced, defensible estate, the composite scorecard, the floor-weighted scoring, the identity-first remediation sequence, the quarterly re-audit, are where programmes either hold the balance or oscillate between exposure and paralysis. That is where Volge concentrates. We would welcome the chance to run this audit against a specific Microsoft estate in detail.
Appendices
Appendix A: The audit scorecard
The scorecard is the instrument this guide exists to supply. Run it against the estate, pillar by pillar, scoring each dimension from 0 to 4 against the anchors below. The pillar score is the lowest of its dimensions where a dimension is a dependency for the others, or the mean of its dimensions otherwise; Appendix B sets out the composition. Score on evidence, not on intention: a dimension scores at the level the organisation can demonstrate, not the level it plans to reach.
The 0-to-4 anchor scale is common across all three pillars:
- 0: Absent. The control does not exist; the estate is exposed on this dimension.
- 1: Ad hoc. The control exists in places, inconsistently, with no standard.
- 2: Defined. The control is standardised and applied, but not yet enforced or measured estate-wide.
- 3: Managed. The control is enforced, measured, and calibrated to risk across the estate.
- 4: Optimised. The control is continuously re-tuned on evidence and spans in-house, third-party, and cross-vendor scope.
Pillar one: identity and access security
| # | Assessment dimension | What "good" (score 4) looks like | Microsoft anchor |
|---|---|---|---|
| I-1 | Inventory and identity | Every agent, in-house, third-party, shadow, holds a first-class Entra Agent ID identity and appears in the Agent 365 registry | Entra Agent ID; Agent 365 |
| I-2 | Delegated vs own-access | Every agent is classified; own-access is reserved for genuinely autonomous agents and scoped from first principles | Entra Agent ID object model |
| I-3 | Least privilege | Access is scoped at the tool call and credential, evaluated per task; delegation inheritance is constrained | Tool-scoped permissions |
| I-4 | Conditional access | Conditional Access evaluates every token; approval-attribute and high-risk blocks enforced after report-only validation | Conditional Access for agents |
| I-5 | Credential lifecycle | Credentials are just-in-time, purpose-bound, and auto-expiring; no standing secrets persist past the task | Entra; just-in-time provisioning |
| I-6 | Audit trail | Every authentication, token, invocation, and policy decision is logged with a unique agent ID; integrated with Purview | Entra logs; Purview |
| I-7 | Revocation | The kill-switch has been tested against a live agent, confirming loss of access, token invalidation, and logging | Identity governance; rapid CA policy |
| I-8 | Secure Score baseline | Identity Secure Score is tracked and the human/tenant identity estate clears Microsoft's recommended baseline | Microsoft Secure Score |
Pillar two: data architecture
| # | Assessment dimension | What "good" (score 4) looks like | Microsoft anchor |
|---|---|---|---|
| D-1 | Unification | Sources consolidated into one logical lake; a single authoritative representation per core entity | OneLake |
| D-2 | Curation | Agents ground on the curated gold layer of a medallion architecture, never on raw ingest | Fabric medallion; OneLake |
| D-3 | Measured quality | Data quality is scored by domain, system, and business rule, visible in scorecards | Microsoft Purview |
| D-4 | Lineage | Every agent answer is traceable through gold, silver, to bronze source; lineage is continuous | Microsoft Purview |
| D-5 | Governance and access | Classification and access policy follow the data lifecycle; the agent cannot surface data the user cannot see | Microsoft Purview |
| D-6 | Recency | Current data is available to ground on where staleness produces wrong answers | Fabric Real-Time Intelligence |
| D-7 | Residency | Regulated data stays inside its permitted boundary across storage, inference, and logging; level matches data class | Sovereign Cloud mechanics |
Pillar three: cloud spend
| # | Assessment dimension | What "good" (score 4) looks like | Microsoft anchor |
|---|---|---|---|
| S-1 | Visibility and attribution | Spend is tagged to team, project, model, and use case via the five-dimension standard | Microsoft Cost Management |
| S-2 | Token telemetry | Telemetry meters at the step level, so the agent multiplier is visible, not buried | Cost Management; telemetry pipeline |
| S-3 | Model right-sizing | Requests are routed so simple work reaches cheaper models; no large-model default | Azure Advisor; model routing |
| S-4 | Discount discipline | Prompt caching and the Batch API are applied to every qualifying workload; discounts stacked | Azure OpenAI; Foundry |
| S-5 | Commitment discipline | PTU and GPU commitments validated against 30–60 days of telemetry and reviewed quarterly | Azure Advisor; Cost Management |
| S-6 | GPU optimisation | Idle GPUs auto-shut-down; serverless for variable load; capacity pooled, not hoarded | Azure Advisor; Well-Architected |
| S-7 | Unit economics | Spend governed on cost per unit of business value, not absolute budget | Well-Architected cost pillar |
Compliance overlay (scored across all three pillars)
| # | Overlay dimension | What "good" (score 4) looks like | Anchor |
|---|---|---|---|
| C-1 | EU AI Act human oversight | Oversight pattern selected per decision by risk class; override built; rationale documented | EU AI Act Art. 14 |
| C-2 | Reconstructible audit trail | Decision logs are write-once, read-many, tamper-evident, and reconstructible | EU AI Act traceability |
| C-3 | DPDP fiduciary accountability | Accountability retained under automation; DPIA, audit, fairness assessment current for SDFs | India DPDP Rules 2025 |
| C-4 | Compliance posture score | Purview Compliance Manager score tracked; customer improvement actions remediated, not just Microsoft-managed baseline | Purview Compliance Manager |
Appendix B: Composite scoring rubric
The composite resolves the pillar scores and the overlay into a single band. The rule is floor-weighted within pillars and overlay-gated across them.
