Straight answer: Claude AI can draft replies to routine customer enquiries, sort your inbox by urgency, write your FAQ page for you, and hold a customer’s hand through a return or booking change, but it cannot make refund decisions or promises it hasn’t been told to make, and the businesses that treat it as fully hands-off end up with angry customers and made-up policies in writing. Set it up and it will save a small business owner three to eight hours a week on email alone.
What Claude is decent at in a customer service inbox
I’ve set Claude up for three small business clients in the past year, and the pattern is always the same. It’s brilliant at the repetitive 70 percent of your enquiries and useless, sometimes dangerously so, at the other 30 percent unless you build a fence around it first.
The 70 percent is stuff like: where’s my order, what are your opening hours, do you deliver to Scotland, can I change my appointment, what’s your returns policy. These are questions with a fixed correct answer that lives somewhere in your business already, in a policy document, a spreadsheet, or your own head. Claude can hold that information in what Anthropic calls a Project (a workspace where you upload your policies, price list, tone guide and past good replies) and then draft answers in your voice, referencing your actual rules, not generic customer service waffle.
The 30 percent is anything involving judgement: a customer who’s furious, a borderline refund request, a complaint about staff behaviour, anything legal-shaped. That’s where Claude should draft a first pass for a human to check, never send unsupervised.
A real example from a client’s inbox
I worked with a small skincare business last year, one founder, one part-time assistant, doing around 40 to 60 email enquiries a day at busy periods. Before we touched anything, the founder was spending roughly 90 minutes each morning just clearing the overnight backlog, most of it “is this suitable for sensitive skin” and “when will my order arrive” questions she’d answered a thousand times.
We built a Claude Project with her ingredient FAQ, her delivery timescales for UK and international orders, her returns window (28 days, unopened only), and about twenty examples of her actual past replies so the tone matched hers, not a generic customer service tone. She pasted incoming emails in, Claude drafted the reply, she read it, tweaked a line or two, sent it. Within three weeks her morning clear-out was down to around 25 minutes for the same volume. That’s not a magic number I’m inventing for effect, it’s what she told me on a call, and it tracks with what I’ve seen elsewhere: drafting is the slow part of email, not deciding what to say.
What did not change: the two or three tricky emails a week, a customer with an allergic reaction, a lost parcel dispute with the courier, someone asking for a refund outside the policy window. Those still went to her, and rightly so. Claude drafted a starting point for those too, but she rewrote most of it.
Step by step: setting up Claude for your own enquiries
You don’t need a developer or an agency for the basic version of this. Here’s how I’d do it for a solo trader or a five-person team:
- Step 1. Get a Claude Pro account (roughly £16 to £20 a month at the time of writing) or a Team plan if more than one person needs access.
- Step 2. Create a Project and name it something obvious, “Customer Emails” is fine.
- Step 3. Upload your actual policies: returns, delivery, opening hours, pricing, anything customers ask about weekly. Word documents and PDFs work.
- Step 4. Add 10 to 20 examples of your best past replies, the ones you were proud of, so Claude learns your tone rather than a bland default one.
- Step 5. Write custom instructions telling it what it must never do: never promise a refund, never invent a delivery date, never apologise on your behalf for something it hasn’t checked, always flag anything with the word “complaint,” “solicitor,” or “review” in it for a human first.
- Step 6. Run it in parallel with your normal process for two weeks. Read every draft before sending. Note what it gets wrong.
- Step 7. After two weeks, you’ll know which categories of question you can send with barely a glance and which ones still need real editing.
That’s it, no integration, no code. If you want it wired directly into your inbox or helpdesk so drafts appear automatically rather than through copy and paste, that’s where it starts to look more like the setups I write about in how AI customer service bots save small businesses time, which is a step up in complexity and usually worth outside help.
Where this quietly goes wrong (the bit worth being honest about)
Here’s the uncomfortable part that most guides skate past because it doesn’t sound as exciting as “AI handles your customer service.” Claude does not know your business. It knows what you told it, and if you told it nothing, or told it something out of date, it will still answer confidently, in full sentences, with total conviction, and it will sometimes be wrong. A returns window that changed last month, a product that’s been discontinued, a delivery promise you no longer make because your courier put their prices up. If your knowledge base is stale, Claude will cheerfully repeat the stale version to a customer as if it were current policy, and that customer will hold you to it.
