Template · Checking the work
AI output verification checklist template
The cheatsheet for checking AI output applies to any draft. This template is narrower. It builds a checklist for one deliverable you produce again and again, such as a weekly report, a contract summary, or a customer reply.
A recurring deliverable contains the same kinds of claims each time. Decide once how each kind gets checked, and the check becomes routine instead of a judgment made under deadline.
When to use it
- You produce the same kind of document with AI every week or month.
- An error in the document would reach a client, a regulator, or leadership.
- More than one person produces the document and the checks differ between them.
The template
Deliverable
Which recurring document does this checklist cover?
What an error would cost
Who relies on it, and what happens if it is wrong?
Claim types
What kinds of specific claims does it contain?
Check for each type
How does each claim type get verified?
Must-check items
Which claims get checked every time, however short the deadline?
When a check fails
What happens to a claim that turns out wrong or cannot be confirmed?
Sign-off
Who confirms the checks were done before it ships?
A filled example
A paralegal builds a checklist for the AI-assisted supplier contract summaries the legal team prepares for the commercial team.
- Deliverable
- One-page summaries of incoming supplier contracts, sent to the commercial team before signature.
- What an error would cost
- The commercial team signs on the strength of the summary. A missed renewal term could lock us into a second year.
- Claim types
- Dates and notice periods. Payment amounts. Liability caps. Renewal and termination terms. Quoted clause text.
- Check for each type
- Dates, amounts, and caps: compare with the clause in the signed copy. Quotes: compare word for word. Renewal terms: read the full clause, not only the summary line.
- Must-check items
- Auto-renewal terms, notice periods, and the liability cap.
- When a check fails
- Correct it from the source. If one claim was wrong, recheck every other claim from the same section. If a claim cannot be confirmed, remove it and flag the gap.
- Sign-off
- The paralegal who prepared the summary initials the checklist. The supervising lawyer reviews any summary with a flagged gap.
What each field is for
Deliverable
Skill: Checking the work
A checklist tied to one deliverable can be specific. A general one ends up saying “check the facts,” which tells nobody what to do.
What an error would cost
Skill: Checking the work
The cost of an error sets how much checking the document deserves. Writing it down keeps the checking in proportion to the stakes.
Claim types
Skill: Checking the work
Errors cluster in specific claims such as figures, dates, and quotes. Listing the types tells the reviewer where to look.
Check for each type
Skill: Checking the work
A named method makes the check independent of the model. Asking the model whether it is sure re-runs the same blind spots.
Must-check items
Skill: Checking the work
Under deadline pressure, people check whatever is easiest. Naming the claims a decision rests on protects them first.
When a check fails
Skill: Checking the work
One false claim is a sign the model was guessing nearby. A fixed rule for failures stops a single correction from hiding a wider problem.
Sign-off
Skill: Getting real work done
Checked work needs a named owner. A sign-off line turns “someone looked at it” into a record of who did and what they checked.
Common mistakes
Checking the parts that are easy to check
Typos and formatting are quick to fix and rarely matter. Start with the claims the decision rests on, even when they take longer.
Using the model as the check
A second pass by the same model in the same conversation is not verification. Use the source document, the data, or your own calculation.
Never updating the list
When an error gets through, add its claim type or tighten its check. A checklist that never changes stops matching the errors you actually see.