Skip to content

Template: Agent Action Authorization Matrix (AAM)

Purpose: Provide an auditable, enforceable definition of what an agent may and may not do, including hard limits and escalation triggers.
Applies to: Zone 3 agents by default; Zone 2 agents when they can initiate workflows/actions or touch sensitive systems/data.
Related controls (examples): 1.14 (Agent scope control), 1.18 (RBAC), 2.12 (Supervision & oversight), 3.1 (Inventory/registration).


1) Agent identity

  • Agent name:
  • Agent ID (unique):
  • Agent type: (M365 Copilot Agent / Copilot Studio Agent used with M365 Copilot)
  • Owner (business):
  • Owner (technical):
  • Approver (Compliance/Risk):
  • Zone: (Zone 1 / Zone 2 / Zone 3)
  • Governance level: (Baseline / Recommended / Regulated)
  • Production status: (Draft / Pilot / Production / Retired)
  • Last reviewed date:
  • Next review date: (recommended: quarterly for Zone 3)

2) Intended purpose and boundaries

Intended purpose (one paragraph)

Describe what the agent is supposed to do and for whom.

Out-of-scope behaviors (bullet list)

  • Examples: “Cannot provide investment advice”, “Cannot approve credit”, “Cannot contact customers directly”, “Cannot change records”.

3) Authorized capabilities (Allowed actions)

List actions the agent is explicitly authorized to perform.

Action category Action Target system Preconditions Evidence/log requirement
Read Read documents SharePoint site(s): ___ User has access + label allowed Log: resource ID + label
Generate Draft summary Word/Outlook Non-confidential only Log: prompt category + output category
Workflow Create a draft case Case mgmt tool Must assign to human reviewer Log: case ID + reviewer
Notify Notify compliance queue Teams/Email Only approved distribution list Log: notification ID

Notes - “Action category” should align to a simple taxonomy: Read / Write / Execute / Notify / Approve / Block / Escalate. - Prefer draft creation over final execution for regulated decisions.


4) Prohibited capabilities (Disallowed actions)

Explicitly list what the agent must never do.

Prohibited action Reason Enforcement mechanism Detection/alert
Execute financial transaction Customer harm + control bypass Technical block (no connector permission) Alert if attempted
Modify official records Recordkeeping integrity Role denies write Sentinel alert
Delete logs Audit destruction WORM store + no delete perms Alert + incident
Access HR/medical data Privacy/HIPAA concerns DLP + restricted discovery Alert if access attempted

5) Connector / tool authorization (technical scope)

Allowed connectors / integrations

List only approved connectors/tools.

Connector/tool Environment Allowed operations Data classes allowed Notes
Microsoft Graph Prod Read only Internal / Confidential? Restrict endpoints
SharePoint Prod Read + Draft write By label rules Restrict sites
Dataverse Prod Create draft row Internal only No delete

Denied connectors / integrations

  • List explicitly (e.g., “No external web browsing”, “No personal email”, “No consumer storage”).

6) Data access policy (high-level reference)

Point to the agent’s detailed data policy (Template: Per-agent Data Handling Policy). - Data policy link: ../agent-lifecycle/per-agent-data-policy.md


7) Hard limits (guardrails that cannot be overridden)

Limit type Limit Rationale Enforcement
Rate limit Max ___ actions/min Prevent runaway automation Platform throttling
Dollar limit Max $___ per day Risk control Workflow approval gate
Scope limit Only business unit ___ Least privilege Environment scoping
Time limit Only business hours Reduce unattended risk Scheduler constraint
Volume limit Max ___ notifications/hour Avoid alert fatigue Rate limiter

8) Human oversight + escalation triggers

Escalation triggers (examples)

  • Any “Execute” action attempted
  • Any access to Confidential/Restricted labels
  • Any confidence level = Low
  • Any scope drift detected
  • Any policy violation

Escalation routing

  • Primary reviewer role:
  • Secondary reviewer role:
  • On-call / after-hours route:
  • SLA for response:
  • Fallback if SLA missed:

(Reference the Escalation Decision Matrix template.)


9) Logging and evidence requirements

Copilot Studio activities can be audited via Microsoft Purview audit logs, with additional transcript access via DSPM for AI in some cases.

Minimum evidence for Zone 3: - Audit logging enabled and verified (Purview audit). - Decision logs (schema from Decision Log template) stored immutably. - Scope drift logs + alerts are retained and reviewable. - Quarterly AAM review evidence (sign-off).


10) Testing plan (required before production)

Negative tests (must pass)

  • Attempt prohibited connector access → blocked + logged.
  • Attempt prohibited action → blocked + logged.
  • Attempt to exceed rate limits → throttled + logged.

Positive tests (must pass)

  • Allowed actions work only within defined scope.
  • Alerts route correctly.
  • Logs are complete and retrievable.

11) Approvals and change control

  • Approved by (Compliance/Risk):
  • Approved by (Platform owner):
  • Approval date:
  • Change ticket ID / PR link:
  • Version:

Any update to allowed/prohibited actions should go through formal change management (align to your existing management controls).