Microsoft AI governance for Copilot and the agents your teams build
Microsoft AI governance is the set of rules, controls and records that decide which AI assistants and agents run in your tenant, what data they can reach, who owns them and what they cost. Microsoft Copilot (formerly Microsoft 365 Copilot) only shows people what they can already open, so AI governance starts with the governance of your Microsoft 365 tenant.
AI and agents overview
- Copilot licences
- 1,850
- Active users, 30 days
- 1,320
- Agents published
- 146
- +12 this month
- Agents without an owner
- 19
Copilot sessions, last 6 months
| May | 9,800 |
|---|---|
| Jun | 11,200 |
| Jul | 10,900 |
| Aug | 12,600 |
| Sep | 14,100 |
| Oct | 15,400 |
Top policy findings
- High 6
- Medium 23
- Agent grounded on a public website 4 High
- Agent shared with everyone 2 High
- Unlabelled site in Copilot scope 23 Medium
Example data, for illustration only
The risks
Why AI governance cannot wait
Copilot and agents work with what is already in your tenant. Five gaps tend to show up first.
-
Agents without an owner
People with the right licence can build agents in Copilot Studio or from SharePoint sites. An agent can then exist without a named owner, an end date or a review.
-
Data exposure
Copilot and SharePoint agents answer from content each user can already open. Years of broad sharing turn into answers, and outdated or duplicate files turn into confident but wrong ones.
-
Unchecked connections
Makers can ground an agent on public websites, add connectors and share it with the whole organisation, often before anyone in IT has looked at it.
-
Spend nobody owns
Copilot licences stay on disabled or inactive accounts, and some users hold a second AI subscription on top. Nobody is accountable for the bill.
-
Gaps in your records
GDPR expects a record of processing, and an impact assessment where the risk is high. The EU AI Act expects AI literacy measures and labels on some content. Without an inventory, both are hard to answer.
What we put in place
What does Microsoft AI governance cover?
Five controls, run every day in your tenant by a governance tool that we select, implement and support with your team.
Agent inventory
Every agent is listed, with its maker, its audience and its sources.
Agents are built in several places: Copilot Studio, SharePoint sites and Power Platform. Microsoft lists them in its own admin centres. We add one inventory that ties each agent to its maker, the people it is shared with, the knowledge it reads and its recent activity, refreshed on a schedule.
What it covers
- Microsoft Copilot use across the Microsoft 365 apps
- Copilot Studio and SharePoint agents
- Power Automate flows and Power Apps
- Maker, audience and knowledge sources of each agent
- Recent activity of each agent
- ChatGPT Enterprise and Claude Enterprise
How we deliver it
- We select the governance tool that fits your tenant and your licences.
- We connect it with permissions reviewed by your security team.
- We run the first inventory and walk through it with you.
- Your admins learn to read and filter it themselves.
AI and agents overview
- Copilot licences
- 1,850
- Active users, 30 days
- 1,320
- Agents published
- 146
- +12 this month
- Agents without an owner
- 19
Ownership and lifecycle
Every agent has an owner, a purpose and an end date.
An agent nobody owns keeps its access long after the project ends. Rules flag agents without a named owner, agents with no recent activity and agents that were never published. The owner confirms before anything is archived or deleted, and each decision is recorded.
What it covers
- A named owner for every agent
- Inactivity detection, for example after 90 days
- Agents created but never published
- Archiving or deletion after the owner confirms
- A record of every decision
How we deliver it
- We agree the owner rules and thresholds with you.
- Owners get a plain request and confirm, extend or retire.
- Retiring an agent follows a documented runbook.
- Your admins see open requests and decisions in one list.
Agents to review
- Expense helper Owned
- Onboarding FAQ No activity 97 days
- Contract summary No owner
Policy-based controls
Rules check agents and data every day, and send findings to owners.
Copilot Studio governance rests on a few checks: knowledge sources, connectors, sensitivity labels, sharing scope and roles. Checks run daily, findings are ranked by severity, and the owner of the agent or the site is asked to fix them. Where the tool supports it, the fix is applied once the owner approves.
What it covers
- Knowledge sources, including public websites
- Connectors used by each agent
- Sensitivity labels on content in Copilot scope
- How widely each agent is shared
- A severity for every finding
How we deliver it
- We start from tested baseline rules and adapt them to you.
- Rules run in report mode first, and we review the results together.
- Enforcement is switched on in waves, with owners informed.
- Your admins learn to add and tune rules.