Step 1: pillar scores. For each pillar, take the mean of its dimension scores, then apply the dependency cap: if dimension I-1 (inventory and identity) scores below 2, the identity pillar is capped at the I-1 score, because an uncounted estate cannot be governed on any other identity dimension. The same cap applies to D-1 (unification) for data and S-1 (attribution) for spend.
Step 2: composite. The composite is the lowest of the three pillar scores, not the mean, because the pillars gate each other and the weakest pillar sets the estate's true exposure.
Step 3: overlay gate. If the compliance overlay (mean of C-1 to C-4) scores below 2, the composite band is capped at controlled regardless of the pillar composite, because an estate that cannot pass an audit is undeployable in regulated scope.
The bands.
| Band | Composite | Meaning | Priority |
|---|---|---|---|
| Exposed | 0.0–1.4 | One or more pillars absent or ad hoc; blast radius uncontrolled; overlay fails | Priority one: identity foundation immediately |
| Controlled | 1.5–2.4 | Foundations in place but uneven; audit-approaching, not audit-ready | Calibrate control; complete the overlay |
| Balanced | 2.5–3.4 | Control calibrated to risk across pillars; overlay clears; agility preserved | Hold position; re-audit on cadence |
| Optimised | 3.5–4.0 | Governed end-to-end; control continuously re-tuned; balance is a managed variable | Sustain; extend cross-vendor |
The bands map onto the chapter-one quadrants: exposed and paralysed both resolve to the exposed or controlled bands by different routes, and balanced is the target. A paralysed estate, high control, low agility, scores well on pillar dimensions but should read its low band as over-control, and remediate by calibrating control to risk rather than adding more.
Appendix C: Remediation-priority matrix
The matrix prioritises remediation by blast radius against remediation effort. Address the high-blast-radius, low-effort gaps first; sequence the high-effort gaps behind the identity foundation they depend on.
| Gap (low scoring dimension) | Blast radius | Remediation effort | Priority | Sequence |
|---|---|---|---|---|
| No first-class identity / registry (I-1) | Critical: whole estate invisible | Low–medium | P1 | Stage 1 |
| Untested revocation (I-7) | Critical: cannot stop a rogue agent | Low | P1 | Stage 1 |
| No conditional access (I-4) | High: any resource reachable | Low–medium | P1 | Stage 1 |
| Standing credentials (I-5) | High: immortal exposure | Medium | P2 | Stage 1–2 |
| Policy does not follow data (D-5) | Critical, agent tunnels access; DPDP breach | Medium–high | P1 | Stage 2 |
| No residency control (D-7) | Critical: unlawful processing | Medium–high | P1 | Stage 2 |
| Fragmented estate (D-1) | High, caps accuracy; contradiction | High | P2 | Stage 2 |
| No lineage (D-4) | High, undefensible; traceability fail | Medium | P2 | Stage 2 |
| No attribution / tagging (S-1) | Medium: spend ungovernable | Low | P2 | Stage 3 |
| Forecast-based commitments (S-5) | Medium: wasted reserved spend | Low–medium | P3 | Stage 3 |
| No unit economics (S-7) | Medium: profitability unknown | Medium | P3 | Stage 3 |
| Overlay below threshold (C-1 to C-4) | Critical: undeployable in regulated scope | Medium–high | P1 | Woven through 1–3 |
Priority reads P1 (immediate, high blast radius) to P3 (scheduled). Compliance-overlay gaps are P1 wherever the estate operates in regulated scope, because they cap the composite band irrespective of pillar strength.