I’ve seen this happen with a client who updated her Etsy shop policy but forgot to update the document sitting in her Claude Project. A customer got promised a 60-day return window that had shrunk to 30 days weeks earlier. She honoured it because arguing with a customer over an AI’s mistake is a worse look than eating the cost once. That’s the trade-off nobody tells you about upfront: the tool is only as current as the documents you feed it, and small business owners are busy people who forget to update documents.
The fix isn’t complicated, it’s just unglamorous: put a recurring 15-minute slot in your calendar once a month to check the documents in your Claude Project are still accurate. That single habit prevents most of the embarrassing mistakes I’ve seen.
Claude versus a fully automated chatbot
People often ask me whether they should just get a chatbot that answers customers on the website without a human touching it. That’s a different tool for a different job. Claude, used the way I’ve described, keeps a human in the loop reading every reply before it goes out, which is slower but safer for a small business where one bad public reply can end up as a screenshot on social media. A fully autonomous setup, the kind covered in how AI agents handle customer service without human input, removes that human check entirely, which only makes sense once you’ve got months of data showing exactly which questions are safe to answer without a person looking first. Most small businesses I meet aren’t there yet, and honestly, don’t need to be. The time saved going from “no AI help” to “Claude drafts, I approve” is already the biggest jump. The jump from “I approve” to “nobody approves” saves you another hour a week and adds a real risk of an unsupervised reply going out wrong. That trade isn’t worth it for most small teams.
What it costs and when to get help setting it up
The AI tool itself is cheap: £16 to £20 a month for one person, more if several staff need seats. The real cost is time: building the Project, writing good custom instructions, and training whoever handles email to check drafts rather than just hit send. Done badly, that setup takes an afternoon of frustration. Done with someone who’s done it before, it takes an hour or two.
If you’d rather not learn this by trial and error the way my skincare client’s first week went (she sent one draft that quoted a delivery time from a document she’d forgotten to delete), working with an AI consultant for your small business to set the whole thing up in one sitting is usually the cheaper route once you count your own hours. It’s a similar scoping exercise to what I cover in what an AI audit covers for a small business and what it should cost, where the first job is always finding out what’s eating your time before you buy any tool at all.
None of this is theoretical or moving fast in some abstract sense either. I cover new developments most weeks, including a run of customer service specific updates in the 9 August 2026 AI news roundup, and the pace hasn’t slowed down since.
The questions to ask before you switch anything on
- Do I have my policies written down anywhere, or only in my head? If it’s only in your head, write it down first, that’s the real bottleneck, not the AI tool.
- Who is going to read every draft for the first two to four weeks? Someone has to, don’t skip this.
- What’s my rule for anything that mentions a complaint, a refund outside policy, or a legal word? Decide this before you start, not after the first tricky email arrives.
- Who owns updating the knowledge base when a policy changes? Put one name against this, not “the team.”
Get those four things sorted and Claude will take the boring 70 percent of your inbox off your plate within a couple of weeks. Skip them and you’ll get exactly the kind of embarrassing, avoidable mistake that ends up as a bad review instead of a saved afternoon.
Frequently asked questions
Can Claude AI answer customer emails without a human checking them first?
It can, technically, but I wouldn’t recommend it for most small businesses until you’ve run it with a human checking drafts for at least a month and know which categories of question it reliably gets right.
How much does it cost to set up Claude for customer service in a small business?
The tool itself runs £16 to £20 a month per user on the Pro plan, with Team plans costing more per seat for multiple staff; the bigger cost is the hour or two it takes to build a proper knowledge base of your policies and past replies.
What’s the biggest mistake small businesses make when using Claude for customer service?
Uploading policy documents once and never updating them, so the AI keeps confidently repeating outdated returns windows, prices or delivery times to customers who then hold the business to them.
Is Claude better than a chatbot for small business customer service?
They solve different problems: Claude used with a human checking drafts is safer and slower, while a fully automated chatbot is faster but removes the human safety net, which only makes sense once you have solid data on which questions are safe to answer unsupervised.
