AI Governance Institute
All governance templates →How do we govern AI agents that take autonomous actions?

Implementation Kit

AI Agent Governance Templates and Sign-Off Checklist

Governing an agent that takes actions: an autonomy boundary spec that says what it can do without a human, an identity and permissions register, an action log schema, and a sign-off checklist covering business, risk, and legal.

Who this is for: The owner deploying an AI agent that can read, write, or trigger actions in real systems.

Download the kit (Markdown) ↓4 artifacts. Every table also copies as CSV.

1. Autonomy boundary specification

Spreadsheet

Every action type the agent can perform, its risk level, and whether it needs a human.

Template

Action typeSystems / scopeRisk levelAutonomy settingRationale
<action><where>read / write / irreversible externalautonomous / requires human confirmation / prohibited

Worked example

Action typeSystems / scopeRisk levelAutonomy settingRationale
Read knowledge baseinternal KBreadautonomousno data leaves; low impact
Read customer ordercurrent ticket's customer onlyreadautonomousscoped to the ticket; logged
Draft a replydraft only, not sentwrite (draft)autonomoushuman sends
Send a replycustomer-facingirreversible externalrequires human confirmationreputational and contractual exposure
Issue a refundpayments systemirreversible externalprohibited (v1)revisit after 3 months of clean operation
Modify a knowledge base articleinternal KBwriterequires human confirmationaffects future answers

Acceptance criteria

  • Every action the agent can technically perform has a row, including read actions.
  • Irreversible external actions default to human confirmation or prohibited, not autonomous.
  • Each setting has a rationale, and "prohibited" rows note what would change the decision.

2. AI agent identity and permissions register

Spreadsheet

Every agent has its own identity with least-privilege scope, tracked like a service account.

Template

AgentNon-human identityPermissions grantedScoped to taskGranted byExpiryLast access review
<agent><service principal / API identity><specific scopes>Y / N<name>YYYY-MM-DDYYYY-MM-DD

Worked example

AgentIdentityPermissions grantedScoped to taskGranted byExpiryLast review
support-copilotsvc-copilot-prodkb:read; orders:read (ticket-scoped); drafts:writeYPlatform IAM2027-03-012026-09-01
billing-reconcilersvc-billrec-prodledger:read; report:writeYPlatform IAM2027-01-152026-09-01

Acceptance criteria

  • No agent shares a credential with a human or another agent.
  • Permissions map to the autonomy boundary spec, with nothing broader.
  • Every identity has an expiry and is in the access review cycle.

3. Agent action log schema

Spreadsheet

Every agent action is attributable: who, on whose instruction, with what authority, and what happened.

Template

FieldRequiredNotes
timestampYes
agent_identityYesthe non-human identity used
actionYesthe operation and target
instruction_sourceYesuser request / scheduled job / another agent (which one)
authorityYesautonomous / human-confirmed (who) / policy-permitted
inputsYesparameters or a reference to them
outcomeYessuccess / failure / blocked by gate
reversibleYesY / N, and rollback reference if applicable

Worked example

FieldValue
timestamp2026-09-05T09:22:41Z
agent_identitysvc-copilot-prod
actionorders:read order=A-88213
instruction_sourceagent run for ticket 88213 (agent: support-copilot)
authorityautonomous (within ticket scope)
inputsticket_id=88213, customer=cust-4471
outcomesuccess
reversibleY (read only)

Acceptance criteria

  • Every action logs the instruction source, including when one agent instructs another.
  • The authority field distinguishes autonomous actions from human-confirmed ones.
  • The log feeds the same audit trail and retention as other security-relevant events.

4. Agentic AI governance sign-off checklist

Spreadsheet

The gate before an agent goes live, with business, risk, and legal each accountable for part.

Template

ItemOwnerSignedNote
Autonomy boundary spec complete and approvedBusiness + Risk
Dedicated identity provisioned with least privilegePlatform / IAM
Action logging verified end to end in stagingSecurity
Kill switch tested, reaches sub-agents and spawned processesSecurity
Irreversible actions gated or prohibitedRisk
Regulatory review (does the agent make Art. 22 decisions, take regulated actions?)Legal
Monitoring and alerting for scope drift configuredSecurity
Rollback and incident path definedBusiness + Security

Worked example

ItemOwnerSignedNote
Autonomy boundary approvedHead of Support + CROYes2026-09-02
Least-privilege identityPlatform IAMYessvc-copilot-prod
Action logging verifiedSecurityYesstaging trace attached
Kill switch testedSecurityYesreached the 1 sub-process; log attached
Irreversible actions gatedCROYessend + refund gated/prohibited
Regulatory reviewLegalYesno Art. 22 decisions; drafts only
Scope-drift monitoringSecurityYesalert on any new scope
Rollback / incident pathHead of Support + SecurityYesdisable feature flag; IR runbook linked

Acceptance criteria

  • Business, risk, and legal each sign the items they own; no single approver covers all.
  • The kill switch is tested, not assumed, with evidence attached.
  • A live agent has a completed, dated checklist in its file.

Governance controls this kit produces evidence for

Completing the artifacts above gives you a head start on the evidence requirements for these controls.

AGT-001
AGT-001

The autonomy boundary spec and identity register are the agent permission boundary definition.

AGT-005
AGT-005

The "requires human confirmation" and "prohibited" settings are the human approval gate for irreversible agent actions.

AGT-006
AGT-006

The action log schema is the agent action audit trail.

AGT-007
AGT-007

The autonomy boundary spec defines the agent's scope and task boundaries.

AGT-016
AGT-016

The sign-off checklist is the agentic deployment readiness assessment.

This kit backs one playbook. Read the full guidance for the reasoning behind each artifact.

Decide what to implement next

Assess your governance gaps, then create an action plan with owners and target dates. Build and export without an account; sign in when you want to save your plan.

Start the AI governance assessment →