The governance pyramid for Copilot Studio agents

Six layers, from a one-page charter down to the records that prove it, with the Microsoft control that enforces each one. Built by a small consultancy that runs governed agents for itself first and for clients next.

In short

Use six layers: Charter, Policies, Standards, Procedures, Guidelines, Records. Policies say what and why and rarely change; standards hold the numbers and named settings and change with the platform; every standard names the Microsoft control or record that proves it. If you are starting from nothing, write five documents first: an autonomy policy, a data and privacy policy, a build standard, a promotion procedure with a release certificate, and an incident, suspension and retirement procedure. All five are published in this library.

Why a pyramid at all

Most small firms govern agents with one long document that mixes beliefs, rules, settings and how-to notes. It fails in a predictable way: the settings change every month, so the whole document goes stale, and nobody can tell which sentence is a rule and which is advice.

A pyramid separates things by how often they change and who must approve them. The charter changes almost never. Standards change whenever Microsoft ships a new control. Records change every day. Keeping them apart means a platform update touches one standard, not the policy a client has signed.

The rule that holds it together is simple. Authority flows down: a standard may only make a policy concrete, never invent a new obligation. Evidence flows up: every standard names the Microsoft control that enforces it or the record that proves it. Whatever names neither is a guideline.

AUTHORITY FLOWS DOWN | EVIDENCE FLOWS UPCharterOwner: founder | Review: yearlyEnforced by: nothing directly; every layer below cites itPolicieswhat is allowed, and whyOwner: founder | Review: yearly or on triggerEnforced by: tenant settings, Purview, Conditional AccessStandardsmeasurable minimumsOwner: founder as platform owner | Review: quarterlyEnforced by: env groups, ACP, data policies, agent settingsProceduresstep by step, with gatesOwner: builder | Review: after every use that hurtEnforced by: admin audit, publish blocks, Manage agentsGuidelinesrecommended practiceOwner: builder | Review: continuousEnforced by: review and evaluation, not by the platformRecordsthe evidenceOwner: whoever acts | Review: per event, register monthlyEnforced by: Purview audit, Entra logs, analytics, Logos
Figure 1. Six layers. Each layer may only say what the layer above allows, and each layer must name the control or record that proves it is being followed. If a rule names neither, it is a guideline, not a standard.

The six layers

Six layers is the smallest set that still separates the things that change at different speeds. Fewer, and settings leak into policy. More, and a one-founder company spends its week maintaining paper.

LayerPurposeApprovesReviewEnforced by (Microsoft)
0. CharterStates why Temei uses agents and the non-negotiables: autonomy is earned, a named human owns every agent, nothing starts above L1.FounderYearlyNo control enforces a charter. It is enforced by being cited: every policy must trace back to a charter line.
1. PoliciesShort statements of what is allowed and why. They change rarely and are written so a client director can sign them.Founder; client accountable owner for client tenantsYearly, or after an incident or a law changeTenant scope: tenant-wide data policy, the setting that allows or blocks publishing of generative AI agents, Purview sensitivity labels and DSPM for AI, Entra Conditional Access on agent identities.
2. StandardsThe measurable minimums that make a policy true: numbers, lists, named settings. They change with the platform, which changes monthly.Founder as platform ownerQuarterly, and whenever Microsoft ships a governance changeEnvironment groups and rules, advanced connector policies, classic data policies (including Copilot Studio virtual connectors and endpoint filtering), Agent access channels, Authentication for agents, Managed Environments sharing limits, embedded maker credential sharing control, agent credit limits.
3. ProceduresThe ordered steps for every moment that changes risk: create, change policy, publish, promote, suspend, retire. Each ends in a record.Founder; executed by the builderAfter any run that went wrong, otherwise yearlyPower Platform admin activity logs in Purview (policy and group changes are logged), publish blocked on data policy or sharing violations, Manage agents in the admin centre to turn an agent off, Agent ID lifecycle in Entra.
4. GuidelinesAdvice that makes good agents more likely: how to write instructions, design tools, curate knowledge. Optional by design.BuilderContinuousNone by design. Checked indirectly by agent evaluation test sets and the activity map during review.
5. RecordsThe evidence that the layers above were followed: cards, contracts, decisions, test results, certificates, incidents.Whoever acts writes; founder reviewsPer event; register reviewed monthlyPurview audit for Copilot Studio activities and agent interactions, Entra sign-in logs for Agent IDs, Copilot Studio analytics and transcripts, the Temei decision log (Logos).

A policy says what and why. A standard says how much. A procedure says in what order. A record proves it happened.

Where ENGRAM sits in the pyramid

ENGRAM is not a seventh layer. It runs through four of them, and that is exactly why it works as a method rather than a poster.

