AI Human Approval Matrix
Decide which AI work still needs a person to sign off. Score each workflow's risk, set none, one or two approvers, and loosen review only with evidence.
What's inside this template
Who it's for
AI Operations Leads, CIOs, COOs and business owners putting Claude agents and AI workflows to work, plus the security and data protection teams who need to know where people stay in the loop
When to use
Before an AI workflow goes live; when an approval queue starts to build; when a connector gets write access; when you want to move a workflow to lighter review
Key benefit
Human attention goes where the risk is: low-risk work flows on automated checks, high-risk work gets two people, and every change of tier is backed by evidence and logged
Sections included
- Workflow details
- Score the risk: six questions and a tier rule
- The tiers: approvals, checks and sampling
- Always-a-person list
- Approval rules
- Moving between tiers
- Guardrail measures
- Approval register
- Re-tiering opportunities (ICE)
- Steering group lines
- Implementation notes
Complete template content
To use this matrix, copy it with the "Copy page" button above or open the Google Doc version. Read the companion guide to setting AI approvals by risk for the reasoning behind each tier.
AI human approval matrix
A free template for AI Ops leads, CIOs and COOs putting agents to work. Decide how many people approve each AI workflow, set by risk: none, one or two. Use one copy per workflow, or keep one register for the whole programme (section 7). Start strict, and loosen only with evidence.
0. Workflow details
- ☐ Workflow name:
- ☐ Business owner (accountable for the outcome):
- ☐ AI Ops lead or reviewer:
- ☐ AI tool or agent used (e.g. a Claude skill, connector or scheduled task):
- ☐ Systems it reads:
- ☐ Systems it writes to, and which fields:
- ☐ Volume (items per week):
- ☐ Approval today (who signs off, and how long it takes):
- ☐ Assessment date and next review date:
1. Score the risk
Answer each question for the workflow.
| # | Question | Yes / No | Notes |
|---|---|---|---|
| 1 | Is it hard to undo once done (sent, paid, deleted or published)? | ||
| 2 | Does the output leave the organisation (customers, suppliers, the public)? | ||
| 3 | Does it move money, change a price or change a contract? | ||
| 4 | Does it touch personal data beyond what the requester's role already allows? | ||
| 5 | Is there a legal, regulatory or contractual duty attached (a regulator, an auditor, a complaint)? | ||
| 6 | Is it a decision with a legal or similarly significant effect on a person (claims, credit, hiring, disciplinary)? |
Tier rule
- No to every question: tier 0 candidate.
- Yes to question 1 or 2 only: tier 1.
- Yes to question 3, 4 or 5: tier 2.
- Yes to question 6: always a person, plus tier 2 approvals if questions 3 to 5 also apply. Involve your DPO.
2. The tiers
| Tier | Human approvals | Checks before a person looks | Sampling |
|---|---|---|---|
| 0: Low | None once checks pass | Automated checks against rules or a template; a second agent reviews with fresh context | The business owner samples a fixed number each week (e.g. 20) |
| 1: Medium | One person, who can be the person who asked the agent | As tier 0, plus a named owner for the workflow | Monthly sample by the business owner |
| 2: High | Two people, one of them the business owner or a delegate | As tier 1, plus a record of what the agent saw and did | Every item approved |
| Always a person | At least one person with the authority to change the outcome | As tier 2, plus your DPO's review of the process | Every item decided by a person |
3. Always-a-person list
Agree this list with the business owner, security and your DPO. Nothing on it moves to a lighter tier without a decision recorded here.
- ☐ Payments and refunds above £____
- ☐ Pricing, contract and commercial terms
- ☐ Decisions with a legal or similarly significant effect on a person
- ☐ Statements in the organisation's name to regulators, the press or the public
- ☐ Deleting records
- ☐ Changes to permissions, connectors or the agent's own instructions
- ☐ Other:
4. Approval rules
- ☐ The work sets the tier, not the person who built or runs the agent.
- ☐ An agent's approval never counts as a person's. Automated checks can clear tier 0, but they can't fill a human slot.
