AI agents in customer service: what works and what's theatre

Everybody wants an AI agent in customer service and almost nobody is clear on what they're asking for. The demo always goes well: someone types "what are your hours?" and the bot answers. The trouble starts in production, when a customer writes "you charged me twice for March and I want my money back" and the bot, with tremendous confidence, invents a refund policy that doesn't exist.

That's the difference between a chatbot and an agent worth having. It isn't the model — it's everything around it.

What doesn't work

A generic model bolted onto your chat channel. With no access to your data, a model can do exactly two things: generalise or invent. And because it's trained to sound helpful, it will invent before it says "I don't know." That's the most expensive failure mode in support, because the customer believes it.

Uploading the FAQ PDF and calling it done. Better than nothing, but it breaks on the first question that isn't phrased the way the PDF phrases it. Real customer questions are never phrased the way the PDF phrases it.

A bot that can't surrender. Without a clean exit to a human, the customer will try three times, get angry, and you'll have converted a simple question into a complaint. A bot that can't say "a person needs to handle this, transferring you now" destroys more value than it creates.

What does work

1. Ground it in your information, not the internet

The agent has to read from your real sources: your knowledge base, your policies, the customer's history, the status of their order. The technique is retrieval (RAG): before answering, the system searches your documents and data, and the model writes only from what it found.

The practical consequence is the one that matters: if it found nothing, it doesn't answer. It says it doesn't have that information and hands off. An agent that can say "I don't know" is an agent you can put in production.

2. Give it explicit limits

Before a line of code gets written, three lists have to exist — and this conversation matters more than the choice of model:

• What it can answer alone. Hours, order status, how to complete a process, what paperwork is needed.
• What it can answer using customer data. Their balance, recent orders, next delivery — always after verifying who they are.
• What it never answers. Money disputes, cancellations, legal questions, exceptions to policy, anything that amounts to a promise. That goes to a human, every time, without trying to solve it first.

3. Make it sound like your company

An agent that reads like a manual translated from English announces itself in the first message and drags down trust in everything after it. Tone, formality and industry vocabulary get defined and tested against real transcripts of your past conversations — not guessed at.

4. Integrate it, don't park it alongside

If the agent can't look up the real status of an order, all it can do is generalise. The value shows up when it's connected to the system where the data lives: the ERP, the storefront, the customer database. That's when it stops being an FAQ search box and starts resolving things.

The reverse matters too: everything it handles has to be recorded where your team already works. If the bot's conversations live in a separate console nobody opens, you've lost half the benefit.

5. Measure it

The metrics that matter aren't conversation volume or the "hours saved" the vendor estimates. They're three:

• Resolution rate without a human, measured on cases that actually closed — not on "the customer stopped writing."
• Handoff rate, which should be high at first. A bot handing off 40% in month one is working correctly; one handing off 2% is probably making things up.
• Cases where it got it wrong, reviewed by hand every week. It's tedious, and it's the only thing that makes the agent improve.

How long it takes and what it returns

A tightly scoped agent — one channel, a clear set of question types, integrated with one system — takes four to eight weeks, including the slow part, which is gathering and organising the material it will answer from.

The typical return isn't "we replaced the support team." It's that 50–70% of repetitive questions stop reaching a person, and the team moves to the cases that need judgement. That usually improves human support quality too: nobody is answering the opening-hours question for the tenth time that day.

Where to start

Not with the model. With the log of your last five hundred conversations.

Group them, count how often each type appears, and you'll find that 60% are five or six questions. That's the scope of version one. Everything else — the interesting cases, the strange ones, the ones that look impressive in a demo — is version three, and only if version one worked.

If you'd like us to look at your real conversation log and tell you honestly which share is automatable, get in touch. It's a half-hour conversation and it usually saves months.