Appendix D: Glossary
- Agency, excessive. The OWASP-named risk in which an agent holds more permissions, functionality, or autonomy than its task requires.
- Agent 365. Microsoft's unified control plane for agent identity, governance, and security; generally available 1 May 2026.
- Agility-control balance. The measurable position an organisation occupies between operational speed of agent adoption and the strictness of enterprise control over it.
- Azure Advisor. Microsoft's recommendation engine identifying idle and underutilised resources and right-sizing opportunities.
- Blast radius. The extent of the data, systems, and cost an uncontrolled agent or identity can reach.
- Composite band. The single position, exposed, controlled, balanced, or optimised, resolved from the three pillar scores and the compliance overlay.
- Conditional Access for agents. The Entra policy engine extended to evaluate every token an agent acquires against scope, approval attribute, and agent risk.
- DPDP. India's Digital Personal Data Protection Act, 2023, and the Rules, 2025, that operationalise it.
- Entra Agent ID. Microsoft's mechanism assigning a first-class identity to each agent; generally available April 2026.
- Floor-weighted composite. A scoring rule in which the lowest pillar, not the mean, sets the composite, because the pillars gate each other.
- Medallion architecture. The staged bronze–silver–gold pattern that refines raw data into AI-ready, business-ready data.
- OneLake. Microsoft Fabric's single, unified, logical data lake for the whole organisation, provisioned per tenant.
- Provisioned Throughput Unit (PTU). A reserved unit of model-processing capacity at fixed cost, delivering savings on sustained, high-utilisation workloads.
- Purview. Microsoft's governance control plane for lineage, data quality, classification, and policy across the estate.
- Purview Compliance Manager. Microsoft's compliance-posture tool, scoring from 0 to 900 against regulatory assessment templates.
- Secure Score. Microsoft's quantified security-posture measure; the Identity Secure Score is its identity portion, expressed as a percentage.
- Sovereign boundary. The permitted residency and compliance perimeter inside which all regulated data, inference, review, and logging must remain.
- Unit economics. The cost of an AI workload per unit of business value delivered, the measure distinguishing profitable from unprofitable deployments.
- Well-Architected Framework. Microsoft's five-pillar architecture review covering cost optimisation, reliability, security, operational excellence, and performance efficiency.
- WORM. Write-once, read-many storage, used to make audit logs immutable and tamper-evident.
Appendix E: Sources
This capstone draws on the following, accessed June 2026. Figures are attributed in-text to their primary source, and figures established in the companion papers are reused for consistency rather than re-sourced. Where a statistic appears across multiple secondary outlets, the most authoritative attribution available is given.
Microsoft assessment tooling (verified June 2026):
- Microsoft Learn: Identity Secure Score (percentage indicator; 2026 recategorisation of identity recommendations).
- Microsoft Learn: Compliance Manager scoring and Microsoft Security: Purview Compliance Manager (0–900 improvement score; 320+ assessment templates); EPC Group, 2026 (30–40% Microsoft-managed baseline).
- Microsoft Learn: Azure Advisor cost recommendations and Azure FinOps; nOps, 2026; EPC Group, 2026 (20–30% compute reduction from right-sizing; Cost Management lifecycle).
- Microsoft Learn: Azure Well-Architected Framework pillars; Microsoft Community Hub: Well-Architected Review assessment (five pillars; ~60-question self-assessment).
- Microsoft: Entra Agent ID (GA April 2026), Agent 365 (GA 1 May 2026, USD 15-per-user add-on), AI Agent Guardrails enforcement layer; Microsoft Learn and Microsoft Community Hub agent-governance documentation, 2026.
Compliance regimes (verified June 2026):
- Gibson Dunn; Hogan Lovells, EU AI Act Digital Omnibus provisional agreement, 7 May 2026; Annex III deferred to 2 December 2027, Annex I to 2 August 2028; Article 14 substantive standard unchanged; 2026 enforcement cycle.
- Ministry of Electronics and Information Technology (India), Digital Personal Data Protection Rules, 2025; Significant Data Fiduciary obligations.
Companion papers in the Voice of Volge series (June 2026):
- From copilot to autopilot: operationalising autonomous AI agents in the enterprise, Gartner 40%+ cancellation by 2027; agents-as-digital-labour framing; Agent 365 GA.
