AI customer support is easy to switch on and easy to get wrong. Two well-known stories show both halves of the problem.
In February 2024, Klarna said its new AI assistant handled 2.3 million chats in its first month, the work of about 700 full-time agents. By May 2025, its CEO was telling Bloomberg that the push to cut costs had lowered quality, and the company began hiring human agents again. The AI kept handling a large share of chats. What changed was the assumption that it could handle all of them.
The second story is about liability. In February 2024, a Canadian tribunal ordered Air Canada to pay a customer after its website chatbot gave him wrong information about bereavement fares. The airline argued the chatbot was responsible for its own words. The tribunal disagreed. Whatever your bot tells a customer, you are on the hook for it.
So the first rule is to limit what the AI is allowed to say. Point it only at your own help articles and policies, and keep those current. An AI that answers from an outdated refund page will confidently repeat the outdated policy.
The second rule is to make escalation easy. A customer should be able to reach a person without typing "agent" five times. Hand over the conversation history so they do not have to explain twice. The frustrating part of bad bots is rarely the bot itself. It is being trapped inside one.
The third rule is to keep people on the sensitive cases. Refunds above a set amount, cancellations, complaints and anything involving grief, health or money should go to a human by default.
Pricing has moved toward paying for results. Intercom's Fin, for example, charges $0.99 for each conversation it resolves, on top of seat fees. That makes it easy to compare against the cost of a human-handled ticket, but read the fine print on what counts as resolved.
Finally, read the transcripts every week. The AI's mistakes are patterns, and the fix is usually a missing or unclear help article.
The short version: let AI answer the repetitive questions from your own documentation, make the path to a person short and obvious, and route anything emotional or expensive to people by default. Treat every answer the bot gives as one your company gave.
