The Agent Excellence Framework

Inventory is a start. Governance is the work.

Seven steps take you from finding your agents to running them responsibly. They all rest on the same thing: Azure, Power Platform, data and cost foundations that are in order.

For platform teams, security teams and leaders · A guide, not an assessment · Checked against Microsoft's documentation on

Start here

Where are you today?

Pick the level that sounds most like your organisation. Then take one next step. Just one.

Your current level. Your next step appears below the options.

Select a level to see where to go next. Most organisations are somewhere around 1 and 2, so you're in good company.

From Ad hoc to Visible

Your next step: Discover

Find what you already have: the agents, the data they read and the platforms they run on. You can't govern what you can't see.

From Visible to Governed

Your next step: pick one of Prove, Protect or Own

  • Prove, if nobody can say what an agent is worth.
  • Protect, if agents read sensitive data or take actions.
  • Own, if agents have a maker but no owner in the business.

From Governed to Measured

Your next step: Run, then Review

Put numbers behind your agents: usage, quality, security and scope. Then decide what happens when an agent stops earning its place.

From Measured to Scaled

Your next step: Grow

Turn what you have learned into skills and reusable patterns, so the next agent doesn't start again at step one.

Staying Scaled

Your next step: go round again

Repeat the cycle for every new platform and every new kind of agent, and recheck the foundations whenever the platform changes.

These are the five levels from the talk. They start a conversation. They are not a Microsoft maturity model or a compliance score, and your choice isn't saved.

Step 1 of 7 · Discover

Find every asset you already have.

What do we already have?

Most organisations I visit have more agents than they think. You can't govern what you can't see.

From the talk

It all starts with a foundation: your data and your platform. Then the agents themselves. Four things to find before you decide anything.

All platforms

#

Build the agent inventory

Start with the agent registry in the Microsoft 365 admin centre. It lists agents from Microsoft, from partners and from your own teams, with their owner, usage and risk. Then add what the registry can't see.

Where to look

  • Microsoft 365 admin centre. The agent registry and Agent Map, part of Agent 365.
  • Power Platform admin centre. Your environments, the Copilot Studio agents in them, and who made them.
  • Azure. Subscriptions and resource groups that host Microsoft Foundry projects or custom agents.
  • Other platforms. Agents on Amazon Bedrock, Salesforce and others. See agents outside Microsoft.
  • People. Ask makers and business teams what they built, bought or switched on.

What one record looks like

An example. AskHR is a made-up agent that returns throughout this guide.
Name and purposeAskHR. Answers leave and benefits questions for employees.
Platform and placeCopilot Studio, HR production environment, West Europe.
Functional ownerHR service lead, in the business.
Technical operatorPower Platform engineering team.
Identity and usersIts own agent identity. All employees, signed in with their own account.
Data, tools and actionsReads the HR policy library. Creates leave requests through the HR system connector.
Autonomy and riskActs with approval. Handles personal data. Internal users only.
FundingHR cost centre. Credits allocated to its environment.
LifecycleLive. Verified this quarter. Next review in six months.

Watch out A registry entry proves an agent exists. It doesn't prove that anyone owns it, reviewed it or governs it.

Who
Platform admin and business owner
Evidence
One reconciled register, with the gaps you know about and the agents that have no owner

ReadMicrosoft Agent 365 overview

Other platforms

#

Agents that live outside Microsoft

Agent 365 can connect to other agent platforms and add their agents to your registry. That is useful, and it has limits you should know before you rely on it.

Supported today: Amazon Bedrock, Google Vertex AI, Salesforce Agentforce, Databricks Genie, Anthropic Claude Managed Agents, Oracle Generative AI Agents and Snowflake Cortex.

  • Sync is manual. Run it again after agents change, and before any review that depends on the registry.
  • Deleted agents can linger. An agent removed on the other platform may stay in your registry until you remove it by hand.
  • It is read-only. Agent 365 lists and observes these agents. It doesn't run or change them.
  • Every connection needs a key or secret. Decide who rotates it.
  • Activity data varies. Coverage differs per platform, part of it is in preview, and collecting it can add charges on the other platform's bill.
Who
AI administrator and the other platform's administrator
Evidence
List of connections, last sync date, and a check that synced agents match the source

ReadConnected platforms in Agent 365

Under every step

Get the platform in order. Then add the agent.