- An identity for every agent: who is accessing your data?, non-human identity ratios (80:1, 100:1); 84% cannot pass an agent-behaviour audit; Entra Agent ID object model; breach pipeline; conditional access mechanics.
- The data estate is the accuracy budget: architecting for the agent era, Gartner 60% abandonment without AI-ready data; 63% lack data-management practices; OneLake, medallion, Purview, Real-Time Intelligence.
- The meter is always running: taming the AI bill, Gartner USD 2.59 trillion / 47% AI spend 2026; 5–30× agentic token multiplier; 73% costs exceed projection; FinOps levers and PTU economics.
- Oversight is the enabler, not the brake: human-in-the-loop guardrails and data sovereignty for regulated AI, EU AI Act Article 14 and Digital Omnibus; India DPDP Rules 2025; sovereignty mechanics; WORM audit-trail standards.
Underlying primary sources (via the companion papers): Gartner (2025–2026); McKinsey (2025–2026); CyberArk (2026); Cloud Security Alliance: Agentic Trust Framework (2 February 2026); OWASP: Top 10 for Agentic Applications (9 December 2025); FinOps Foundation: State of FinOps 2026; Rack2Cloud, Spendark, CloudZero, Spheron, Index.dev (2026); European Union AI Act; MeitY (India).
Addendum: How Volge's forward-deployed engineers shall help: Assessment and advisory
The Volge engagement team
Volge delivers through one accountable team that runs from the boardroom to the production branch. Four roles carry the engagement, and they hold their remit from the first workshop to operational handover rather than handing off between phases.
| Role | What they own |
|---|---|
| Principal | The client relationship and the account. Sets the engagement's business outcomes, the governance cadence, and the value tracking, and stays the single point of accountability throughout. |
| Architect | The target architecture and the transformation roadmap. Sequences the work, sets the standards, and makes the platform, build-versus-buy, and control-design calls. |
| Principals | The senior engineering leads. Each owns a workstream, makes the hard technical calls, sets the quality bar, and mentors the embedded engineers. |
| Forward-deployed engineers | The embedded build. They sit inside the client's team, build and ship in the client's own environment, and own the last mile to production. |
The Principal holds the outcome, the Architect holds the design, the Principals hold the workstreams, and the forward-deployed engineers hold the keyboard, one team, accountable from strategy to shipped software.
This capstone has argued a single proposition: the balance between agility and control is a measurable position, scored across identity, data, and spend under one compliance overlay, and the failure in most estates is one of measurement and discipline rather than procurement. Volge runs that audit and the remediation it points to through a forward-deployed engineer. The scorecard in Appendix A is built to be run against a live Microsoft estate, the composite is built to be remediated against a sequenced plan, and both are work the forward-deployed engineer delivers end to end: from the first assessment through to the re-baseline that keeps the position from decaying.
What a forward-deployed engineer is
A forward-deployed engineer is a senior Volge engineer who runs the audit and then embeds in the client's own team to remediate inside the client's environment, staying through to a re-baselined posture. The model is deliberately not a posture scan handed to the client to action: the same engineer who scores the estate against the Appendix A scorecard owns the real-world engineering that closes the gaps, standing up Entra Agent ID, unifying sources into OneLake, wiring Purview governance, right-sizing the spend, and clearing the compliance overlay. They work in the client's tenant, against the client's regulated boundary, alongside the client's platform and security teams, and they do not leave when the report is written. The engagement ends at a measurable re-baseline, not at a slide.
How the forward-deployed engineer helps the client
The forward-deployed engineer owns the engineering across all three pillars rather than scoring and departing. The assessment-and-advisory practice spans the same surfaces the audit scores, identity and security, cloud, data, and compliance, and grounds in the Microsoft instruments the estate already owns. The engineer converts a self-reported position into an evidenced one, then carries the remediation through: identity hardening, data-estate remediation, cost and FinOps optimisation, compliance wiring, and the remediation backlog driven down to a measured re-baseline.