  • Policy. The Agent Autonomy Policy makes the Trust Pyramid binding: nothing starts above L1, nothing rises without evidence, any reviewer may demote, only evidence promotes.
  • Standard. The thresholds live here because they will be tuned. ENGRAM sets floors (for L2, at least three scored Quality Ledger sessions with R×U of 0.75 or more and no scope violation). Temei adds its own, stricter floor: twenty correct recorded decisions before L2. Floors may be raised, never lowered, by a standard.
  • Procedure. The Agent Creation Gate is ENGRAM Step 0 to 2 (Register, Declare, Scope). The Promotion Procedure is Steps 4 to 6 (Score, Log, Promote).
  • Records. The Registration Card, Trust Level Declaration, Quality Ledger and Logos entries are the evidence base.

What Microsoft actually enforces

The platform enforces standards, not policies. Knowing which control backs which line is the difference between governance and good intentions. Points worth building around, as of November 2026:

  1. Two connector policy systems can apply at once. In mixed mode, a connector or action is usable only when the advanced connector policy allows it and no classic data policy blocks it. Agent access channels and Authentication for agents also both apply alongside the matching Copilot Studio virtual connectors; if either blocks, access is blocked. Your standard must list both sides.
  2. Group rules override environment settings. A rule published on an environment group takes precedence and the environment-level setting becomes read-only. Standards are therefore written per group, not per environment.
  3. Changes are not instant. Microsoft says sharing rules can take up to an hour to be enforced. We observed data policy changes taking up to about 90 minutes to reach Copilot Studio, so the change procedure applies policy changes the day before a build.
  4. Sharing limits only bite on authenticated agents. Managed Environments sharing limits (Editor and Viewer rights, individuals only, maximum viewers) apply only to agents that require authentication. The Agent Build Standard must therefore require Entra ID authentication before it can rely on sharing limits.
  5. Maker credentials are a sharing risk. A separate control can block sharing and publishing of agents whose connections reuse the maker's credentials. It governs sharing and publishing, not runtime, so it complements the Write-Path Standard rather than replacing it.
  6. Every agent now has an identity. New Copilot Studio agents get a Microsoft Entra Agent ID automatically, with the owner added as sponsor and connector permissions visible as API permissions that Conditional Access can target. On the Microsoft Copilot harness the Agent ID is provisioned at first publish rather than at creation. Retirement must delete the agent so its Agent ID goes with it.
  7. Audit is there, if you look. Copilot Studio activities and agent interactions land in the Purview unified audit log, and Power Platform admin activity (data policies, groups and rules) is logged by default. Agent evaluation test sets give repeatable test evidence that can run from the API or a connector.

The document set

Twenty-eight documents make up the full pyramid. The ones marked with a link are published in this library as working drafts you can read in full and adapt. The rest are on our list.

LayerDocumentWhat it fixesIn this library
CharterTemei AI CharterPurpose, principles, ENGRAM as the method, the red linesRead it
PolicyAI Use and Acceptable Use PolicyWho may build and use agents, for what, what is forbiddenNot yet published
PolicyAgent Data and Privacy PolicyWhich data classes an agent may read or write, UK GDPR basis, client data separation, retention of transcriptsRead it
PolicyAgent Autonomy PolicyTrust Levels L1 to L5 as binding rules, promotion on evidence, automatic demotion, no generic write toolsRead it
PolicySupplier and Model PolicyWhich models, harnesses and third-party connectors are acceptable and who decidesNot yet published
StandardEnvironment and Data Policy BaselineEnvironment groups, ACP plus classic data policy in mixed mode, allowed connectors and blocked actions, change lead timeNot yet published
StandardAgent Build StandardNaming, owner and sponsor, authentication, channels, sharing limits, knowledge sources, logging switched onRead it
StandardWrite-Path StandardAgents never hold generic write tools; writes go through a constrained workflow that enforces allowed fields and deduplicationRead it
StandardEvaluation and Promotion ThresholdsTest set size, pass marks, the 20 correct recorded decisions before L2, demotion triggersRead it
StandardLogging and Audit StandardWhat must be logged where, retention, who reviews and how oftenNot yet published
ProcedureAgent Creation GateRegistration Card, Trust Level declaration, scope, non-goals, acceptance contract before any buildNot yet published
ProcedureData Policy and ACP Change ProcedureChange one thing, record it, apply a day before the build because propagation lagsNot yet published
ProcedurePublish and Share ProcedureChannel choice, authentication check, sharing within limits, first publish provisions the Agent ID on the Microsoft Copilot harnessNot yet published
ProcedurePromotion Procedure (L1 to L2 and up)Evidence pack, reviewer, release certificate, Logos entryRead it
ProcedureIncident, Suspension and Demotion ProcedureWho turns the agent off, how, who is told, what reinstatement needsRead it
ProcedureRetirement ProcedureRemove sharing, delete agent and its Agent ID, archive records, close register entryRead it
GuidelineAgent Operating ModelThe 44 roles, guardian agents, how agents and humans hand offNot yet published
GuidelineInstruction and Prompt Writing GuideStructure, refusal wording, grounding, examplesNot yet published
GuidelineTool and Workflow Design GuideNarrow tools, Copilot Studio workflows as triggers on the Microsoft Copilot harness, idempotent writesNot yet published
GuidelineKnowledge Curation GuideSource selection, freshness, labels, what never goes into knowledgeNot yet published
RecordRegistration CardStep 0 of ENGRAM, one per agentNot yet published
RecordAcceptance Contract and ScopeAcceptance criteria, scope control, non-goalsNot yet published
RecordDecision Log (Logos)Every decision with date, evidence, reviewerNot yet published
RecordAgent RegisterOne row per agent: owner, level, environment, channels, statusNot yet published
RecordTest Evidence and Quality LedgerEvaluation runs, scored sessions, decision countsNot yet published
RecordTrust Level DeclarationSigned level per agent per versionNot yet published
RecordRelease CertificateOne page that says this version, at this level, passed these tests, approved by this personRead it
RecordIncident and Exception RegisterIncidents, waivers from a standard and their expiryRead it

