AI customer service: what it costs and what it actually replaces.

Customer service is the highest-volume repetitive work in most businesses, which makes it the obvious first automation — and the easiest to do badly, because a bad support bot is worse than no bot at all.

What it should and should not handle

Should: hours, location, pricing, policy, availability, order or appointment status, the forty questions your team answers every week from an answer that already exists in writing.

Should not: anything where the customer is upset, anything involving money moving, anything where the right answer depends on judgement rather than policy. Those should route to a person immediately with the conversation attached.

The quality of a support assistant is decided almost entirely by how well it hands off, not by how much it answers.

Setup, honestly

The build is not the hard part. Gathering the content is. Most businesses discover during week one that their policies live in four places, two of them contradict, and one is three years out of date. Cleaning that up is genuinely useful work you should have done anyway — and it is why the first week is scope rather than code.

What to measure

• Deflection rate — the share of conversations resolved without a human. Expect 40–70% on well-chosen question sets.
• Escalation quality — when it hands off, does the human have full context, or does the customer repeat themselves?
• After-hours capture — enquiries answered outside working hours that previously went unanswered entirely. This is usually the biggest surprise.
• Accuracy against your own cases, checked monthly against the evaluation set, not assumed.