AI in customer service works well on a narrower slice than vendors suggest, and the projects that fail nearly always fail the same way: automating the conversations where being wrong is expensive.
Start with the split
Assisting agents — drafting replies, summarising history, suggesting the next step, with a human sending. Fast payback, low risk, almost always the right first move.
Automating conversations — the bot talks to the customer directly. Higher ceiling, much higher risk, and the thing everyone starts with by mistake.
If you do nothing else from this page: do the first one first. It delivers most of the value with almost none of the exposure, and it teaches you which conversations are actually repetitive before you automate any of them.
What automates safely
Triage and routing. Read the message, categorise it, send it to the right queue with a suggested reply. Reliable, and a human still sends.
Documented questions. Opening hours, shipping policy, password resets. If the answer is in your help centre, a bot can retrieve it.
Information gathering. Order number, account email, description of the problem — collected before a human picks it up. Saves a round trip every time.
Out-of-hours acknowledgement. Better than silence, provided it is honest about being a bot.
Status lookups, where the bot reads a real system rather than guessing.
What does not
Billing, refunds and account status. The exact conversations where a confident wrong answer costs money and generates a complaint. The model cannot know it is wrong.
Anything about legal terms. Cancellation rights, liability, contractual obligations.
Complaints and anything emotional. A customer who is already angry and gets a bot becomes a customer who posts about it.
Edge cases. Which is, by definition, what reaches support after your FAQ has absorbed the easy ones.
The design decision that matters most
Be obvious that it is a bot, and keep the route to a human visible at all times.
Customers tolerate bots. They do not tolerate being trapped by one, or realising after five minutes that they have been explaining a problem to software. Almost every bot people complain about publicly failed on one of those two points.
The escape hatch should be visible from the first message, not offered after three failed attempts.
Grounding, or your bot invents your policy
A general model knows nothing about your business. Ask it your refund window and it will produce a plausible one.
Grounding means giving it your actual help centre and policies, restricting it to answering from them, and instructing it to escalate when they do not cover the question. This is not optional — it is the difference between a support bot and a liability.
Practical consequence: most of the work is preparing your documentation, not configuring the bot. Teams consistently underestimate this. If your help centre is thin or out of date, fix that first — you will get most of the deflection from the help centre alone.
Measuring it honestly
The metric vendors quote is deflection rate: how many conversations the bot closed. It is easy to game and easy to misread — a customer who gives up is counted the same as a customer who was helped.
Better measures:
- Resolution rate: how many were actually resolved, checked by sampling
- Escalation rate, and whether escalated customers were more annoyed
- Repeat contact within 48 hours — the clearest signal a "resolution" was not one
- CSAT split by bot-handled vs human-handled, which is the number that tells you whether this is working
If repeat contacts rise while deflection rises, the bot is not deflecting. It is delaying.
Costs
- Agent assistance: per-seat subscription. Cheapest, fastest payback.
- Customer-facing: usually per conversation or per resolution. Model your volume before signing — this is where costs surprise people.
- Custom builds: engineering plus maintenance as models change. The maintenance is what gets underestimated.
Common questions
Can AI replace customer service agents? No. It removes repetitive triage so agents handle cases needing judgement. Treating it as headcount reduction is how these projects fail.
What should AI never handle in support? Billing, refunds, account status, legal terms, and complaints — the places a confident wrong answer costs real money.
Do customers mind AI support? Not if you are upfront and a human is one click away. They mind being trapped or deceived.
How do I stop it giving wrong answers? Ground it in your own documentation, restrict it to answering from that, and make it escalate rather than guess.
What is a realistic deflection rate? Depends entirely on how repetitive your volume is. Be sceptical of vendor figures and measure repeat contact alongside deflection.
How long does it take to set up? Agent assistance, days. A grounded customer-facing bot, weeks — most of it spent on your documentation.
Where should I start? Agent assistance. It delivers most of the value, carries almost no risk, and shows you which conversations are worth automating.