Customer support is where most companies start with AI, and where the experience degrades fastest when the objective is set wrong.
The deflection metric problem
The metric almost always chased is the percentage of queries resolved without human involvement. It is easy to measure, rises quickly and presents well at the committee.
It is also the most reliable way to build something your customers detest. A system optimised not to escalate learns not to escalate, which is not the same as resolving. The result is the experience we have all suffered: three rounds with a bot before getting to a person, and arriving at that person angrier than you started.
The metric we recommend is time to correct resolution, counting the full path including escalation. A system that resolves 40% in thirty seconds and hands the other 60% to a person with the context already gathered beats one that “resolves” 75% with half of it resolved wrongly.
What works well today
Model capability is jagged, and in support that shows up sharply:
Works very well:
- Classifying and routing to the right team from the first message.
- Gathering context before a person steps in — order number, version, steps already tried.
- Answering questions that have a documented, stable answer.
- Drafting the reply for a human agent to review and send.
- Summarising long threads for whoever picks the case up.
Works badly:
- Anything requiring the live state of several systems plus reasoning about exceptions.
- Cases where the customer is angry. Added friction multiplies the damage.
- Situations with financial consequence: refunds, compensation, cancellations.
- Anything the system cannot resolve but can appear to resolve.
That last point is the dangerous one. A model that confidently states something incorrect about a returns policy creates a bigger problem than not answering at all.
The design we recommend
Assist before replace. The first phase should put AI on the agent’s side, not the customer’s. Draft replies, summarise history, suggest articles. The agent reviews and sends. Low risk, immediate gain, and it generates the data that will tell you what can be fully automated later.
Frictionless escalation. The option to reach a person must be visible from the first message. Hiding it does not cut cost: it moves the cost to your reputation.
Context that travels. When it escalates, the human receives the whole conversation. Making the customer repeat what they already explained is the most expensive and most avoidable failure of all.
And the legal obligation
Since 2 August 2026, Article 50 of the EU AI Act requires telling people they are interacting with an AI. If your chatbot is called “Laura” and signs off as though it were on the support team, you have a compliance problem on top of a trust problem.
And if the system was deployed before that date, the deadline to bring it in line with Article 50(2) is 2 December 2026.