When a Client Corrects the AI, Update the Source

Treat the correction as new evidence

Keep the client’s original message and the correction. The correction becomes the current fact; the earlier statement becomes superseded history. Record who made the correction, when it applies and which statement it replaces.

That distinction matters because assistant memory is often a summary built from several sources. OpenAI’s current Memory guide lets a person correct a remembered item, while warning that hiding or deleting the memory does not necessarily remove the underlying chat, file, email or connected source. Anthropic’s guide similarly lets a person edit or delete a memory topic, and keeps citations back to past chats when search supplies the context.

An edited answer is only one output. The source record is what the next answer, draft or handoff needs.

Write the current fact so it can survive a handoff

Turn the correction into one plain statement another operator can use without reopening the whole thread:

Dana owns commercial routing as of November 1. Marco continues to cover Ontario through November 6. This replaces the October 12 note naming Marco as the routing owner.

The sentence carries the current owner, effective date, temporary exception and earlier version. Link it to the client’s message or meeting note so the next person can inspect the wording.

Keep interpretation separate from what the client said. If the consultant believes Dana also owns the fallback inbox, record that as an open question until the client confirms it.

Trace where the old fact traveled

The old owner may have reached more than the assistant’s last answer. Search the work created after the earlier note and before the correction.

• Open scope and request drafts that name the owner or use that person’s approval.
• Client updates, follow-ups and handoff files still waiting for review.
• Tasks, reminders or schedules assigned from the old statement.
• Anything already shared with the client that now needs a short correction.

Do not silently rewrite a sent update or an approved scope. Keep the original artifact and add the correction beside it. If the change affects authorization, timing or the work itself, reopen that decision with the right person.

This is ordinary change control applied to AI-assisted work. NIST’s generative-AI profile recommends tracing the origin and modification of digital content, reviewing sources and citations, and maintaining feedback loops between provenance records and human reviewers.

Keep incomplete corrections uncertain

Client corrections often close one gap and expose another. Dana owns routing now. Marco covers Ontario through Friday. Nobody has said who covers Ontario on Monday.

Save the two confirmed facts and mark Monday’s coverage as unresolved. The assistant can then ask for the missing decision instead of promoting the most likely person. A useful draft might say, "Ontario coverage after November 6 still needs an owner."

Keep the correction narrow. A new owner does not change the deadline, approval or fallback rule unless the client said it does. This prevents a small edit from spreading into unrelated facts that merely appeared in the same conversation.

Run the next-answer test

Once the record is current, start a fresh conversation in the same client workspace or project. Ask who owns the work, when the change took effect and what remains unresolved. Open the cited source instead of judging the answer by tone.

• The answer uses Dana as the current owner and keeps Marco’s Ontario exception time-bound.
• The old note remains visible as history but is not presented as current.
• A draft that used Marco is marked for review or is found in the manual search.
• A conversation for another client does not receive either version of the fact.

Also test the controls on the plan your practice uses. OpenAI notes that its source display may not show every factor behind a response. Claude says a topic edit applies from the next conversation onward, while deleting a conversation does not automatically remove memory created from it. Those boundaries are easy to miss when the correction appears to work in the current chat.

Fix the record before you ask for another draft

The next time a client writes "small correction," save the message with the fact it replaces. Review the open work that used the old version, then run one fresh answer with the source visible.

It takes a few minutes. It also keeps Friday’s update from reopening Tuesday’s mistake.