A customer forwards an email your team sent yesterday. It says the service includes a replacement part at no additional charge. Your actual agreement excludes that part. An employee accepted an AI draft without catching the error, and the customer has planned around the promise.
The next decision concerns the customer and the affected workflow. Give one person authority to contain the problem, another clear responsibility for correcting it, and a named owner for the restart decision.
A response timeline you can adapt
The following hypothetical timeline illustrates an operating response. Its times are examples, not a service-level promise or legal requirement.
| Point in the response | Action | Owner and evidence |
|---|---|---|
| 09:10: customer flags the promise | Preserve the exact sent message, related agreement, and source material available to the drafter. | Account manager creates the incident record without changing the original message. |
| 09:20: scope the affected use | Pause AI-assisted price and inclusion statements for this service while staff use the approved reference. | Operations lead records the pause and tells the people who prepare these messages. |
| 10:00: identify other outputs | Review the messages produced from the same template or source version. | Team lead records the bounded affected set and unresolved cases. |
| Customer correction | Explain the specific error and propose a resolution the authorized business owner approves. | Account manager sends the approved correction and records the customer's response. |
| Restart review | Test the corrected source, drafting behavior, and staff acceptance process. | Sponsor decides whether to resume, narrow the use, or continue manual handling. |
The pause concerns the affected use. Your team may continue unrelated drafting if the evidence supports that boundary. If you cannot establish the boundary, widen the containment while investigating.
Find the source of the wrong promise
Compare the sent wording with the actual approved material. Did the agent retrieve an outdated price sheet? Did it infer an inclusion from an ambiguous paragraph? Did the employee change the draft after review? Those causes require different repairs.
Preserve the source version and the available execution record before updating the workflow. Your operator needs enough evidence to reproduce the error. Limit access to the people handling the incident and avoid copying customer details into an unrestricted troubleshooting document.
Check the related outputs by a defensible rule, such as the same template version and service. “Review every email we ever sent” may be impractical; “review only the reported email” may miss the shared fault. Write why the selected set fits the evidence and record where uncertainty remains.
Correct the customer-facing fact
The customer needs to know which statement was wrong and what your company will do about it. A long explanation of model behavior does not resolve their scheduling or cost decision.
The account manager can prepare concise wording: “Our message incorrectly stated that the replacement part was included. The agreement lists that part separately. We are reviewing the impact on your appointment and will confirm the approved resolution with you.” The business owner should approve any commercial remedy before staff promise it.
Use appropriate specialist advice when an agreement, regulated activity, or affected customer's rights require it. This operating example does not determine legal obligations or justify withholding a material correction.
Require a restart decision, not a patched prompt
Run the original failure case against the repair and include a missing-source example. The team should show that staff can identify the approved inclusion and that the workflow routes uncertainty for review. Test the acceptance step as well as the generated text.
The sponsor's restart record should state the approved use, remaining exclusions, test evidence, reviewer, and monitoring owner. If the team cannot prevent unsupported inclusion statements, it can keep AI for formatting while a person writes the commercial terms.
NIST's generative AI profile offers voluntary risk-management context. Our decision-rights guide addresses ownership, and fractional CAIO support can help turn an incident into a defined operating correction.

