Claude for business: a customer reply desk that knows your prices

A chat bubble beside a laptop, standing for customer messages answered with AI

It is 10pm and there are eleven unanswered WhatsApp messages. Four ask the same thing (do you deliver to Punggol?), two want a price you have quoted a hundred times, and one is a complaint you need to think about properly. You open a chatbot, paste in the first message, and get back a cheerful, fluent reply that quotes a delivery fee you have never charged.

That is the usual first experience of using AI for customer replies, and it is why a lot of small business owners give up after a week. The chatbot writes well. It just doesn’t know your business, and it fills the gap by guessing.

The fix is not a cleverer prompt. It is giving the tool your actual facts once, so you stop retyping them and it stops inventing them. In Claude, the feature for that is called a Project.

Why this is worth an evening

You would not be early. IMDA’s Singapore Digital Economy Report found AI adoption among Singapore SMEs tripled from 4.2% in 2023 to 14.5% in 2024, driven mainly by micro and small firms picking up off-the-shelf generative AI tools such as ChatGPT, Gemini and Copilot. In IMDA’s pulse survey of firms already using AI, customer service was one of the top three business functions they used it for.

So the question is less “should I” and more “how do I stop it making things up about my prices”. Here is the setup.

Step 1: gather the boring documents

Before you open anything, put together the material you already answer from. For most small businesses that is three things:

  • A price list, with what is and isn’t included.
  • An FAQ, written the way you would actually answer. Delivery areas, lead times, payment methods, what happens if an order is late.
  • Two or three past replies you were happy with, with the customer’s name and number removed. These teach tone better than any description of it.

Plain text or a simple document is fine. If your FAQ only lives in your head, typing it out is the most valuable part of this whole exercise, AI or not.

[[photo:1]]

Step 2: create the Project

  1. Go to claude.ai/projects (or click Projects in the left sidebar).
  2. Click + New Project in the top right corner.
  3. Give it a name and description, such as “Customer replies”. Anthropic’s help pages note that Claude does not see the name or description, so they are only for you.
  4. On the project page, find the knowledge panel on the right and click the + button. Upload your price list, FAQ and sample replies.
  5. Click Set project instructions, paste in something like the text below, and click Save instructions.

You draft replies to customer messages for my business. Use only the facts in the uploaded price list and FAQ. If a customer asks something those files do not answer, do not guess: write [CHECK: what you need to know] in the reply so I can fill it in. Keep replies short, friendly and in Singapore English. Match the tone of the sample replies. Never promise a discount, refund or delivery date that is not in the files.

The [CHECK] line is the important one. A blank you can see is far cheaper than a confident wrong fee you can’t.

Projects are available on Claude’s free plan, which allows up to five projects, so you can try this without paying. On paid plans, Anthropic says Claude switches to a retrieval mode when project knowledge gets close to its limit, which lets a project hold a much larger set of files. (Why there is a limit at all is covered in our explainer on what a context window is.)

Step 3: test it on messages you have already answered

Do not start with tonight’s eleven messages. Take five old enquiries where you know the right answer, remove names and phone numbers, and paste each one into a new chat inside the project. For each draft, check:

  • Is every price and date in it actually in your files?
  • Did it use [CHECK] where it should have, or did it guess anyway?
  • Would you be comfortable sending it with only small edits?

When a draft is wrong, the tempting move is to correct it in the chat. Do that, but also fix the source. Anthropic’s help page is explicit that context is not shared across chats within a project unless the information is added to the project knowledge. A correction typed into Tuesday’s chat is gone by Wednesday. A correction added to the FAQ file sticks.

Do this for a couple of weeks and your FAQ quietly gets better too.

Step 4: check one privacy setting before real customers

On Claude’s consumer plans (Free, Pro and Max), Anthropic may use your chats to train future models if the Model Improvement setting in your Privacy Settings is on. You can switch it off at any time, though Anthropic notes that turning it off stops future use rather than removing data already used. Business plans such as Team and Enterprise are covered by commercial terms instead.

Whatever the setting, the safer habit is to paste the question, not the customer. Strip out names, phone numbers, addresses and order details unless the reply truly needs them. Singapore’s Personal Data Protection Commission published advisory guidelines on personal data in generative AI on 20 July 2026, and a summary by the law firm Allen & Gledhill notes that organisations deploying these tools carry primary responsibility for safeguarding the personal data they hold. It is guidance, not a new law, and this is not legal advice. But “don’t paste what you don’t need to” costs nothing and covers most of it.

What it will and won’t do

This setup drafts. You still send. It handles the repetitive messages (prices, delivery, opening hours) and gives you back the time for the complaint that deserves a human answer. It is not a reason to stop reading what goes out under your name.

If you only do one thing tonight, do Step 1. Type out the FAQ. It makes the AI useful, and if you later decide AI isn’t for you, you will still have the document. And if spreadsheets are the other thing eating your evenings, our guide to using Claude for Excel on an inherited spreadsheet uses the same principle: give it your real material, then check its work.

Was this useful?

Sources

  1. Singapore Digital Economy Report 2024/2025 · Infocomm Media Development Authority (IMDA)
  2. What are projects? · Anthropic (Claude Help Center)
  3. How can I create and manage projects? · Anthropic (Claude Help Center)
  4. Is my data used for model training? · Anthropic (Claude Privacy Center)
  5. Updates to Consumer Terms and Privacy Policy · Anthropic
  6. PDPC issues Advisory Guidelines on Use of Personal Data in Generative AI · Allen & Gledhill