What to Do When Your AI Agent Gets It Wrong

Direct answer

Run four steps. 1) Contain: pause or constrain the flow that produced the error. 2) Correct: contact the customer yourself, fast, with the right information and a clear apology. 3) Fix the source: update the knowledge base or instructions that led to the mistake. 4) Tighten the guardrails: add human approval for that topic going forward. Document the incident and review it monthly.

Why this is a legal issue, not just a quality issue

In February 2024, the British Columbia Civil Resolution Tribunal ruled in Moffatt v. Air Canada (2024 BCCRT 149) that Air Canada was liable for misinformation from its own website chatbot. The chatbot had invented a retroactive bereavement fare policy; the customer relied on it and was refused a refund. Air Canada argued the chatbot was a separate entity responsible for its own actions. The tribunal rejected that argument and ordered the airline to pay damages. The practical rule for any business: your company owns everything its AI says to customers.

Step 1: Contain the error

Pause the affected conversation flow or restrict the topic. You do not need to shut down the whole agent; you need to stop the specific wrong answer from reaching the next customer. Most agent platforms let you disable a single intent or add a temporary human-review step.

Step 2: Correct it with the customer

Speed wins trust. Reach out personally, state the correct information clearly, and make the customer whole where reasonable. Customers forgive honest corrections quickly; they remember silence and blame-shifting. Your message should be short: what was wrong, what is right, and what you are doing about it.

Step 3: Fix the source

Every agent error traces back to one of three places: the knowledge base is missing or outdated, the instructions are ambiguous, or the agent had no approved answer and improvised. Find which one, fix it, and add the corrected Q&A as an approved response. Then test the same question five ways to confirm the fix holds.

Step 4: Tighten the guardrails

Guardrails are cheap compared with liability. The Air Canada case turned on a policy question the chatbot should never have answered freely.

The monthly incident review

Keep a simple log: date, what happened, customer impact, root cause, fix applied. Review it monthly with your agent owner. Error patterns repeat until someone reads the log; the review is what converts individual mistakes into system improvements.

Related reading

Frequently asked questions

Who is responsible when an AI agent gives wrong information?

Your company is. In Moffatt v. Air Canada (2024), a tribunal ruled the airline liable for its chatbot's invented policy. "The AI said it" is not a legal shield.

What should I do first when the agent makes a mistake?

Contain it: pause or constrain the affected flow, then contact the customer yourself with the correction. Speed and ownership protect trust better than any excuse.

How do I prevent repeat errors?

Fix the source: update the knowledge base, add the correct answer as an approved response, and add a guardrail rule (human approval for that topic). Log the incident for the monthly review.

Should I turn off the agent after an error?

Only for the affected topic. Pause that flow, fix the source, test it, and re-enable. Full shutdowns usually punish customers for a knowledge problem, not a system problem.

Want AI agents with guardrails and human oversight built in from day one? Talk to inicIA.

Talk to inicIA on WhatsApp

By inicIA