Is the ground solid enough to build on?

  1. Azure
  2. Power Platform
  3. Cost
  4. Data and identity

Azure, Power Platform, cost and data are not setup details for later. They decide whether the other six steps hold.

Fit the foundation to the agent. An agent that only lives in Microsoft 365 doesn't need its own Azure landing zone. An agent hosted in Azure does. Start discovery now, and check the foundations that apply before anything goes to production.

AzureMicrosoft Foundry

#

Azure: a landing zone, not a loose subscription

An Azure landing zone is a prepared environment where identity, network, policy and monitoring are already in place. The platform team runs the shared part. The workload team runs the agent.

The eight design areas to settle

  1. Billing and tenant. A named owner and cost centre for every subscription.
  2. Identity and access. Least-privilege roles, and managed identities instead of API keys.
  3. Network and connectivity. Known routes in and out, private endpoints and working DNS.
  4. Resource organisation. Management groups and subscriptions that separate platform from workload.
  5. Security. Controls for every part of the environment, with secrets in Key Vault.
  6. Management. Logs, metrics and alerts that reach a named person.
  7. Governance. Azure Policy enforces the rules and flags drift.
  8. Automation. Infrastructure as code, so test and production match.

Who runs what

Platform team

The hub network and firewall, DNS, policies on the management group and the shared connection to everything else.

Workload team

Microsoft Foundry, the app with its search and storage, Key Vault, workload monitoring, and its own subnets and private endpoints.

Test before go-live

Private DNS resolves. Blocked routes stay blocked. Roles grant only what is needed. Alerts reach a person.

Watch out Treat Microsoft's reference architecture as a starting point, not a shopping list. Fully deployed, its chat baseline landing zone is likely to cost more than $100 a day.

Who
Azure platform team and workload team
Evidence
Approved architecture, policy and role checks, network test results and a handover to operations

ReadIntroduction to AI landing zonesCloud Adoption Framework: govern

Power PlatformCopilot Studio

#

Power Platform: decide where makers build

Every tenant has one shared default environment, and without a plan that is where agents pile up. An environment strategy gives makers a safe place to build and gives production a controlled way in.

Get in order

  1. Name the admins. A Power Platform admin for the tenant, and environment admins where it matters.
  2. Secure the default environment. Limit sharing, restrict connectors with a data policy, and keep it for personal productivity.
  3. Route makers away from default. Environment routing sends each maker to their own developer environment.
  4. Group your environments. Development, test and production groups, each with its own rules.
  5. Release through pipelines. Solutions move from development to production, with an approval in between.
  6. Control who creates environments. Only admins create production environments, so capacity stays planned.
  7. Tell makers the rules. Where to build, which connectors they can use and how to ask for more.

Microsoft's common starting points

  • 20people at most that one app, flow or agent is shared with in the default environment
  • 5people at most in a maker's own developer environment
  • Blocksolution checker in default and production. Warn in developer and sandbox.

Watch out Environment groups don't apply data policies for you, so scope each policy to the right environments yourself. A new sharing limit only affects future sharing, so check what is already shared. Groups, rules and routing need Managed Environments, and those need premium licences.

Changed this month. Dataverse for Teams environments are deprecated. You can't create new ones since 15 November 2026, so upgrade the ones you want to keep.

Who
Power Platform admin and environment admins
Evidence
Environment map, routing and group rules, a release approval and a capacity review

ReadPower Platform governance overviewEnvironment strategy

CostAll platforms

#

Cost: name the payer before the pilot

Plan, control, optimise. Every agent needs someone who pays for it, a forecast, and an agreed answer to one question: what happens when it reaches the limit?