Where to start

If you have agents running and none of this written down, do not try to write twenty-eight documents. Five carry most of the weight, in this order:

  1. An autonomy policy. Turns trust levels, the no-generic-write rule and automatic demotion into something a director can sign. Every procedure assumes it. Ours.
  2. A data and privacy policy. Which data classes an agent may touch at each level, lawful basis under UK GDPR, transcript retention, separation between your data and your suppliers'. The first question a data protection lead will ask. Ours.
  3. A build standard. The exact settings every agent must have: naming, owner, authentication, channels, sharing, knowledge sources, how writes are guarded, the tests it must pass. Ours.
  4. A promotion procedure with a release certificate. The evidence an agent needs before it touches live data, and the one page someone signs. Ours.
  5. An incident, suspension and retirement procedure. Who switches an agent off, where, who is told, and how an agent is removed completely. Ours.

Then write the charter that sits above them all. Ours.

Ownership in a small team

Today one person holds every role. That is acceptable only if the roles are written down separately, so that each decision is made deliberately in one capacity, and so that each role can be handed to someone else later without rewriting the pyramid.

RoleTodayWith clientsHolds
Governance ownerFounderFounder for Temei; client's named accountable owner for their tenantApproves charter, policies, exceptions; accountable for every agent
Platform ownerFounderClient IT admin, advised by TemeiApproves standards; makes data policy, ACP and group changes
BuilderFounder, with AI assistanceTemeiWrites procedures and guidelines; builds; writes the records
ReviewerFounder after a cooling-off period of at least one day, plus guardian agent checksA client reviewer who did not build the agentSigns promotion evidence and release certificates; may demote
Agent owner and sponsorFounderNamed business owner in the clientAnswers for one agent; is the Entra sponsor; receives incidents

Two rules make this honest. First, the person who built an agent never approves its own promotion beyond L2 once a second person exists; until then, a one-day delay and a written evidence pack stand in for independence. Second, in a client tenant the client owns policies and the agent; Temei owns its methods and its records of what it did.

How it maps to frameworks

The pyramid is built for practice, not certification. It lines up with the main frameworks so that a client already working towards one recognises the structure. It does not make anyone compliant on its own.

FrameworkWhere it lands
NIST AI RMF 1.0GOVERN: charter, policies, roles. MAP: Registration Card, scope, non-goals. MEASURE: evaluation thresholds, Quality Ledger, test evidence. MANAGE: promotion, incident, suspension, retirement.
ISO/IEC 42001:20235.2 AI policy (policy layer); 5.3 roles (ownership note); 6.1 risk and AI system impact assessment (Registration Card and Trust Level declaration); 7.5 documented information (the whole pyramid); 8 operation (procedures); 9 performance evaluation (records and reviews); 10 improvement (incident and change procedures).
EU AI ActRelevant only where agents serve people in the EU. Article 4 (AI literacy) points to the AI Use Policy; Article 50 transparency, telling people they are talking to an AI, belongs in the Agent Build Standard for customer-facing channels.
UK approachPrinciples-based, applied by existing regulators. The five principles (safety, transparency, fairness, accountability, contestability) map to the charter; UK GDPR and ICO guidance drive the Data and Privacy Policy.

Sources