AI customer service for small business: a worked example

Most small businesses don’t need an AI chatbot. They need AI behind the inbox: sorting messages, finding the right help article and drafting replies that a person approves. That’s cheaper to run, safer when it’s wrong and far easier to measure than a bot talking to customers on its own.

The bar isn’t a perfect bot. It’s answering better than your team does today, and proving it with a number. If nobody can say what share of customers the inbox actually helps, start there.

“Your accuracy is never going to be 100%, because that’s the wrong target.”

What follows is a worked example. The company is hypothetical, but the plan is the one we’d put in front of a real operations lead.

What does AI customer service look like at a 60-person business?

Imagine a 60-person UK company that rents commercial coffee machines to cafés on a monthly subscription. Support is four people sharing one inbox and a phone line. Customers write about late deliveries, billing, machine faults and cancellations.

The workflow before

• Every email lands in one shared inbox, unsorted.
• Whoever is free picks the next message, usually the newest rather than the most urgent.
• The agent opens the order system, the billing system and the courier’s tracking page in separate tabs.
• They search an internal wiki that’s half out of date, or ask a colleague on Slack.
• They write the reply from scratch, or paste a saved snippet and fix the details by hand.
• A complaint that should reach a manager gets spotted late, if at all.

Nothing here is broken. It’s just slow, and the slow parts are reading, looking things up and typing.

The workflow after

• A model tags each new message with a topic and an urgency level. The system looks up the customer from the sender’s address.
• The model also tags likely refunds, complaints, legal threats and safety issues. Coded rules route those tags, plus any keyword matches, to a named person’s queue.
• The order, the latest invoice and the tracking status appear in one panel beside the message.
• The system searches your help articles and past resolved tickets, then shows the three closest matches.
• It drafts a reply and lists the sources it used underneath.
• An agent reads the draft, edits it if needed and presses send.

The agent still owns every reply. What changes is that they start from a sourced draft instead of a blank page and five open tabs.

What should be automated and what should stay human?

Draw the line by asking what it costs to be wrong. A wrong tag costs a few seconds. A wrong refund promise costs money and trust.

Notice the third row. If a question has one right answer sitting in a system, fetch it. Don’t ask a model to remember it.

Will AI damage the trust you’ve built with customers?

It’s the fear we hear most from owners. You’ve built trust and a customer base over years. You don’t want a bot to undo that in an afternoon.

The answer is to give the machine the cases it’s good at and keep the rest human. Picture a plane coming in to land in thick fog on autoland.

Autoland copes with fog because the conditions are tightly defined and tested, and the pilots still monitor it. The 2009 Hudson River ditching needed human judgement, because nothing in the checklist fitted it.

Support splits the same way:

• Common, well-documented questions have hundreds of past answers. Let AI draft those.
• Unusual, emotional or new situations have little history. Hand them to a person straight away.

That gives you a practical rule. If a message looks unlike anything in your resolved tickets, the system should flag it rather than guess.

How does knowledge-base search work?

Most of the value sits here, and most of the failures too. The technique is called retrieval: the system finds relevant passages first, and the model answers only from those.

You’d index three things: published help articles, internal process notes and a few hundred resolved tickets with good replies. Each is split into short passages and stored as embeddings. Embeddings are numerical fingerprints that let you search by meaning rather than exact words.

When a message arrives, the closest passages go to the model with a plain instruction. Answer from these, cite them, or say you don’t know.

The awkward truth is that your help content is usually the bottleneck. If the wiki says one thing and your agents do another, the model will confidently repeat the wiki. Fix the content before you blame the model. Our guide to getting your data ready for AI goes into this properly.

When is a chatbot the wrong answer?

A customer-facing chatbot answers people directly, with nobody checking first. That’s a very different risk from drafting replies for your staff. We’d steer you away from one if any of these are true:

• Most of your contacts need an account change, a refund or a judgement call, not information.
• Your help content is thin, contradictory or out of date.
• Your customers are often upset, vulnerable or in a hurry.
• You can’t staff a fast hand-off when the bot gets stuck.

A chatbot earns its place when lots of people ask the same factual questions out of hours. Think delivery windows, opening times or how to reset a password. Even then, put a clear “talk to a person” option on every screen.

In our coffee-machine example, most messages are about faults and billing. Those need the account and a decision. So we’d build the inbox assistant and leave the website chat alone for now.

What does the build look like?

There are four parts. None of them is exotic, and the integration work outweighs the AI work.

Data sources

• Your helpdesk or shared inbox, such as Zendesk, Freshdesk, Help Scout, Gmail or Outlook.
• Your order, billing and delivery systems, connected read-only.
• Help articles, internal process notes and a sample of well-handled past tickets.

Model

A hosted large language model does the reading, tagging and drafting. You call it over an API, which is a programmatic connection, rather than through a chat window.