Get in order

  1. Name the payer. A billing owner and a cost centre for every agent.
  2. Forecast the whole bill. Model usage, search, hosting, storage and logging in Azure. Copilot Credits and licences in Microsoft 365 and Power Platform. Charges from other providers.
  3. Set alerts, and limits where they exist. Know which of the two you actually have.
  4. Bound the agent itself. Cap retries, tool calls, loops and parallel runs.
  5. Decide who acts on an alert. Write it down and test it once.

Three places to manage cost, three behaviours

WhereWhat it coversAt the limitGood to know
Azure Cost ManagementAzure charges for a management group, subscription or resource group.Sends alerts and can trigger an action group. Nothing stops.Cost data usually arrives 8 to 24 hours later. Budgets are checked once a day.
Power Platform admin centreCopilot Studio agents, with credits and capacity per environment.Depends on how credits are allocated and billed. As presented at Ignite, prepaid credits are capped per environment and pay-as-you-go is not.The place for capacity allocation and detail per workload.
Microsoft 365 admin centreCopilot Credits by spending policy, user, group, agent and service.Spending policies set limits for users and groups. A task that is already running can finish past a per-user limit.Refreshes every 2 to 4 hours. Shows an estimate that may differ from your bill.

Two ways to fund agents

Fund a critical agent

"We need to manage AskHR's costs, so users don't have to self-fund." Put the spending policy on the agent or on its environment.

Fund people to experiment

"I want my makers to build, with a personal spending limit." Put the spending policy on users or groups.

Watch out An Azure budget is an alarm, not a brake. When it is reached, Azure sends alerts. Resources keep running and the bill keeps growing.

Moving cost management to the Microsoft 365 admin centre in five steps
  1. Assign roles. AI Administrator to set policies, AI Reader to review spend.
  2. Review what you have. For each environment: credits allocated, billing plan, agent limits and who owns it.
  3. Decide the policy type. A department's agents: an environment policy. One critical agent: a resource policy. Makers: a user or group policy.
  4. Pilot one environment. Set limits and alerts, and link it to the department's Azure subscription for chargeback.
  5. Expand. Environment by environment.

These steps were presented at Microsoft Ignite 2026 in session BRK262. Parts of it are new or still rolling out, so check what your tenant has before you plan around it.

Who
Finance or FinOps, platform admin and business owner
Evidence
A funding decision, a forecast, one tested alert and one tested limit

ReadAzure budgetsMonitor Copilot Credit spending

Data and identityAll platforms

#

Data: fix access before you speed it up

An agent finds whatever its users, or its own identity, can already reach. If a salary sheet was shared too widely years ago, an agent will find it faster than any person would.

Get in order

  1. Fix oversharing. Find the sites and files that are shared too widely, and repair the permissions.
  2. Choose the sources. Decide which data an agent may use, who owns it and how fresh it must be.
  3. Settle the identity. Does the agent act with the user's permissions or with its own? Test what it can reach, and what it must not.
  4. Set guardrails. Sensitivity labels, data loss prevention, retention and audit, where your licences cover them.
  5. Treat the conversation as data. Prompts, answers, memory and logs: who can read them, how long are they kept, and where?

Watch out Read-only is not low-risk. An agent that only reads can still surface something sensitive. And hiding a site from Copilot buys you time. It doesn't repair the permissions.

Who
Data owners, the identity team, and security and compliance
Evidence
Oversharing fixed or accepted, an approved source list, and access tests that pass and fail as expected

ReadSecure and governed data foundationData access governance reportsOn this blog: sensitivity labels first

Step 2 of 7 · Prove

Build the business case first.

Is this agent worth building?

If you can't explain the value in one sentence, you're not building an agent. You're building a hobby project.

From the talk

Every agent needs a business case. Then test it: start small, measure, and only then decide to scale.

  • ProblemWhat problem does it solve?
  • ValueWhat value does it bring?
  • ROITime saved, errors prevented, faster service.
  • Pilot and measureStart small, measure, then decide to scale.

All platforms

#

One sentence, one baseline, one small pilot

The one-sentence test

For who, this agent does what, so that a result you can measure.

