AI human review matrix
Not every AI output needs the same checking. This free AI human review matrix sets the level of human review each output needs, from a quick self-check to expert sign-off.
Runs in your browser. Nothing you enter is sent anywhere or saved. A starting draft only, not legal advice.
Your AI outputs
List each kind of work AI helps produce, then say who sees it, how bad a mistake would be, and whether it can be undone. The examples are only samples, so change them to fit your work.
Your review matrix
| AI output | Review level | Who reviews | What to keep |
|---|---|---|---|
| Brainstorming and first drafts | 1. Self-check | The person who used the AI | None needed |
| Meeting summaries for the team | 2. Sample check | The author, plus a colleague who samples | Sample log |
| Customer support repliesHigher because customers or the public will see it. | 3. Second-person review | A colleague who didn’t produce it | Reviewer’s name and date |
| Code that goes liveHigher because a mistake could be costly. | 3. Second-person review | A colleague who didn’t produce it | Reviewer’s name and date |
| Financial reports and forecastsHigher because a mistake could cause severe harm. | 4. Expert sign-off | A qualified expert, such as legal, finance, clinical, or security | Signed approval kept on file |
| Screening job applicantsHigher because it affects decisions about a person, a mistake could be costly. | 5. Human decides | A named, accountable person | Decision and reasons recorded |
What each level means
- 1. Self-check
Read it, check it makes sense, and edit before use.
- 2. Sample check
The author reviews every output. A colleague checks a sample, such as 1 in 10, on a regular schedule.
- 3. Second-person review
A second person reviews each output before it is sent, published, or relied on.
- 4. Expert sign-off
An expert checks and approves each output, and nothing goes out without that approval.
- 5. Human decides
AI may inform the decision, but a person makes it, weighs it against other evidence, and can explain it. Affected people can ask for a review.
How to use it
- List your AI outputs. Add each kind of work AI helps produce, such as customer replies, code, or reports. Start from the examples if you like.
- Rate each one. Say who sees it, how bad a mistake would be, and whether it can be undone. Check the box if it informs a decision about a person.
- Copy the matrix. Paste the table into your AI policy or runbook, or download it as a Markdown file. Review it with the people who own the risk.
Human review of AI, explained
- What is human review of AI output?
- A step where a person checks what an AI produced before it is used, sent, or acted on. How much checking is needed depends on the stakes: a list of ideas needs far less than a customer contract or a hiring decision. People often call this keeping a human in the loop.
- What is a human review matrix?
- A table that lists the kinds of work AI helps with and sets the level of human review each one needs, who does it, and what record to keep. It turns a vague rule such as “review high-risk output” into something people can follow.
- How does the tool choose a level?
- It scores each output on three things: who sees it, how bad a mistake would be, and how hard it is to undo. The cost of being wrong counts for the most. Anything with severe harm gets at least expert sign-off, and anything that informs a decision about a person gets a named human decision-maker, whatever the score.
- Which AI outputs always need review?
- Anything customers or the public will see, anything that spends or commits money, code that goes live, legal and financial material, and anything that informs decisions about people. Even low-risk work benefits from the author reading it before use.
- Is reviewing every output realistic?
- Not always, and the matrix doesn’t ask for it. Lower levels rely on the author’s own check and regular sampling. Save second-person and expert review for the work where a mistake is costly or hard to undo.
- What should a reviewer actually check?
- Facts, figures, quotations, and sources against reliable references. Whether code runs and is safe. Whether the tone, claims, and commitments are ones your organization can stand behind. Reviewers should be able to say no and have it stick.
- Does this meet legal requirements for human oversight?
- Not by itself. Some laws and regulations require specific human oversight for certain decisions, such as those about employment or credit. Use the matrix as a starting draft, adapt it to your rules, and ask legal counsel about regulated uses.
I’m Courtney Stephenson, an AI strategy and implementation leader, and I built this tool. Get in touch if you’re setting up review for AI work. To put the rules in writing, try the AI acceptable use policy generator.
Want a second opinion on your results?
Book a free 15-minute call and we’ll talk through what this means for your team.