Agent & Copilot Governance

Know which agents are in your tenant and what they can reach

Copilot Studio makes it easy for anyone with a licence to build an agent, ground it on SharePoint content and publish it. Months later, nobody can say who owns it, whether anyone uses it, or what it can read. T365 keeps an inventory of the agents it can see, records how each one changes, and scores the risk each one carries.

Agent inventory — discovered and risk-scored at each scan
Scheduled scan

Surfaces scanned

  • Copilot Studio agents
  • Azure Bot Service bots
  • Leave policy assistantCopilot StudioNo rules raisedlow · 0
  • Quote builderAzure Bot ServiceOwner no longer exists + 1 moremedium · 25
  • Project helperCopilot StudioUnused for 90 dayslow · 5
  • Ticket triageAzure Bot ServiceGrounded on anonymous-link content + 2 morehigh · 55
  • Contracts searchCopilot StudioGrounded on anonymous-link content + 4 morecritical · 80

Illustrative — sample data. Findings come from T365’s agent risk rules; your own approved-vendor list feeds the score.

What you get

Why teams choose Agent & Copilot Governance

01

Know which agents you have

One inventory of the Copilot Studio agents and Azure bots T365 can see, with owner, surface and publishing state.

02

See what each agent can reach

Rules flag knowledge sources shared by anonymous link or with guests, permissions beyond the manifest and webhooks to outside hosts.

03

Find agents nobody owns

Agents whose owner has left, or that nobody has used in 90 days, are flagged so they can be reassigned or retired.

04

Review the riskiest agents first

A score and a level for each agent, from rules you can tune, puts the agents that need attention at the top of the list.

05

Evidence for your AI register

Exports of the inventory, its change history and PSPF classifications give you a record to work from, not a spreadsheet kept by hand.

How it works

From connection to evidence

  1. 1

    Connect

    Grant T365 read access to Power Platform and Azure for each tenant.

  2. 2

    Scan

    Scheduled and on-demand scans list the agents found and record what changed since the last scan.

  3. 3

    Score

    Each agent is scored against the rules you have turned on, with a level from low to critical.

  4. 4

    Classify

    Record a PSPF classification against each agent and check it against your approved-vendor list.

  5. 5

    Report

    Export the inventory as CSV or JSON, and the classification report as CSV.

In detail

How Agent & Copilot Governance works

Inventory: the agents T365 can see

T365 reads Copilot Studio agents through the Power Platform inventory and Azure Bot Service bots through Azure Resource Manager, and keeps one inventory per tenant with each agent’s owner, surface, publishing state and the scan that last saw it.

What T365 finds depends on the permissions and consent your tenant has granted it. An agent on a surface T365 does not yet read, or behind a permission it does not hold, is not in the inventory, so treat the inventory as a record of what was found.

Change history: added, modified, retired

Scans run on a schedule for each surface, and you can start one on demand. Each scan records which agents were added, which changed and which are no longer seen, and keeps a snapshot of each version so you can compare an agent with how it looked before.

T365 also reads the Entra directory audit log to record when each agent was last used, so agents nobody has invoked for months stand out.

Risk scoring you can tune

Each agent is scored against a set of rules: an owner who has left the directory, no use in 90 days, a draft visible in production, a bot acting with the calling user’s permissions, permissions beyond its manifest, a webhook to a host outside Microsoft, a shared service principal, and knowledge sources shared by anonymous link or with guests. The weights of the rules that fire add up to a score and a level from low to critical.

A rule with no data for an agent is shown as having no signal, not as a pass. You can switch rules off, change their weights and set the score thresholds for each level for each tenant.

An optional AI review can be turned on for agents whose score falls within a range you set. It may suggest moving an agent up one level, with its reasoning recorded beside the score. It is off by default and uses the AI provider your organisation configures.

How a risk score is built — Ticket triage (Azure Bot Service)
high · 55
  • Grounded on anonymous-link content+25
  • App permissions beyond what it uses+20
  • Posts to an external webhook+10

Rules can be switched off or reweighted, and thresholds set per tenant. Defaults shown.

Illustrative — sample data.

Vendors, classification and reporting

Keep your own list of approved vendors. A vendor-published agent whose vendor is not on the list is flagged in its risk score, so it is reviewed rather than missed. The list flags; it does not stop an agent running.

Record a PSPF classification against each agent, one at a time or in bulk. T365 flags agents grounded on content labelled PROTECTED or higher, and reports those that carry no sensitivity label of their own. The inventory exports as CSV or JSON, and the classification report as CSV or a printable report.

