“Someone will check it” is an incomplete workflow. Before an AI-assisted draft reaches a customer, decide who reviews it, what they are checking and what happens when the output is not ready. A short responsibility sheet makes those decisions visible without pretending that a small checklist solves every AI risk.
Start with one narrow task
Describe a task in one sentence: for example, preparing a draft response from an approved product FAQ. Keep the boundary clear. Drafting a response, deciding whether a refund is allowed and sending the final message are different actions; do not quietly bundle them into one instruction.
NIST describes its AI Risk Management Framework as voluntary guidance for considering trustworthiness in AI design, development, use and evaluation. The worksheet here is a small editorial planning aid inspired by that general risk-management approach. Completing it does not establish compliance, certification or endorsement by NIST.
Name the people and the decision
Write down a task owner who maintains the workflow and a reviewer who decides whether a particular output is ready. One person may hold both roles in a small business, but the review still needs an explicit step.
Add an escalation contact for questions the reviewer cannot resolve. “Ask the product owner about conflicting specifications” is more actionable than “use good judgment” when a draft gives two different answers.
Define three stop conditions
Choose conditions that a reviewer can recognize. For the FAQ example, a team might stop a draft when:
- It makes a promise that is absent from the approved source.
- It includes a factual detail that the reviewer cannot verify.
- It answers a request outside the task's agreed scope.
These are illustrative checks, not a complete safety policy. Adapt them to the consequences of the task. If the reviewer lacks the expertise to assess an output, changing a label from “AI draft” to “reviewed” does not solve that gap.
Record the correction, not just the approval
Use four fields for each reviewed example: input reference, issue found, correction made and final decision. Keep confidential information out of a shared example log unless the organization has approved that use.
A correction log can show whether the same problem recurs. If drafts repeatedly invent a delivery date, the owner can change the source material or narrow the task before simply asking reviewers to be more careful.
Try the process before increasing volume
Use a small batch of approved examples to see whether the reviewer can apply the checks consistently. Do not describe the process as tested until that work has actually happened. If review takes longer than writing the response directly, record that observation rather than counting only generation speed.
Use the repeatable AI comparison worksheet to keep a later trial structured.
Source and scope
Background: NIST AI Risk Management Framework. The responsibility sheet, stop conditions and FAQ example are ShamsMag editorial suggestions, not an official NIST template. No product performance result is claimed. Cover: AI-generated illustration.
