What to Do When Your AI Agent Gets It Wrong (2026 Guide)
Last updated: August 2026.
Your AI agent will get something wrong. Not might — will. It will quote a price that changed, book someone into a slot that was not available, or say something confidently incorrect. The question worth preparing for is what happens in the hour afterwards.
On This Page
Written by Dr Priya Jaganathan — Go High Level Certified Admin, Certified AI Tech Stack Consultant and keynote speaker. This is an operations guide, not legal advice — take advice on your specific circumstances.
The first hour
Honour it if you reasonably can. If the agent quoted $180 and the correct price is $220, the cheapest resolution is usually to do the job for $180 and fix the rule. The margin lost is smaller than the review you avoid.
Apologise without blaming the system. "Our system gave you the wrong information and that's on us" lands well. "The AI made a mistake" sounds like an excuse and invites the customer to wonder what else it is getting wrong.
Find out who else was affected. This is the step most businesses skip and it is the one that matters. An agent that gave one customer a wrong price gave it to everyone who asked that question since the price changed. Search your transcripts.
Switch off the affected path if the error is serious. If the system is giving out incorrect safety, health or compliance information, disable that path immediately rather than waiting for a fix.
Who is responsible for what it said
Businesses sometimes assume an automated statement carries less weight than one from an employee. That assumption is not safe.
Australian Consumer Law prohibits misleading or deceptive conduct in trade or commerce. It does not contain an exception for statements generated by software you deployed. If your agent told a customer something incorrect about your goods or services, that is your business making the statement.
Treat anything your system says as though a staff member said it, because commercially and reputationally that is how it will be treated.
This is also why the configuration boundaries discussed throughout this blog matter. A system that never gives advice cannot give wrong advice, which is a far better position than relying on it to give correct advice consistently.
| Error type | Usual cause | Response |
|---|---|---|
| Outdated information | Business changed; system did not | Honour it, update source data |
| Booking error | Calendar or duration misconfigured | Accommodate, fix the rule |
| Answered something it should not | Escalation trigger missing | Correct with the customer, add trigger, test |
| Invented a detail | Asked beyond its data | Constrain to known data; teach it to say "I don't know" |
| Tone failure with a distressed customer | No sentiment escalation | Human contact now; add escalation |
If you'd like your escalation rules stress-tested before they matter, book a CRM transition call.
Fixing the cause, not the conversation
The instinct after an error is to fix that customer's situation and move on. That guarantees a repeat.
Every mistake points at one of four causes: the source data was wrong, a rule was missing, an escalation trigger did not fire, or the system was asked something outside its knowledge and answered anyway.
Identify which, then fix it at that level. A wrong price means your price data is stale — which means every price the agent quotes is suspect until you check. A missing escalation means there is a category of question you have not thought about, and there are probably adjacent ones.
Then test the fix adversarially. Ask the same question five different ways and confirm the new behaviour holds. A fix that works for the exact phrasing that failed is not a fix.
The four error categories
Stale data is the most common by far and the easiest to prevent — a quarterly review of what changed in your business, reflected in the system.
Missing rules surface as the agent handling an edge case badly. Each one you find is worth adding to a growing list rather than patching individually.
Missing escalations are the serious category, because they mean the system answered something it should have handed over. These are the ones to test hardest for.
Invention happens when a system is asked something beyond its data and produces a plausible answer anyway. The fix is constraint: explicitly instruct it to say it does not know and offer a human, and test that it does.
Our guides on auditing after 90 days and the real cost in year two cover the review disciplines that catch these before customers do.
Preventing repeats
Three habits prevent the large majority of agent errors.
Read twenty transcripts a month. Not a dashboard — actual conversations. You will find problems customers did not bother reporting.
Run a quarterly business-change review. What changed in pricing, services, policies, staffing or availability, and does the system know?
Keep a running error log. What happened, which of the four causes, what you changed. After six months this document is the most valuable thing you own about your automation, and it is what lets a new staff member take it over.
Frequently Asked Questions
What should we do first when an AI agent gives wrong information?
Honour what the customer was told where reasonable, apologise without blaming the software, then immediately search your transcripts to find out who else received the same wrong answer. The second and third steps matter more than the first.
Are we legally responsible for what our AI agent says?
Australian Consumer Law prohibits misleading or deceptive conduct in trade or commerce and does not carve out statements generated by software you deployed. Treat anything your system says as a statement by your business, and take advice on your specific situation.
Should we tell the customer it was an AI error?
Acknowledge the error as your business's error. Saying "our system gave you incorrect information and that's on us" is honest without sounding like an excuse — blaming the software invites the customer to distrust everything else it told them.
Should we honour an incorrect price?
Usually yes, where the amount is reasonable. The margin lost is generally smaller than the cost of a dispute and a negative review, and it is also the response most consistent with your obligations around representations you made.
What causes most AI agent mistakes?
Stale data — the business changed and nobody updated the system. The others are missing rules, escalation triggers that did not fire, and the system inventing an answer when asked something beyond its knowledge.
How do we stop it inventing answers?
Constrain it explicitly to known data and instruct it to say it does not know and offer a human when a question falls outside that. Then test the constraint by asking obscure questions, because plausible invention is the hardest failure to spot.
How do we prevent the same error recurring?
Fix the cause rather than the individual conversation, test the fix with several different phrasings, and keep a running error log recording what happened and what you changed. Monthly transcript review catches most issues before customers report them.
If your agent has never been stress-tested, that's worth doing before a customer does it for you. Book a CRM transition call, or see how we work at Pivot 2 Thrive.
Related Articles
Services
Related guides
- AI and Australian Consumer Law
- How to Audit Your AI Automation After 90 Days
- The Real Cost of AI Automation in Year Two
- AI Disclosure: What to Tell Your Customers
More
