What Does “Done” Mean When AI Creates the First Draft?
Updated: 3 days ago

The AI writes that a vendor renewal is approved. The source note says finance will review it Friday. Your team can forward a clean summary in seconds, and the first person to notice the difference might be the vendor.
Define “done” as a decision a named reviewer can make: this exact draft meets these checks and can move to this destination. Give the assistant a separate status for “draft ready.” If a check fails, leave a useful note for the reviewer and keep the draft where it is.
An approval that hasn’t happened yet
Imagine an operations coordinator using AI to prepare a vendor renewal brief. The assistant has a current quote and notes from an internal meeting. The quote says it expires November 30. The notes say the finance lead will review the cost on Friday. The renewal workflow needs three things before the account manager can send a recommendation: the verified cost, an unambiguous deadline and an approval record.
What the first draft says: “Finance approved the renewal. Please confirm the vendor’s new term before November 30.”
What the reviewer can verify: November 30 is the quote’s expiration date. The notes name a reviewer and a meeting date. They don’t contain a finance decision, and the internal deadline for the recommendation hasn’t been set.
The coordinator marks the draft Needs evidence, rewrites the status as “Finance review pending,” and asks the finance lead to confirm the internal deadline and record a decision. The brief stays internal. Even if the cost and expiration match the quote perfectly, those checks don’t make the approval real.
Put the checks where the decision happens
The Scrum Guide describes a Definition of Done as a shared description of the state a deliverable reaches when it meets its quality measures. The GOV.UK service manual describes acceptance criteria as outcomes you use to check whether a service meets a user need. A small workflow can borrow the distinction: use the same basic quality rules every time, and name any extra conditions for this particular output.
A finish line you can copy
Output: Renewal brief, version 3, in the internal vendor record.
Check against: The current quote for the cost and expiry date; the finance approval record for the decision; the project record for the internal deadline.
Reviewer: Operations checks the sources and wording. Finance owns the approval.
Pass: Each claim matches its record. The finance decision and internal deadline are present. An authorized person clears the exact version and recipient before anything goes out.
Hold: Label the missing evidence, give it to the person who can provide it, and keep the draft internal.
Apply this to one output before standardizing the whole workflow. The permission check for connected AI helps when a draft may be visible to more people than the reviewer expects.
A review queue can make that finish line visible. The ClickUp dashboard decision example separates a completed review from completed customer work, with two included tasks and two exclusions you can check.
Give the reviewer two versions to catch
Take an old, redacted brief whose quote, approval record and internal deadline the reviewer can open. It should pass. Give the reviewer a second version with the same layout but one important fact changed, such as replacing the approved amount with a different one. This one should stay on hold. Watch which source they open and where they hesitate.
If they catch the change only because they remember the old job, improve the check. For example, require a link to the decision record beside the approval sentence. Then try a case where the decision record exists but the reviewer cannot open it. The right response is to hold and ask the owner for appropriate access. Use the exception queue guide to give unresolved cases an owner and a next step.
If the same missing fact keeps reaching the reviewer, use the guide to rewriting an intake question to test whether you can collect it earlier.
Once reviewers use the same finish line, measure the rework after a draft comes back. Count the extra review and correction time before deciding whether the workflow saves anyone work.
Questions teams ask about AI review
Can AI tell me whether its own draft is done?
It can compare the draft with a checklist and point to missing evidence. Keep the acceptance decision with a person when the output will commit the business, reach a customer or change a record.
Do I need a new checklist for every task?
Reuse short checks for source accuracy, permissions and destination. Add specific criteria when the output has its own risk, such as approval evidence or a customer-facing deadline.
What should the team do with a failed draft?
Keep it internal. Record which check failed, link to the source and assign the next decision to someone who can make it.
Bring one draft people keep handing back
Take a redacted first draft, the evidence behind it and the person who can accept it. Ask Bell Consulting Solutions to review the workflow, or explore the AI process optimization guide.





Comments