For example: for employees, AskHR answers leave questions, so that HR spends less time on repeat tickets.

Before you build

  • Write down today's baseline. Volume, time, errors and cost, before the agent exists.
  • Check a simpler fix first. Better search, a flow or a process change may do the job.
  • Count the full cost. Building, credits or tokens, licences, support, and the people correcting the agent.

Run the pilot

  • Keep it small. Limit the users, the data, the actions, the time and the budget.
  • Test with real questions. Include the ones it should refuse or hand to a person.
  • Agree the outcome up front. What makes you expand, change or stop?

Watch out Usage is not value. A busy agent can still cost more than it saves, and estimated time saved is not money in the bank.

Who
Business sponsor and functional owner
Evidence
The baseline, the pilot scope, the results and a recorded go or no-go

ReadMicrosoft Copilot DashboardIgnite 2026: adoption and impact resources

Step 3 of 7 · Protect

Set risk and guardrails.

What is it allowed to do?

These are new digital employees. They need identity security and data security, just like your human colleagues.

From the talk

Before an agent goes live, put it through the same five questions. Your answers set the risk level, and the risk level sets the guardrails.

  1. What will it access?Decides the data protection: approved sources, labels and least privilege.
  2. Does it run on its own?An agent that acts without a person starting it needs tighter limits and closer monitoring.
  3. Where is the human in the loop?Decides which actions need approval, and who is allowed to give it.
  4. How is it allowed to answer?Decides the tone, the topics it must refuse and when it hands over to a person.
  5. Who is allowed to use it?Decides the audience: one team, the whole organisation, or people outside it.

All platforms

#

More capability, more guardrails

A simple chat, an agent that reads your data and an agent that acts for people do not need the same controls. Match the guardrails to what the agent can do.

The agentWhat it doesAdd at least
Observes and advisesReads and answers.Approved sources, tested answers, and a clear route to a person.
Acts with approvalProposes an action. A person approves it.Show what will change, check the approver has the authority, and log the decision.
Acts on its ownActs without asking.Narrow permissions, limits on runs and spend, monitoring, and a tested way to stop it.

Test it like an attacker would

Try prompt injection, poisoned content in its sources, sensitive data leaking into answers, and tool calls nobody asked for.

Watch out Instructions are not a security boundary. If an agent must not do something, take away the permission. Don't just tell it not to.

Who
Security and compliance reviewer, with the functional owner
Evidence
Risk assessment, the list of allowed actions, test results and a release sign-off

ReadGovern agents using Agent 365

Power PlatformCopilot Studio

#

Power Platform: govern connectors and release

Three questions for the Power Platform admin: where can makers build, what can they build with, and who can release and run an agent? Data policies answer the second one.

  • There is no data policy by default. Until you create one, nothing restricts which connectors makers combine.
  • Three groups. Each connector is business, non-business or blocked. One app or flow can't mix business with non-business, and data policies apply to Copilot Studio agents too.
  • Policies add up. When several apply, the most restrictive combination wins. An environment policy can't override a tenant policy.
  • Go finer when you need to. Allow a connector's read actions and block its write actions, or limit the endpoints it may call.
  • Microsoft's suggested start. One strict policy for all environments that allows Microsoft 365 and standard connectors, plus wider policies for the environments that need more.

Watch out A data policy controls which connectors can be used. It doesn't inspect the content, it can't tell a test connection from a production one, and it doesn't replace permissions in the system on the other end.

Who
Power Platform admin and the owner of the connected system
Evidence
Policy scope per environment, a blocked-connector test and a release approval

ReadImplement a data policy strategy

Step 4 of 7 · Own

Pair a business owner with engineers.

Who is responsible?

Owner without engineers means good intentions and no support. Engineers without an owner means a technically perfect agent nobody steers. You need both.

From the talk

Every agent needs a functional owner: someone in the business who answers for what the agent does and says. Not IT. And the platform needs engineers who can support it.

