A customer asks "can I use my 20% discount code on sale items?" Your AI responds "Yes, discount codes can be applied to any item in your store." The customer places a $300 order, applies the code, and gets a confirmation email showing the discount.

Problem: your policy says discount codes don't apply to sale items. The AI was wrong. Now you have a customer who made a purchase decision based on incorrect information from your official support channel.

What do you do?

This scenario plays out thousands of times a day across companies using AI support. The AI gives a confidently wrong answer, the customer acts on it, and the company has to decide between honoring the wrong answer (losing money) or correcting it (losing trust).

Types of AI Errors in Support

AI makes mistakes in customer support that fall into four categories, each needing a different response.

Misclassification. The AI thinks the customer is asking about billing when they're actually asking about a feature. The customer gets a billing-related response to a feature question. This is confusing but usually not harmful. The customer just asks again or escalates.

Misclassification is detectable. If the customer's follow-up message doesn't match the expected pattern for the classified intent, the classification was probably wrong. AI systems can detect this and auto-escalate.

Wrong information. The AI gives factually incorrect information about your product, policies, or processes. This is the dangerous one. Customers make decisions based on what your support tells them. Wrong information leads to wrong decisions, which leads to disputes.

Common sources of wrong information: outdated policy data, AI extrapolating from incomplete knowledge, LLM hallucination (generating plausible-sounding information that's entirely fabricated), and edge cases that don't match any training pattern.

Wrong action. If your AI can take actions (process refunds, update accounts, create tickets), it might take the wrong one. Refunding the wrong amount, updating the wrong field, creating a duplicate ticket. This is the highest-risk error type because it changes real data.

Wrong tone. The AI responds with cheerful boilerplate to a customer who's angry or distressed. The content might be technically correct, but the mismatch between the customer's emotional state and the AI's tone makes them feel dismissed. This is especially common with template-based systems that prepend "Great question!" to every response.

Detection: How to Catch Errors

You can't fix what you don't see. Build detection into your AI support system.

Confidence monitoring. Every classification or response has a confidence score. Set a threshold (usually 85 to 90%). Below that threshold, don't respond automatically. Route to a human. This catches the cases where the AI isn't sure.

Customer feedback signals. If a customer says "that's not what I asked" or "that doesn't answer my question" or rates a response negatively, the AI probably got it wrong. These signals should trigger automatic escalation and flag the interaction for review.

Post-resolution audits. Randomly sample 5 to 10% of AI-resolved tickets weekly. A human reviewer checks: was the classification correct? Was the response accurate? Was the customer satisfied? Track error rates over time.

Repeat contact detection. If a customer contacts you about the same issue within 24 hours of an AI resolution, the AI probably didn't actually resolve it. Flag these for review.

Recovery: What to Do When AI Is Wrong

The recovery playbook depends on the impact of the error.

Low impact (misclassification, wrong FAQ answer, no action taken). Correct the information. Apologize briefly. Move on. "I'm sorry about the confusion. Here's the correct information about [topic]." Don't over-apologize. Don't make it a bigger deal than it is.

Medium impact (wrong policy information, incorrect quote, misleading guidance). Correct the information. Acknowledge the mistake explicitly. If the customer made a decision based on the wrong info, honor it. "Our AI gave you incorrect information about the discount code. That's our mistake. We're honoring the 20% discount on your order, and we've corrected the response so this doesn't happen to anyone else."

Honoring the AI's wrong answer costs money. But the alternative (saying "the AI was wrong, so your discount is void") destroys trust. The customer will never trust your support again, human or AI.

High impact (wrong action taken, financial error, data corruption). Fix the error immediately. Escalate to a senior agent. Contact the customer proactively (don't wait for them to discover it). Explain what happened, what you've done to fix it, and what you're doing to prevent it in the future.

Repeated errors (same mistake happening to multiple customers). This signals a systemic problem, not a one-off. Pull the AI off that intent category until you fix the root cause. One customer getting a wrong answer is a bug. Ten customers getting the same wrong answer is negligence.

"We discovered that our system processed an incorrect refund on your account. The refund was $50 more than it should have been. We've corrected the amount. If this causes any issues with your bank, please contact us and we'll sort it out immediately."

Prevention: Reducing AI Errors

The best error handling is fewer errors in the first place.

Use classifiers instead of generators. A classification model that routes to pre-verified responses can't hallucinate. It might misclassify (send the wrong pre-verified response), but it won't invent a policy that doesn't exist.

Supp's purpose-built classifier achieves 92% accuracy on 315 intents. The 8% that it gets wrong are usually close misses (classifying a "change shipping address" as "update account info"). The responses for similar intents are similar enough that the customer still gets useful help.

Compare that to LLM chatbots, which hallucinate at rates of 3 to 15% depending on the model and the complexity of the question. And hallucination errors are much harder to detect because the wrong answer sounds confident and coherent.

Narrow the AI's scope. Don't let AI handle everything. Define a clear list of intents that AI resolves automatically. Everything outside that list goes to a human. It's better to handle 50% of tickets perfectly with AI than 80% with a 10% error rate.

Version your responses. When you update a policy, update the AI's response library the same day. Stale responses are the most common source of "wrong information" errors. If your refund policy changed last Tuesday and the AI is still giving last Monday's answer, that's an organizational failure, not an AI failure.

Test with adversarial examples. Try to break your AI. Send it ambiguous messages, multi-topic messages, messages with typos, messages in unexpected formats. Find the failure modes before your customers do.

The Error Budget

Zero errors isn't achievable. A 92% accuracy rate means 8 out of 100 messages get the wrong classification. At 500 messages per day, that's 40 mistakes. Per day.

The question is: are those 40 mistakes caught and corrected quickly? And are they low-impact mistakes (wrong FAQ category) or high-impact ones (wrong refund amount)?

Set an error budget: the maximum acceptable error rate and the maximum acceptable impact. "We tolerate up to 10% misclassification as long as the customer gets a correct resolution within 2 additional messages." Or: "Zero tolerance for errors that result in financial impact to the customer."

Track against the budget weekly. If you're within budget, the AI is working. If you're over, either narrow the AI's scope, improve the model, or add more human checkpoints.

Perfection isn't the goal. Being better than the alternative (human-only support, which has its own error rates of 5 to 15%) is the goal.