An AI agent should not be approved to act independently simply because its demo works. Before scaling, a cross-functional team should be able to show what the agent improves, what it may decide, what it can access and how someone can intervene.
This is an author-proposed discussion worksheet, not a certified audit, validated maturity index or substitute for legal and security assessment. Evidence is more useful than an inflated score.
Download blank evidence register (CSV) ↗
Step 1. Identify the deployment
Agent / workflow: ____________________ Business owner: ____________________
Review date: ____________________ Reviewer(s): ____________________
Scope: ____________________ Deployment stage: Pilot / Limited / Scale
Step 2. Record the evidence, not just the answer
For each check, record missing / partial / evidenced, link to supporting documentation, assign an owner and capture any mitigation. A positive response without verifiable evidence is still partial. Do not add the statuses into a single maturity score.
| Check | Question | Expected evidence |
|---|---|---|
| Contract 1: Outcome | ||
| O1 | Is there a measurable end-to-end outcome, including a baseline? | KPI, baseline date, counterfactual or comparison method. |
| O2 | Is a business owner accountable for the result and process redesign? | Named owner, authority to change workflow, review cadence. |
| O3 | Are full unit economics defined beyond model or token spend? | Per-completed-outcome cost, human review, rework, platform, controls and exceptions. |
| Contract 2: Decision | ||
| D1 | Are allowed actions separated into recommend, execute and escalate? | Decision-rights matrix and thresholds. |
| D2 | Are sensitive actions protected by explicit human approval? | Approval flow, exception cases, identified approvers. |
| D3 | Is cross-agent and cross-team delegation traceable? | Delegation map, transaction identifiers, logs. |
| Contract 3: Access | ||
| A1 | Is the agent individually identifiable, with least-privilege permissions? | Service identity, access review, inventory. |
| A2 | Are data sensitivity and permitted purposes explicitly enforced? | Data classification, access policy, technical controls. |
| A3 | Can privileges be revoked promptly and independently? | Revoke / kill-switch test and log. |
| Contract 4: Assurance | ||
| S1 | Has the agent been evaluated on representative real-world cases and failure modes? | Evaluation design, results, owner and test scope. |
| S2 | Are operational monitoring, incidents and recovery responsibilities explicit? | Alerting, response runbook, named responsible parties. |
| S3 | Are there periodic reviews with authority to change, pause or stop deployment? | Review minutes, escalation rules, reapproval triggers. |
Step 3. Decide what happens next
- Proceed with defined scope: evidence supports the stated use case; accountable owners accept residual risks; applicable reviews are complete.
- Limit or redesign: open gaps require reduced permissions, human checkpoints, additional validation or clearer business ownership.
- Do not authorize autonomous action yet: required approvals, lawful basis, access restrictions, safety controls or stop/recovery capabilities cannot be demonstrated.
These are proposed decision prompts, not an exhaustive compliance checklist or legal rule. Organizations should define their own mandatory gates according to risk, sector and jurisdiction.
Step 4. Test whether value survives scale
Cost per successful completed outcome is often more meaningful than cost per token or task. Record the same metric before and after deployment, including human supervision, retries, exception handling, technology, and controls. Track quality and customer impact alongside cost so apparent efficiency does not hide harm.
Illustrative calculation (fictional): 1,000 requests processed at £2,400 total cost with 800 successfully completed to an agreed standard gives £3.00 per successful outcome. If an AI-enabled approach costs £1,800 but completes only 450 satisfactorily, its effective unit cost is £4.00, not an improvement, even though the total spend is lower. The figures illustrate the denominator problem; they are not empirical performance results.
Sources and prior work
This worksheet builds on established governance concepts. It does not claim to invent agent authorization, accountability or human oversight. The WEF/Capgemini AI Agents in Action playbook (2026) introduces the ACAP authorization profile for deployment governance. NIST's agent identity and authority work covers identification and permissions. McKinsey's 2026 AI survey documents the gap between individual productivity and enterprise EBIT reporting.
The distinct contribution proposed here is to make business ownership and full workflow-level unit economics first-class evidence alongside operational delegation, access and assurance.