AI Strategy

What to do when an AI tool gets something wrong at work

By HelpUsWith.ai Team

A customer receives an AI-assisted reply promising a refund outside the published policy. The employee who sent it checked the tone but missed the promise. What happens next?

The answer should be clearer than “ask the person who knows the AI tool.” A short response plan helps staff contain the mistake, take care of the customer, and find out whether the same problem could happen again. You can write one before your team deploys a new assistant, or add it to an existing incident process.

Make it easy to report a problem

Tell staff where to report an AI-related mistake and what information to include. Use a channel they already know, such as a service desk or an incident form. Give them a named backup if the usual owner is unavailable. Reports should include mistakes caught before release as well as those that reached a customer or another system.

An employee might report: “The assistant cited our returns policy but invented a 90-day refund window. I caught it before sending. The source article is linked in the ticket.” That gives the owner something concrete to investigate.

Ask for the task, time, tool and account used, the output, the source material available, and who may have received or acted on it. Follow your normal privacy and security rules when collecting examples: a screenshot or prompt may contain customer information. Avoid copying sensitive material into a general chat channel.

Make reporting a routine part of the work. If people expect blame for catching a problem, they may quietly fix the answer and the team will miss a pattern.

Contain the immediate problem

Decide what needs to stop while someone investigates. A reviewer may simply discard an incorrect draft. If the answer was sent, the response may include contacting the customer and correcting the record. If an assistant can update records or trigger actions, pause the affected workflow or disable the relevant permission when appropriate. Use your existing security or privacy incident process when the issue involves exposed information or unauthorized access.

Give someone authority to make that call and a backup who can act outside normal hours if the workflow runs then. Keep a manual route available for urgent work while the affected process is paused.

Not every mistake deserves the same response. A clumsy internal summary can be corrected locally. A false price, an exposed customer record, or an incorrect action in another system needs faster escalation. Agree on examples with the teams who know the work rather than relying on a single severity label for every use.

Preserve enough evidence to investigate

Record the original output and the circumstances in which it appeared, subject to your retention and access rules. Useful details include the relevant prompt or request, the source documents and their versions, the tool and configuration, and any review or action that followed. If you cannot safely retain an input, record an approved description of it instead.

AI responses can vary, so repeating the same request later may produce a different answer. A successful retry does not explain what happened in the original case. Check the source material and workflow as well as the generated text: an outdated policy article, a misleading instruction, or a reviewer without enough time may each contribute.

The NIST AI Risk Management Framework 1.0 includes post-deployment monitoring, incident response, recovery, and change management among its recommended practices. It is a voluntary framework; use it to inform a process that fits your organization.

Fix the contributing factors you find, then test the change

A correction might mean updating a source article, changing an instruction, limiting a tool’s access, or giving reviewers a more specific checklist. Sometimes the right decision is to stop using AI for that task. Record what changed and who approved it.

Before restarting, test the original scenario and a handful of related cases. For the refund example, include an eligible return, an ineligible return, and a question that the policy does not answer. Check that the assistant can leave uncertainty for a person to resolve rather than inventing a rule. A few examples can expose a recurring problem, but passing them does not prove the workflow will never fail.

After the change, review a sample of live work and give staff a way to report a recurrence. The owner should look for repeated errors, review the impact of corrections, and revisit the workflow when the tool, its permissions, or its source material changes.

Practice before you need the plan

Try a 20-minute tabletop exercise with the people who would respond. Give them a plausible scenario, such as an assistant sending an outdated delivery promise to several customers. Ask who pauses the workflow, who checks the affected messages, who contacts customers, and where the team records the incident. Note the unanswered questions and assign owners to resolve them.

Your plan does not need to be a new binder. It needs a clear reporting route, people with authority to act, and a way to check whether a fix worked.

If your team needs help preparing staff or putting an AI workflow into practice, explore our AI training and AI strategy services.

Keep exploring AI

Get more perspectives from HelpUsWith.ai.