When a customer believes something about their account that the data contradicts (they think they were charged twice when they were charged once, they think they cancelled when they did not), an AI agent's job is to surface the evidence plainly, offer the next step, and escalate when the gap persists. Arguing is not the AI's job; clear evidence and a graceful path forward is.
A customer messages: "you charged me twice last month." The AI looks at the billing record. The customer was charged once, plus a refund went back to the wrong account three months earlier. The customer is reading the bank statement and seeing two debits in a row from a different concern. The AI has to address the actual situation without making the customer feel small.
What people in the field are saying
CX Decoded's "The AI forgiveness gap" argues that customers extend less forgiveness to AI than to humans when the AI delivers an unwelcome answer, even when the answer is correct. The AI has to be more careful with the framing, not less.
What should the AI do first?
Surface what the data shows, in plain language, with the specifics. "Looking at your account, you were charged once on the 14th, and the second debit you may be seeing is from a previous month. Would you like me to pull up both records?" Not "you are wrong." Not "actually." Just the facts and the offer to walk through them together.
When should the AI not push back?
When the gap involves a customer's emotional reading of a situation, not a factual one. A customer who says "this is the third time I've contacted you about this" may be technically wrong about the count and right about the feeling. Correcting the count adds nothing. Acknowledging the experience and offering the next step is the right move.
When should the AI escalate?
When the customer maintains their belief after the AI has shown the evidence twice. At that point, the AI is not the right tool: it has done what it can, and a person needs to take over. The escalation should carry the audit trail (what the AI showed, what the customer said) so the human starts informed.
Where does this go wrong?
The AI says "no, that's not correct" too plainly. The customer reads it as the company calling them a liar. Or the AI capitulates and apologises for something that did not happen, which sets a bad precedent and leaves the customer thinking the company agrees with their version. Neither serves the customer; both are easier than the right move.
What is the practical pattern?
Train the AI on the evidence-and-offer pattern. Make escalation easy on the second time the gap appears in the same conversation. Read a sample of these cases by hand monthly to make sure the AI is being firm enough to be useful and gentle enough not to inflame. The balance is real and it is part of the brand voice.
Related: how AI handles an emotional contact, empathy at scale, and the glossary explainer on escalation.