All platforms

#

Two names on every agent

Functional owner

The business. Answers for:

  • Why the agent exists and who it is for
  • Whether its answers and actions are right
  • The benefit it should deliver
  • Whether it continues, changes or retires
Platform engineers

IT. Answer for:

  • Where it runs and how it is deployed
  • Its identity, permissions and connections
  • Monitoring, incidents and recovery
  • Changes and releases

Who else has a say

  • Security and compliance. Risk review, data protection, audit and incidents.
  • Finance or FinOps. Who pays, the forecast, and what happens at the limit.
  • Leadership and adoption. Priorities, risk appetite and skills.

House rules

  • One accountable name for each approval, with a deputy.
  • The maker is not automatically the owner.
  • When an owner leaves or changes role, the agent gets a new owner or goes to Review.

Watch out Three things are easy to mix up: the agent's identity in Entra, the person who owns it in the business, and the admins who run the platform. Record all three.

Who
Functional owner and technical operator
Evidence
Named owner and operator with deputies, the responsibility split, and an escalation route

ReadManage Power Platform adoption at scaleOn this blog: AI agents won't start in IT

Step 5 of 7 · Run

Running an agent is operations.

Is it healthy?

Running an agent isn't "publish and forget". It's operations.

From the talk

An agent needs a lifecycle, just like an app does. Once it is live, you monitor it. Four questions to keep asking.

  • UsageIs it being used?
  • QualityIs it answering well?
  • SecurityIs it behaving securely?
  • ScopeIs it still doing what it was approved to do?

All platforms

#

Turn signals into a response someone owns

QuestionSignals to watchWhere to lookWho acts
UsageActive users, sessions and credits used.Agent overview in the Microsoft 365 admin centre, Copilot Studio analytics, cost dashboards.Functional owner
QualityFailed runs, hand-overs to a person, evaluation results and complaints.Evaluations in Copilot Studio or Microsoft Foundry.Functional owner and engineers
SecurityIdentity risk, sensitive data in answers and threat alerts.Defender, Entra and Purview. Their risk signals also show in the agent registry.Security operations
ScopeNew tools, connectors, data sources or audiences since approval.Change records and the registry.Technical operator, with a new approval if the risk changed

Have ready before go-live

  • A runbook. For wrong answers, unsafe actions and running out of budget.
  • A tested way to stop the agent. On the platform that actually enforces it.
  • Versions. Of prompts, tools, models and connections, with a retest after every change.
  • Restricted logs. Decide who can read transcripts, and for how long they are kept.

Watch out The registry gives you the view across your estate. It doesn't replace each platform's own monitoring, and an agent synced from another platform doesn't get every protection automatically.

Who
Technical operator, security operations and functional owner
Evidence
A monitoring baseline, a tested runbook, the incident log and retest results

ReadMonitor and manage agents with Agent 365

Step 6 of 7 · Review

Plan for the day the value fades.

What happens when the value fades?

We offboard employee accounts. We should offboard agents with the same discipline.

From the talk

This is the step almost nobody plans for. When an agent stops earning its place, you have three options.

All platforms

#

Update it, offboard it, or keep it on purpose

  • Option 1Update itAssign the fix, set a fresh test baseline, confirm the funding, and approve again if the risk changed.
  • Option 2Offboard itStop it running, unpublish it, revoke its permissions and credentials, remove its schedules and connections, stop paid resources and clean up the registry.
  • Option 3Keep it runningConsciously, not by default. Record why, who decided and when you will look again.

Review when

  • The date arrives. Pick a rhythm that fits the risk.
  • The owner leaves the company or changes role.
  • Its data, tools or audience change.
  • There was an incident.
  • Cost or usage drifts away from the business case.

Watch out Check what depends on an agent before you delete it, and keep the records your retention rules require. Agents synced from other platforms may need removing from the registry by hand.

Who
Functional owner and technical operator
Evidence
A recorded decision, with a renewed owner and access or a completed offboarding checklist

ReadAgent 365: governance and ownerless agents

Step 7 of 7 · Grow

Grow skills in-house and with a partner.

Are our people ready?

Grow is what keeps the cycle turning. Without it, you'll be back to step one with every new agent.

