You caught the AI mistake. Who else needs to know?

Before the next draft, the person responsible for the task needs to know what failed, how it was corrected and whether the instructions need changing.

For a routine drafting mistake, keep the response simple: hold the draft, check the source and record the lesson in your existing task log.

Issues 001 to 003 covered information boundaries, checking answers and measuring the finished work. This issue deals with what happens when a check fails.

For an ordinary drafting error, keep the message unsent while you check it. Do not copy the same claim into another email, quotation or website update.

EXAMPLE

Approved service text: “Delivery dates are confirmed when you book.”

AI draft: “Your parcel will arrive tomorrow.”

The draft has added a commitment. Hold it until the date is confirmed or remove the unsupported promise.

If the message has already been sent, tell the person responsible for that customer communication so they can decide the correction needed.

Asking AI “Are you sure?” is not a check against your business information.

Compare the claim with the approved source. If that source does not settle the question, ask the person authorised to decide. Leave the draft on hold if nobody can verify it.

EXAMPLE

A checked replacement could be: “We will confirm your delivery date when you book.”

Then review the rest of the reply. Fixing one sentence does not establish that the remaining claims are correct.

For a repeated problem, test one change to the instruction, such as: “Do not state a delivery date unless one is explicitly supplied.” Keep human checking in place. A stronger instruction cannot guarantee a correct answer.

Use a short entry in an existing task log rather than creating another system.

Tell the person responsible for the task about the correction, so they can decide whether the instructions or scope need changing.

For routine drafting mistakes, record:

EXAMPLE LOG ENTRY

Task: public-information reply.
Issue: ADDED PROMISE.
Caught before use: yes.
Change: require a supplied date before including one.
Owner: task supervisor.
Review: after the next three trial drafts.

Keep names, contact details and copied customer messages out of this routine learning log. Store it in an approved business location.

If personal or confidential information may have been exposed, or the output could cause harm, stop the task and contact the person responsible for your business’s incident response promptly. Follow that process for handling evidence. This short log is not an incident response plan.

Evidence: The US National Institute of Standards and Technology (NIST) Generative AI Profile (2024), action MG-4.3-002, recommends procedures for recording and tracking reported errors and near-misses. The three steps and log above are our practical adaptation for a routine drafting task, not a NIST checklist or a UK legal requirement.

Correct the answer and keep the lesson.

A mistake caught before sending is useful feedback. Use it to decide whether to change the task, narrow its scope or stop using AI for it.

For one approved low-risk task this week, add the stop, check and record rule to your existing instructions. During the next three trial drafts, count how often the same issue returns.

If an important claim cannot be checked, stop that use of AI and do the task manually or refer it to someone qualified to decide.

Barbara @ AI Made Human

P.S. If no mistakes appear in three drafts, that is a small trial result. It does not prove the task is reliable.