| Service | Client outcome |
|---|---|
| Estate audit against the scorecard | The estate is scored 0–4 across identity, data, and spend under the compliance overlay, resolving to an evidenced composite band, exposed, controlled, balanced, or optimised, rather than a culture-flattered self-report. |
| Identity and security posture assessment | The Identity Secure Score baseline, Conditional Access coverage, PIM and just-in-time privilege, ID Governance recertification, and Agent 365 registry adoption are assessed against the Zero Trust maturity model, surfacing where licensed controls run un-operationalised. |
| Cloud and FinOps assessment | Spend is assessed against Azure Advisor right-sizing, Cost Management attribution and tagging, reservation and commitment discipline, and the Well-Architected cost pillar, converting a volatile aggregate bill into an attributed, unit-economics view. |
| Data-readiness assessment | The estate is assessed against OneLake unification, the medallion gold layer, and Purview lineage, measured quality, classification, and residency, locating the ceiling on agent accuracy and the floor under compliance risk. |
| Compliance mapping | The Purview Compliance Manager posture, the EU AI Act oversight and traceability obligations, and the India DPDP fiduciary duties are mapped onto the three pillars, separating Microsoft-managed baseline from the customer remediation that remains. |
| Remediation roadmap | The scored gaps are sequenced by blast radius and dependency, identity foundation first, data and compliance second, spend and optimisation third, and delivered as engineered guardrails on a governed landing zone, not committee gates. |
The engagement, scope to re-baseline
The forward-deployed engineer runs a single repeatable engagement from scoping through to a quarterly re-audit, with the three pillars assessed in parallel under a compliance overlay before the composite is scored, prioritised, and remediated in the client's own environment.
How the model helps firm and client together
The forward-deployed engineer model compounds for both sides, which is what separates this practice from a posture scan that ends at a report.
The scorecard and accelerators are reusable IP that improve with every engagement. The Appendix A instrument, floor-weighted composition, overlay gate, 0-to-4 anchors per dimension, plus Volge's scoring harness, are refined across the companion-paper series and sharpened on each estate the engineer remediates. A feedback loop runs both ways: the client gets an instrument hardened on real estates, and the firm carries every lesson forward into the next engagement. The output is an evidenced band an estate can act on, not a maturity grid filled in by impression.
The view is cross-pillar, not single-tool. Identity gates data and spend, and an audit that scores one surface in isolation misses the dependency that caps the composite. The engineer assesses all three pillars and the overlay together, so a strong spend position built on uncounted agents reads as the exposure it is.
The engineering follows the audit through to the fix. Volge is a deep-tech engineering firm, not an assessment-only shop. The same engineer who scores the gap stands up Entra Agent ID and the registry, unifies sources into OneLake, governs through Purview, and lands remediation on a Cloud Adoption Framework landing zone with Azure Policy guardrails. The audit and the remediation are one engagement, and the relationship is outcomes-based, measured against the re-baselined composite, not billed against a report delivered.
Accelerators compress the cycle. The assessment leans on instruments already paid for, Secure Score, Azure Migrate and Advisor, the Well-Architected Review, Purview Compliance Manager, and on Volge's own scoring harness, so the first evidenced position arrives in weeks, not a planning quarter.
Routing the weak pillar to the practice
A low pillar score is not a verdict; it is a routing instruction. The forward-deployed engineer who runs the audit routes each weak pillar to the matching Volge practice that remediates it, drawing in specialist depth where a pillar needs it, while the compliance overlay binds across all three.
Engagement: the first assessment
The concrete first step is a bounded assessment against the Appendix A scorecard, run by a forward-deployed engineer: a fixed-scope audit of one regulated workload or business unit across all three pillars and the compliance overlay, returning an evidenced composite band, a prioritised remediation matrix, and the first ninety days of the identity-foundation roadmap. The same engineer then embeds to drive the backlog down to a re-baseline. The deliverable is a position the leadership team can act on rather than argue about: the honest measurement this guide has held to be the precondition for moving the balance. Any efficiency or savings figure quoted at scoping is illustrative until the assessment measures the specific estate. Volge would welcome the chance to run that first assessment against a defined Microsoft estate.
Prepared by Volge · June 2026. This document is a point of view intended to inform IT, security, architecture, and governance decision-making. It is not legal, financial, security, or compliance advice, and it does not establish an advisory relationship. Volge is not a law firm. Organisations operating in regulated scope should engage qualified legal and compliance counsel before relying on any statement in this guide for a compliance decision. The audit scorecard is a structured assessment aid, not a certification; a high score is not a guarantee of security, accuracy, or compliance, and a completed scorecard does not substitute for independent audit or penetration testing. Statistics are attributed to their sources and were accurate to the best of available knowledge as of June 2026. Product capabilities, availability dates, licensing, pricing, and regulatory timelines are subject to change by the vendor or legislator and should be reconfirmed against current documentation before they are relied upon.
Voice of Volge: the firm's point of view on the technology it delivers. Sourced, year-stamped, and built to be handed to a board.
Volge is an independent deep-tech enterprise service provider and premier Microsoft delivery partner. That interest is disclosed; the analysis in this book is independently sourced.