- ☐ Whoever asked the agent for the output owns the result and counts as one approver.
- ☐ An approval covers one version. If the output changes after approval, approve it again.
- ☐ Approvers must be able to change or stop the outcome, not just acknowledge it.
- ☐ Write access is phased system by system and field by field, inside existing role permissions.
5. Moving between tiers
Start strict. Move a workflow to a lighter tier only when the evidence below is met, and record the decision. The thresholds are starting points: set your own.
| Measure | To move down a tier | To move back up |
|---|---|---|
| Share of outputs edited before approval | Below 10% for 4 weeks (tier 2 to tier 1); below 5% for 4 weeks (tier 1 to tier 0) | Above the threshold for 2 weeks in a row |
| Errors or reversals after release | No increase on the baseline | Any material increase |
| Complaints or escalations linked to the workflow | No increase on the baseline | Any increase |
| Items reviewed in the period | At least ____ | n/a |
- ☐ Decided by:
- ☐ Date:
6. Guardrail measures
Record a baseline before you change a tier, then track it monthly. If errors rise, you haven't saved time. You've moved it.
| Measure | Baseline | Month 1 | Month 2 | Month 3 |
|---|---|---|---|---|
| Approver hours per week | ||||
| Average wait for approval | ||||
| Share of outputs edited | ||||
| Errors or reversals | ||||
| Complaints or escalations |
7. Approval register
One row per workflow. The first row is a fictional example.
| Workflow | Tier | Approvals | Approver role(s) | Edit rate | Owner | Next review |
|---|---|---|---|---|---|---|
| Example: customer claim status updates | 1 | 1 | Handler who asked for the draft | 6% | Head of claims | In 4 weeks |
8. Re-tiering opportunities (ICE)
Treat each change of tier as an opportunity. Score impact, confidence and ease, each from 1 to 10, and multiply them for a score out of 1,000. Do the highest first, and record the approver time it returns as an outcome.
| Workflow | Proposed change | Impact (1-10) | Confidence (1-10) | Ease (1-10) | ICE score | Outcome to record |
|---|---|---|---|---|---|---|
9. Steering group lines
- Workflows re-tiered this month:
- Approver time returned:
- Workflows held at their current tier, and why:
- Decisions needed from the steering group:
10. Implementation notes
- Start with one high-volume workflow where the approval queue already hurts.
- Review tiers when a workflow changes, a connector gets write access or a release changes what your AI tools can do.
- Keep this matrix alongside your AI policy and approved tools list.
- This template isn't legal advice. For decisions about individuals, involve your DPO.
Get started
How to use this matrix
Pick one workflow
Start with a high-volume AI workflow where the approval queue already hurts. Fill in the workflow details, including what it reads and writes.
Score the risk
Answer the six questions. Can you undo it? Does it leave the building? Money, personal data, a legal duty, or a significant decision about a person? The tier rule gives you a starting tier.
Set the approvals
Tier 0 needs no approver once checks pass, tier 1 needs one person, tier 2 needs two. Anything on the always-a-person list keeps a person with the authority to change the outcome.
Start strict, loosen with evidence
Record a baseline, then move down a tier only when the edit rate, errors and complaints meet the thresholds you agreed. Move back up if they rise.
Log it and report it
Add the workflow to the approval register, score re-tiering with ICE, record the approver time returned as an outcome, and put one line in your steering group update.
Questions
Frequently asked questions
Common questions about setting human approvals for AI work
Why not just have a person approve every AI output?
Where does the none, one or two approvers model come from?
Is it legal to remove a human sign-off?
What evidence should we need before loosening review?
How does this fit with use case prioritisation?
Does Claude Code Review approve pull requests?
More templates
Related templates.
Ready when you are
Want help setting tiers for your first agents?
Kowalah is a UK Anthropic Implementation Specialist. We help AI Ops leads and CIOs put Claude agents to work with the right people in the loop, from the first workflow to the steering group. If you'd like help, book a conversation.