Pick a provider whose business terms say your data isn’t used for training. Check which region it’s processed in. You don’t need to train your own model, and for an SME that’s rarely the right call. We compare the options in off-the-shelf AI vs custom.

Guardrails

• The model is told to answer only from retrieved sources and must cite them under the draft.
• A check flags drafts whose claims don’t appear in those sources.
• The model tags likely refunds, complaints and legal issues. Coded rules route those tags, plus any keyword matches, to a person.
• Personal details are stripped from anything kept for testing.
• No draft leaves the building until a person presses send.
• Every draft, edit and send is logged, so you can trace any reply back to its sources.

Human review

Agents see the draft, its sources and a flag when the system wasn’t sure. Once a week, someone samples the tickets where agents rewrote most of the draft. Those rewrites tell you exactly which help content or instructions need fixing.

Where are the quick wins?

Sometimes the best win isn’t in the inbox at all. On a customer call with an HR-tech client, we learned something nobody had flagged.

One of their customers was paying a full-time salary to translate Spanish reports for English-speaking leadership. Sales had said it was too complex.

Machine translation is now good enough for most internal reports, though specialist terms still need a check. We shipped an auto-translate button within 24 hours, behind a feature flag, then rolled it out more widely.

The lesson for support is to listen to the contacts themselves. A request that keeps coming back often points to one small feature that removes it.

How much effort does it take?

We won’t promise you a saving we haven’t measured. The shape of the effort is more predictable than the outcome, so here it is.

1. Discovery. Read a few hundred real tickets, map the categories and audit the help content. This tells you whether the idea is worth building at all.
2. Triage first. Tagging and routing ship first. They’re low-risk, and they show quickly whether the model understands your customers.
3. Search and drafts. Add retrieval and drafting for your two or three biggest categories only.
4. Roll out in steps. Switch drafts on for one agent, then the team. At higher volumes, go 1%, 5%, 25%, then all messages, checking the measures at each step.
5. Widen. Add more categories once drafts in the first ones are usually sent with light edits.

A feature flag makes the steps safe. It’s a switch in the code that turns a feature on for chosen users, and off again in seconds.

Most of the time goes on connecting systems and tidying content, not prompts. We’re a Databricks partner, and our founder was a founding engineer at AI Music, which Apple acquired. Even so, the hard part is usually your billing system’s API.

What are the risks?

• Confident wrong answers. Models fill gaps with plausible text. Retrieval and citations reduce this, and human review catches what’s left.
• Stale content. The system is only as current as your help articles. Someone has to own them, by name.
• Rubber-stamping. Agents can drift into approving drafts without reading them. Watch for approval times that shrink to a few seconds.
• Personal data. Tickets hold names, addresses and sometimes health or financial details. Health details are special category data, which need an Article 9 condition as well as a lawful basis. You also need a data processing agreement with the provider and a retention rule. Our guide to AI and GDPR in the UK covers the detail.
• Automated decisions. UK GDPR has specific rules for significant decisions made solely by automated means. They changed in February 2026, and the ICO’s updated guidance explains them. A person who genuinely reviews refunds and account changes keeps you clear of that question. Waving drafts through doesn’t count.

That’s general information, not legal advice.

How do you measure success?

Choose your measures before you build, and capture a baseline from the current inbox. That baseline is the bar the AI has to beat.

Three numbers matter most. What share of customers were actually helped? How many round trips, meaning back-and-forth messages, did it take? And how many conversations escalated to a person?

“If you don’t know the percentage, you were already in trouble.”

If agents are still rewriting most drafts after a month, stop. Look at the content and the categories before you touch the model.

Have we built anything like this?

The closest thing we’ve built is for Everyone’s Invited. It’s a charity that was receiving over 100 personal stories a day. Each one had to be read and anonymised by hand before it could be published.

We built a system that automatically removes names, dates and locations from each submission. Staff and volunteers still publish the stories, now much faster. It has handled over 50,000 submissions and saved around 12,500 hours of review.

That isn’t a helpdesk, but the pattern is the same. The machine does the reading and the first pass. People keep the decisions that carry risk.

When isn’t this worth doing?

• You get a handful of messages a day. Good saved replies and a tidy inbox will do more.
• Customers can’t find answers on your site. Fix the help pages first, then see what’s left.
• Most contacts are about a problem upstream, like a courier that keeps missing slots. AI helps you apologise faster, but it won’t fix the courier.
• The real goal is cutting the team rather than freeing them for harder work. That plan usually ends with a chatbot nobody trusts.

If you’d like a second opinion on your inbox, talk to us about AI. The first chat is free, and so is an architectural review of your current setup. We’ll tell you honestly whether triage, drafts, a chatbot or nothing at all is the right next step. You can also score the idea yourself first with our AI readiness scorecard.