Most small businesses already have a knowledge base. It sits in the owner's head, a shared drive nobody has tidied in years, and a pinned WhatsApp message about how to process a refund. An AI knowledge base takes those scattered documents, puts them in one place, and lets staff ask questions in plain English instead of interrupting whoever knows the answer.
This guide covers how to set one up with Claude Projects, what to put in it, where it falls short, and the UK data checks to do before anyone uploads anything.
What is an AI knowledge base for a small business?
The difference from a shared drive is the way people search it. A shared drive needs you to know which file holds the answer. A knowledge base built in Claude lets a new starter type "can a customer return a sale item after 14 days?" and get an answer drawn from your returns policy, with the relevant wording quoted back.
In Claude, the tool for this is called Projects. Anthropic's help centre describes Projects as self-contained workspaces with their own chat histories and knowledge bases, where you upload documents and set project-specific instructions.
How many UK businesses are using AI this way?
The Office for National Statistics release on AI in UK businesses, published on 20 July 2026, puts large language models at the top of the list for firms with 10 or more employees, at 18%.
Read those together and the picture is clear. Plenty of owners have opened a chatbot. Very few have given it their own material to work from, which is the step that turns an occasional novelty into something staff use every day.
What should you upload first?
A useful way to pick is to look at the last fortnight of interruptions. If three people asked how to issue a credit note, that procedure goes in. If nobody has asked about an old supplier agreement, it stays out.
A sensible first batch for a ten-person trade or retail business:
- Current price list, with the date it took effect written at the top.
- Returns, refunds and cancellation policy, in the exact wording customers see.
- Booking or job intake rules, such as minimum call-out charges and lead times.
- Two or three written procedures, for example how to log a warranty claim.
- A short "house style" note on tone for emails and phone calls.
If those procedures do not exist on paper yet, writing them is the real first job. The guide to documenting business processes with AI covers how to get them out of people's heads and into a format Claude can read.
What file formats and limits apply in Claude?
Anthropic's file upload guide lists the supported types and notes that XLSX needs code execution enabled. Project files are read as extracted text, except multimodal PDFs, so the layout of a glossy brochure mostly disappears and only the words remain.
On capacity, the Projects help page says Pro, Max, Team and Enterprise plans can switch on a retrieval mode that expands project knowledge by up to 10 times. Free plan users can create up to five projects but do not get that expansion.
How should you write the project instructions?
The instructions box is where you set the rules of the knowledge base. A short, firm set works better than a long one. Something like:
- Answer only from the project files. If the files do not cover it, say so and suggest who to ask.
- Name the document and section each answer comes from.
- Use UK English and £ for prices.
- If two documents disagree, flag the conflict rather than picking one.
The last rule is more useful than it looks. Conflicting documents are a common problem in a first knowledge base, and having Claude point them out is a cheap way to audit your own paperwork.
How do you keep the knowledge base accurate?
Knowledge bases fail slowly. Nobody notices the old price list is still in there until a customer is quoted last year's rate. A short monthly routine prevents this:
- Open the project file list and check every date.
- Ask Claude five questions you know the answers to and compare.
- Ask staff which questions got "not covered" replies and fill the gaps.
Keep a plain list of what is in the project and when each file was last updated. It takes a few minutes a month and is the difference between a tool people trust and one they quietly stop using.
Should it answer customers or only staff?
A customer-facing assistant is a different build. It needs a website widget, logging, a clear handover to a human and limits on what it is allowed to promise. The trade-offs are covered in more detail in whether an AI chatbot is worth it for a small business and the broader guide to AI customer service for UK small businesses. Running the internal version for a few weeks first gives you the evidence to decide.
What UK data protection checks matter?
UK GDPR does not stop you using AI tools, but it does mean you should know what personal data you are processing and why. Most knowledge bases need none. A returns policy contains no personal data. A spreadsheet of last year's complaints, with names and addresses, does.
If staff will share the project, the Projects help page notes that Team and Enterprise members can share by email or across the organisation with "can view" or "can edit" permissions, and that owners can switch sharing off. Decide who needs edit access, which is usually one or two people. The AI data protection checklist for UK businesses walks through the rest, including what to record before you start.
Is it worth the effort for a small team?
The honest trade-off: Claude will not fix messy paperwork, and a knowledge base built on outdated files gives confident wrong answers. The time goes into the documents, not the software. The upside is that once they are clean, dated and in one place, a new starter can get answers on day one without waiting for the one person who knows how everything works.
More guides on putting Claude to work in a UK business are collected on the Claude for business hub.
Claude AI FAQ and Knowledge Base for Your Business — FAQ
Can I build a knowledge base in Claude on the free plan?
Yes, within limits. Anthropic's help centre says Projects are available to all users, including free accounts, and free users can create up to five projects. Each project holds uploaded files plus custom instructions that Claude reads before answering. The main gap is capacity. Anthropic says the retrieval mode that expands project knowledge by up to 10 times is only on Pro, Max, Team and Enterprise plans, so a free project has to fit inside Claude's normal context window. For a small business with a price list, a returns policy and a handful of process documents, that is often enough to test whether the idea works before paying for anything. If staff need to share one project, you will need a Team or Enterprise plan.
What file types and sizes can I upload to a Claude project?
Anthropic's file upload guide lists PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX as supported document types, with XLSX needing code execution turned on. Inside a project, each file can be up to 30MB, and you can upload as many files as you like as long as the total content fits within Claude's context window. Project files are read as text, except multimodal PDFs. In practice, clean Word documents and plain text work best. Scanned PDFs of old handbooks, image-heavy brochures and spreadsheets with merged cells tend to produce weaker answers, so it is worth converting the important ones into simple text before uploading.
Is a Claude knowledge base safe for customer and staff data?
It depends on what you upload and which plan you use. UK GDPR still applies, so the safest default is to keep personal data out of the knowledge base entirely. Upload policies, prices, procedures and product information, not customer lists, payroll files or HR records. Check the data and training settings for your plan before you start, and write down what the project contains and who can access it. On Team and Enterprise plans, