Use Your Experience to Check AI Work

Years of work can teach you which detail will cause trouble: a promise no one approved, a missing exception, or a deadline that ignores how a process actually works. That knowledge can help you review an AI draft.

It does not mean that experience is always right or that AI can never handle complex work. Your memory can be out of date too. The useful skill is knowing what needs checking, where the reliable information is, and who has authority to decide.

Choose a task you understand

Start with a short draft related to familiar work. Examples include a customer reply, a handover note, an explanation of a procedure, or a summary of agreed actions.

Use an approved tool and information you are permitted to share. If you are practicing, use fictional material. Keep the source information beside the AI output so that checking does not depend on memory alone.

The NIST generative AI risk profile describes risks from confidently incorrect content and from people relying too heavily on AI. A fluent answer can still contain invented details or unsupported conclusions.

A worked example: an unapproved promise

Consider this fictional customer-service exercise. The source notes say:

• A customer has asked about a replacement part.
• Availability has not yet been checked.
• Delivery is estimated at three to five business days after dispatch.
• A manager must approve any shipping-fee waiver.

Suppose the AI draft says: “Your replacement will arrive on Friday, and we have waived the shipping fee.” This is an invented example of a faulty draft, not a report of a test we conducted.

A reviewer should catch three problems. The draft assumes the part is available, turns an estimate after dispatch into a guaranteed arrival date, and claims an approval that has not happened.

A more appropriate draft might be: “I will check whether the replacement part is available and confirm the estimated delivery once dispatch details are known. I will also ask whether we can waive the shipping fee.”

Even this version needs a person to accept responsibility for those follow-ups. Editing the wording has not checked stock, arranged a shipment, or obtained approval.

Check your result

Try classifying the faulty sentence before reading the check: which claims need evidence, and which need approval?

Review facts, assumptions, and authority

Work through the draft in three passes:

1. Facts: compare names, dates, quantities, requirements, and source references with current records. Recalculate any important arithmetic.
2. Assumptions: look for information the source did not provide. Check whether words such as “may,” “estimated,” or “subject to approval” disappeared.
3. Authority: identify each promise or action. Who is allowed to approve it, and has that approval actually been given?

For a handover note, a task being discussed is not necessarily a task assigned. For a procedure, a common practice is not necessarily an approved rule. For a customer reply, polite wording does not make a promise accurate.

Ask what changed since the source was written. An old document can be faithfully summarized and still be unsuitable for today’s decision.

Use AI to help expose questions

You can ask for a structured review without treating the tool as the final judge:

Compare this draft with the reference notes. For each statement that is unsupported, uncertain, or dependent on approval, quote the relevant wording and identify the missing evidence. Preserve qualifications and exceptions. Suggest a revision using only confirmed information. If the notes do not resolve something, put it under Questions. Do not contact anyone, approve an action, or change a record. Reference notes: [insert permitted information]. Draft: [insert text].

Then inspect the comparison yourself. Asking the same assistant whether its own answer is correct is not independent verification. Check the actual document, calculation, responsible person, or other relevant source.

When the task is outside your competence, involve someone qualified to assess it. Experience in one area does not establish expertise in another.

Keep people’s circumstances out of guesswork

An AI draft may sound empathetic while guessing why someone behaved a certain way. Replace assumptions about motives or emotions with what the person actually said and a question where clarification is needed.

For example, “The customer is being unreasonable” is a judgment. “The customer says the replacement date will not meet their needs” records useful information. Your next step is to understand the requirement and confirm the available options.

Record the contribution you made

Keep a brief note of the issue, the evidence you checked, and the correction. If you measured time, include the review and follow-up work. Do not count a draft as completed work while its important questions remain unanswered.

These notes can show where your judgment improved a result. With appropriate permission or fictional material, you can turn one example into a small AI skills portfolio.

For your next practice session, take five sentences from a permitted draft and mark each as supported, uncertain, or requiring approval. Our guide to checking an AI answer gives additional steps for verifying factual claims.

Featured photograph by David Travis on Unsplash. Photograph used for illustration.