Policy findings
- High 6
- Medium 34
- Low 7
- Agent grounded on a public website 4 High
- Agent shared with everyone 2 High
- Agent without a named owner 11 Medium
- Unlabelled site in Copilot scope 23 Medium
- Overlapping AI subscriptions 7 Low
Usage and cost
Copilot licences go to the people who use them.
Microsoft Copilot is an add-on licence for an eligible Microsoft 365 or Office 365 plan, so an unused seat is money spent for nothing. We compare assigned and active licences, show sessions by app and cost by department from Microsoft Entra ID attributes, and list the seats to reassign before you buy more.
What it covers
- Assigned and active Copilot licences
- Licences on disabled or inactive accounts
- Sessions by app
- Cost by department, from Entra ID attributes
- Users holding more than one AI subscription
How we deliver it
- We set the activity thresholds with your licensing team.
- Reports go to IT and finance on a schedule.
- Seats are reclaimed through a documented process.
Copilot licences by department
- Sales: Assigned 320, Active 240
- Operations: Assigned 260, Active 150
- Finance: Assigned 140, Active 120
- HR: Assigned 80, Active 35
Evidence for auditors
Show auditors what runs, who decided and when.
When a question comes from audit, legal or a regulator, you need the history of decisions, not a list rebuilt by hand. Decisions on agents and findings are kept, reports are scheduled and exported, and a compliance trend shows progress over time. Microsoft also records Copilot interactions in the Microsoft Purview audit log.
What it covers
- History of decisions on agents and findings
- Owner confirmations kept with each finding
- Scheduled, exportable reports
- Compliance trend over time
How we deliver it
- We agree with you which reports auditors need.
- Exports are scheduled to the right people.
- We review the trend with you after go-live.
Compliance score, last 6 months
| May | 64% |
|---|---|
| Jun | 67% |
| Jul | 71% |
| Aug | 70% |
| Sep | 76% |
| Oct | 81% |
Examples
Examples of AI governance rules we configure
| Rule | Why it matters | Typical severity |
|---|---|---|
| Copilot licence on a disabled account | A paid seat nobody can use, and often a sign of a gap in the leaver process. | High |
| Agent using a public website as a knowledge source | Answers can mix internal content with web pages nobody has checked. | High |
| Agent shared with the whole organisation without approval | Everyone can use it before anyone has checked what it reads and returns. | High |
| Agent without a named owner | Nobody answers for its access, its content or its retirement. | Medium |
| Agent with no activity for 90 days | It keeps its access and connections while nobody uses it. | Medium |
| Site read by Copilot without a sensitivity label | Nothing signals how sensitive the content is or limits how it is reused. | Medium |
| "Anyone" link without expiry on a site in Copilot scope | Whoever holds the link can open the files for as long as the link exists. | Medium |
| User with Copilot and a ChatGPT or Claude seat | Two AI seats for one person, often without a clear reason. | Low |
Examples. Rules and thresholds are agreed with your team.
How it works
How AI governance runs in your tenant
We move quickly without cutting corners: ready-made tooling, tested starting rules and runbooks your team keeps.
- Step 1
Connect
The tool is registered as an app in Microsoft Entra ID, with admin consent. We review its permissions with your security team before anything runs.
- Step 2
Discover
A first inventory of agents, Copilot licences and knowledge sources. We go through it with you to agree priorities.
- Step 3
Baseline
Rules run in report mode only. You see what would be flagged, and we adjust thresholds together.
- Step 4
Enforce
Owners are notified and asked to act. Fixes run after the owner approves, in waves.
- Step 5
Report
A dashboard for your admins, and scheduled exports for security, finance and audit.
Native tools
Native Microsoft controls and what a governance layer adds
Microsoft's own controls are the foundation. We keep them, configure them and add what they leave to manual follow-up.
| Microsoft control | What it does | What a governance layer adds |
|---|---|---|
| Agent Registry and Microsoft Agent 365 | The Microsoft 365 admin center has an agent registry that lists the agents available in your organisation. Agent 365, generally available and licensed per user, adds Microsoft's own registry, access and security controls for agents. | Owner rules, inactivity checks and owner confirmations for each agent, with licence and cost views, next to your other Microsoft 365 rules. |
| Power Platform admin center | Data policies control which connectors and knowledge sources agents may use. Managed Environments limit how widely agents are shared. The inventory lists agents, apps and flows. | Daily checks that flag agents grounded on public websites or shared too widely, ranked by severity and sent to the agent owner. |
| Microsoft Purview | Sensitivity labels protect content. Data Security Posture Management (DSPM) shows how AI apps and agents use sensitive data. The audit log records Copilot interactions. | Rules that find unlabelled content in Copilot scope and ask site owners to label it, with a record of each decision. |
| SharePoint Advanced Management | Available once at least one user has a Microsoft Copilot licence. Data access governance reports show where sites may be overshared. Restricted Content Discovery keeps a site out of Copilot and organisation-wide search while permissions are fixed. | Owner reviews of overshared sites and risky links, repeated on a schedule rather than run once. |
| Microsoft Entra ID | Microsoft Entra Agent ID gives each agent its own identity, so its access can be governed like any other account. | Department and account status from Entra ID, used to assign owners, reclaim licences and report cost by department. |
Microsoft names and features as of the last review date shown above.