From the talk

All of this takes skills: security, identity, data and platform. Upskill your own people, work with a partner who already has the skills, or ideally both, so the knowledge stays in your organisation.

All platforms

#

Four skills, two ways to get them

Upskill your people

Build these skills inside your organisation.

Work with a partner

Bring in people who already have them, and make handing over the knowledge part of the deal.

Who needs to learn what

  • Makers. Where to build, which connectors they can use, and how to ask for more.
  • Engineers. Identity, deployment, cost and incident response.
  • Business owners. How to judge answers, weigh risk and decide when an agent retires.
Who
Leadership and adoption, with team leads
Evidence
A skills plan per role, published guidance for makers, and patterns you reuse

ReadExplore Microsoft Agent 365Train makers and environment admins

The tooling

Start with the question. Then open the right tool.

Different admins govern different layers, and no single tool covers all seven steps. Before you switch anything on, check the licence, the admin role and whether the feature has reached your tenant.

Agent 365 covers tools, not decisions

It is your control plane once the agent is alive.

  1. DiscoverStill on you
  2. ProveStill on you
  3. ProtectPartly covered
  4. OwnStill on you
  5. RunCovered
  6. ReviewCovered
  7. GrowStill on you
What the tool covers
  • Registry and identity
  • Access and security
  • Monitoring
  • Lifecycle
What stays with us
  • Writing the business case
  • Appointing the owner
  • Deciding your risk appetite
  • Training your people

This is where the tool helps. Everything else is on us.

From the talk

Seven questions every admin team gets asked

Based on Microsoft Ignite 2026 session BRK262, Scale agents, not risk.
QuestionAnswer it withWhere
1. Where can makers build?Environment strategy and environment routingPower Platform admin centre
2. What can they build with?Sharing controls and connector policiesPower Platform admin centre
3. Who can release and run an agent?Solutions, pipelines and release approvalsPower Platform admin centre
4. Who can access the agent?Identity and permissionsAgent 365 and Entra
5. What can the agent access?Data, tools and action permissionsAgent 365, Entra and Purview
6. Which content safety guardrails apply?Safe interactions and information protectionAgent 365, Purview and Defender
7. How do we manage cost and track value?Spending policies by resource or by identityPower Platform and Microsoft 365 admin centres

Which tool helps with which step

A dot means the tool helps with that step. No tool finishes a step for you, and Own and Grow stay mostly with people.
ToolDiscoverProveProtectOwnRunReviewGrow
Agent 365YesNoYesNoYesYesNo
Power Platform admin centreYesNoYesNoYesNoNo
EntraNoNoYesYesNoYesNo
Purview and SharePointNoNoYesNoYesYesNo
DefenderNoNoYesNoYesNoNo
Azure Policy, Monitor and FoundryNoNoYesNoYesNoNo
Cost ManagementNoYesNoNoYesYesNo
Copilot and agent dashboardsNoYesNoNoNoYesYes

The tools, one by one

AI admin

#

Microsoft Agent 365

Which agents exist, who owns them, and which need attention?

One view of your agents in the Microsoft 365 admin centre. The registry shows owner, usage and risk signals from Defender, Entra and Purview. You can approve agents before they are published, apply policy templates, find agents that lost their owner and block a risky one.

Check first, and know the limits
  • Licence. At least one user needs a qualifying Agent 365 licence. Microsoft says it works best on top of E5.
  • Other platforms. Sync is manual, and what you can see differs per platform.
  • Not included. Your business case, your owner, your risk appetite and your training.
Helps with
Discover, Protect, Run, Review
Evidence
Registry checked against each platform, a list of ownerless agents, and risk exceptions

ReadOverview and licensingLearning path

Power Platform admin

#

Power Platform admin centre and Copilot Studio

Where can makers build, what can they use, and who can release?

Environments, groups and rules, routing, sharing limits, data policies, capacity and pipelines. The recommendations page flags what needs attention across the tenant.