Who it is for

Built for the people accountable for it

CISOs and ITSAs

See which agents are in the tenant, which carry the most risk, and what changed since the last scan.

AI accountable officials

Work from an exported inventory with owners and change history when you report on your agency’s use of AI.

Microsoft 365 and Power Platform administrators

Find unowned and unused agents and review the ones flagged for sharing or permissions.

Frameworks

The frameworks it helps you evidence

Digital Transformation Agency

Policy for the responsible use of AI in government

The policy expects agencies to know where they use AI and who is accountable for it. T365 gives you an inventory of the agents it finds, with owners, publishing state and change history, as input to that record.

Read the source (opens Digital Transformation Agency in a new tab)

Australian Cyber Security Centre

Engaging with artificial intelligence

The guidance points to risks such as sensitive data reaching an AI system and weak control of who can use it. T365 risk rules flag agents grounded on content shared by anonymous link or with guests, permissions beyond the manifest and webhooks to hosts outside Microsoft.

Read the source (opens Australian Cyber Security Centre in a new tab)

Attorney-General’s Department

Protective Security Policy Framework (PSPF)

T365 records a PSPF classification against each agent, flags agents grounded on content labelled PROTECTED or higher, and reports those that carry no sensitivity label of their own.

Read the source (opens Attorney-General’s Department in a new tab)

T365 measures, maps and records evidence against these frameworks. Your organisation remains responsible for its own assessment and authorisation decisions.

Questions

What buyers ask us

Which agents does T365 find?

Today, Copilot Studio agents through the Power Platform inventory and Azure Bot Service bots through Azure Resource Manager, in tenants that have granted T365 the access it needs. Other agent types, such as Teams apps and Microsoft 365 declarative agents, are not yet in the scan.

Does T365 stop risky agents?

No. T365 reports and scores; it does not change or switch off agents. An agent from a vendor not on your approved list is flagged in its risk score for someone to review.

How is the risk score worked out?

Each rule that fires adds its weight, up to 100, and the total maps to low, medium, high or critical against thresholds you set. You can change rule weights or switch rules off for each tenant, and rules with no data for an agent show as no signal.

Does AI decide an agent’s risk?

No. The score comes from rules. An optional AI review, off by default, can suggest moving a borderline agent up one level, and its reasoning is recorded beside the score.

How often are agents scanned?

Scans run on a schedule for each surface, and you can start one at any time. Each scan records which agents were added, changed or are no longer seen.

Evidence: where each claim on this page comes from
  • Copilot Studio agents read from the Power Platform inventory

    packages/agent-governance/src/connectors/power-platform-inventory.ts

  • Azure Bot Service bots read from Azure Resource Manager

    packages/agent-governance/src/connectors/arm-bot-service.ts

  • Scans scheduled for each surface

    services/scheduler/src/jobs/agent-governance-scan.ts

  • On-demand scan

    apps/web/app/api/agents/scan/route.ts

  • Version snapshots: added, modified, retired

    packages/db/src/repositories/agents.ts

  • Last-used date from the directory audit log

    services/agent-governance-worker/src/audit-replay.ts

  • Rules sum to a score and a level; disabled rules and weight overrides honoured

    packages/agent-governance/src/risk/engine.ts

  • Rule set

    packages/agent-governance/src/risk/rules/index.ts

  • Per-tenant thresholds, allowlist, and AI review off by default

    packages/db/src/schema/agent-governance-config.ts

  • Optional AI review suggests at most a level change

    services/agent-governance-worker/src/ai-sentencing-real.ts

  • Allowlist flag and PROTECTED knowledge-source rules

    packages/agent-governance/src/risk/rules/tier-c-stubs.ts

  • Bulk PSPF classification

    apps/web/app/api/agents/pspf/bulk/route.ts

  • PSPF distribution and PROTECTED-without-label report

    apps/web/app/api/agents/compliance/route.ts

  • Classification report export as CSV or HTML

    apps/web/app/api/agents/compliance/export/route.ts

  • Inventory export as CSV or JSON

    apps/web/app/api/agents/export/route.ts

Buying for government? We are on the DTA Digital Marketplace.

  • Software
  • Cloud
  • Professional services
  • Labour hire
  • Hardware
How to buy

Ready to Secure Your Microsoft 365?

Start your free 14-day trial or talk to our team about your needs.