How Neptune delivers
From tool choice to day-to-day support
One team selects the tool, puts it in place in your tenant, trains your people and stays with you afterwards.
-
Select
We pick the governance tool that fits your tenant, your licences and the agents you already have. We work with several vendors, not one.
-
Implement
We connect it to your tenant, set up starting rules and switch them on in waves, with owners informed.
-
Train
Admins and agent makers learn the rules on their own data, and keep runbooks they can reuse.
-
Support
We stay reachable after go-live, tune the rules and keep pace with Microsoft changes.
FAQ
AI governance questions, answered.
What does an AI agent inventory include?
Each agent's origin (Copilot Studio, a SharePoint site or Power Platform), its maker and named owner, who it is shared with, the knowledge sources and connectors it uses, and its recent activity. Microsoft lists agents in its admin centres; the inventory joins those details in one place, so AI agent governance and Copilot governance work from the same list.
Do we need Copilot readiness before governing agents?
No. Agent governance can start straight away: the first inventory shows which agents exist, and the first rules target owners and sharing. Copilot readiness work on oversharing and labels runs alongside it and benefits both.
Which agents can be governed?
Agents built in Copilot Studio and from SharePoint sites, Power Automate flows and Power Apps, and the use of Microsoft Copilot across the Microsoft 365 apps. ChatGPT Enterprise and Claude Enterprise are covered too. Other AI platforms are not covered today.
How do you govern Copilot Studio agents?
In the Power Platform admin center, data policies control which connectors and knowledge sources agents may use, and Managed Environments limit how widely agents are shared. We configure those first. The governance layer then checks every agent each day: a named owner, knowledge sources such as public websites, how widely it is shared and its recent activity, with findings sent to the maker.
How are SharePoint agents governed?
People with a Copilot licence can create agents from SharePoint sites and files, and an agent only uses content its user is allowed to open. Governing SharePoint agents therefore starts with the sites they draw on: sharing, sensitivity labels and owners. The inventory lists each agent with its site, its maker and its activity, so an unused or unowned agent can be retired.
How does this help with the EU AI Act?
The AI Act does not require a formal AI inventory, but keeping one is the practical way to know which tools are used and who needs training. Organisations using AI are expected to take measures that support AI literacy among their staff, and the rules for high-risk uses will apply from 2 December 2027. The Act can also apply to UK and Swiss organisations whose AI output is used in the EU.
A governance tool gives you the inventory, the records and the reports these questions call for. No tool makes you compliant by itself: that assessment stays with your legal and compliance teams.
Can we see who uses Copilot and agents, and what it costs?
Yes. For Copilot: assigned and active licences, sessions by app and cost by department from Microsoft Entra ID attributes. For agents: recent activity and who each one is shared with. People holding more than one AI subscription are listed too, so seats can be reassigned before you buy more.
Does this cover ChatGPT Enterprise and Claude Enterprise?
Yes. Their users are matched to Microsoft Entra ID, so ChatGPT and Claude accounts, projects, custom GPTs, seats and cost appear next to Copilot, including people who hold more than one assistant. The ChatGPT and Claude governance page, linked below, covers accounts, connectors and audit history in detail.
Related pages
More on Microsoft 365 governance
-
Copilot readiness
Copilot only shows people what they can already open. The readiness assessment finds oversharing, missing labels and idle licences before rollout, then keeps checking.
-
Compliance
Policies on paper do not check the tenant. We turn them into daily checks across Microsoft 365, send findings to owners and export the evidence auditors ask for.
-
ChatGPT and Claude governance
ChatGPT Enterprise and Claude Enterprise reach your people and your files. We list their accounts, GPTs, projects and connectors, and read seats and spend next to Copilot.
Insights
Further reading
Talk to us about AI governance.
Tell us which agents and AI tools are in use today. We will explain which tool fits your tenant and how we would put it in place with your team.