Check first, and know the limits
  • Roles. Power Platform admins set the policy. Environment admins apply it in their own environments.
  • Licences. Managed Environments features need premium licences.
  • Limits. Data policies govern connectors, not content.
  • New for makers. Credits shown in preview chat, in evaluations and in the agent monitor, as announced at Ignite 2026. Check whether your tenant has it.
Helps with
Discover, Protect, Run
Evidence
Environment rules, a blocked-connector test, a release approval and a capacity review

ReadGovernance overviewGovern at scale

Identity and access admin

#

Microsoft Entra

Who can use the agent, and what can its identity reach?

Agent identities and access packages control what each agent can reach. Decide whether an agent acts with the user's permissions or with its own, and review that access the way you would for a person.

Check first, and know the limits
  • Identity type. Check which kind of identity the agent uses, and which Entra features your licences include.
  • Built for people. Controls made for user accounts don't automatically cover agents.
  • The other end. Permissions granted inside the connected system still count.
Helps with
Protect, Own, Review
Evidence
The identity and access model, a permission review, a test of what it must not reach, and a record of revoked access

ReadIdentity and access in Agent 365

Compliance admin

#

Microsoft Purview and SharePoint Advanced Management

Is the information agents use and produce protected?

Sensitivity labels, data loss prevention, audit and retention in Purview. Oversharing reports and access controls for SharePoint sites in SharePoint Advanced Management, which comes with your Copilot licence.

Check first, and know the limits
  • Coverage. Check which workloads and licences each feature covers.
  • Hiding is not fixing. Restricting a site from Copilot doesn't repair its permissions.
  • Two kinds of DLP. Purview data loss prevention and Power Platform data policies are separate controls.
Helps with
Protect, Run, Review
Evidence
Access reviews and fixes, policy test results, audit and retention settings, and recorded exceptions

ReadDeployment blueprintData access governance reportsOn this blog: DLP for Copilot

Security admin

#

Microsoft Defender

Which agent activity needs investigating or stopping?

Threat detection and response for agents. Defender watches for threats, raises alerts and incidents, and can block unsafe activity.

Check first, and know the limits
  • Switch it on. Enable security for AI agents and check what it covers.
  • Listed is not protected. An agent in the registry doesn't automatically have runtime protection.
  • Test the block. Decide and test where a block actually takes effect.
Helps with
Protect, Run
Evidence
What is covered, how alerts are triaged, a tested block and an incident record

ReadEnable security for AI agents

Azure platform team

#

Azure Policy, Azure Monitor and Microsoft Foundry

Is the hosted agent compliant, observable and behaving as intended?

Azure Policy and role-based access for the baseline. Azure Monitor, Logs and Application Insights for operations. Microsoft Foundry to build, test and evaluate the agent.

Check first, and know the limits
  • Split logs. Workload logs and central platform logs often sit in different workspaces. Plan how you will read them together.
  • Compliant is not correct. A deployment that passes every policy says nothing about the quality of the answers.
Helps with
Foundations, Protect, Run
Evidence
Deployment checks, diagnostic coverage, alert tests and evaluation results

ReadAzure AI landing zones

Finance and admins

#

Cost Management in Azure and Microsoft 365

Who pays, what is used, and which limit really applies?

Azure Cost Management for Azure charges. Copilot Cost Management in the Microsoft 365 admin centre for Copilot Credits. Capacity views in the Power Platform admin centre for each environment.

Check first, and know the limits
  • Estimates. Dashboards show estimated cost that can differ from the invoice.
  • Overlap. People can be in several groups and policies, so don't add group totals together.
  • Alerts. Azure budgets warn you. They don't stop anything.
Helps with
Prove, Run, Review
Evidence
Usage reconciled with funding, the policies that apply, a delivered alert and a tested response

ReadCopilot Credit spendingAzure budgets

Leadership and adoption

#

Copilot Dashboard and Agent Dashboard

Is adoption turning into outcomes?

Adoption and usage signals for Copilot and agents, in Viva Insights. Combine them with the quality, cost and process measures you agreed in the pilot.

Check first, and know the limits
  • Access. Check who can see the dashboards, and which people and metrics they cover.
  • Usage is not ROI. Compare against the baseline from your business case.
  • Not for appraisals. These reports are not meant for judging individual employees.
Helps with
Prove, Review, Grow
Evidence
Outcomes against the baseline, the effort spent correcting the agent, and a recorded decision

ReadCopilot DashboardDashboard settings

Follow an agent

Three agents. The same seven steps.

Made-up examples that show how the steps play out on each platform. Every step links back to the part of the guide it leans on.

Microsoft 365SharePoint

A policy knowledge agent

  1. Discover the agent, its sites, its owner and who it is for.
  2. Prove it cuts repeat policy questions, in a small pilot.
  3. Protect it by fixing oversharing and approving its sources.
  4. Own: the policy team owns accuracy, IT owns support.
  5. Run: watch wrong answers, hand-overs and stale content.
  6. Review when policies change: update, retire or keep.
  7. Grow: teach content owners and users its limits.

Start with data and permissions, not with an Azure deployment.

Copilot Studio

AskHR, an agent that takes actions

  1. Discover its environment, connections and the HR flows behind it.
  2. Prove faster service, counting the time spent on corrections.
  3. Protect its actions with identity, connector policies and approval.
  4. Own: HR is the functional owner, engineers own releases.
  5. Run on funded credits, with limits and an alert someone answers.
  6. Review permissions, value and owner. Approve changed actions again.
  7. Grow makers and operators who can support it.

Controlling who can publish and protecting what runs are two different layers.

AzureMicrosoft Foundry

A custom agent that prepares contracts

  1. Discover where it is hosted, its identities, its data and its APIs.
  2. Prove it speeds up drafting, without handing over legal approval.
  3. Protect it with a landing zone, data access and narrow tool permissions.
  4. Own: legal owns the outcome, workload and platform teams run it.
  5. Run evaluations, telemetry, token limits and recovery.
  6. Review scope and cost. Retire resources and revoke access safely.
  7. Grow identity, data, evaluation and operations skills.

Adding it to Agent 365 doesn't replace the controls of the workload itself.

Take it into your next meeting

A small reference kit.

Four lists to copy into your own tools. Keep your real agent records there, not here.

What to record for every agent
  • Name and purpose
  • Platform, environment and region
  • Functional owner and technical operator
  • Identity and who can use it
  • Data, tools and actions
  • Autonomy and risk
  • Funding and cost centre
  • What it depends on
  • Last verified and next review
Five questions before go-live
  • Is the business case approved, with pilot results?
  • Are the foundations that apply checked?
  • Are risk, permissions and approvals tested?
  • Are owner, operator, funding and runbook agreed?
  • Is the review date set?

A list to talk through together. It doesn't certify that you are ready.

What to record for an exception
  • What is asked for, and why
  • The control and platform it affects
  • The risk, and what compensates for it
  • Who approved it
  • When it expires
  • Who fixes it for good
Agenda for a review meeting
  1. Purpose and value
  2. Usage and quality
  3. Incidents and access
  4. Cost and funding
  5. Owner and dependencies
  6. Open exceptions
  7. Decision: update, offboard or keep
  8. Next review date

Printing includes every topic and opens every section, whatever filter is on.

Your homework

Know your next step.

Agents aren't an IT project. They're new colleagues, so treat them that way. Discover what you have. Prove what it's worth. Protect what matters. Own it, run it, review it, and grow your people along the way.

Where this comes from

Sources, and how current this is.

The seven steps and the five levels come from my talk, From Agent Inventory to Agent Governance. The admin questions and the cost section draw on the Microsoft Ignite 2026 session BRK262, Scale agents, not risk: Governance and FinOps for the enterprise. The rest was checked against the Microsoft documentation listed here on .

Watch out Agent tooling changes every month. Anything described as new, in preview or announced can change before it reaches your tenant. Check licences, admin roles and rollout status before you plan around a feature.

This guide is my own synthesis. It is not an official Microsoft framework and it doesn't certify anything.

Microsoft documentation

More on this blog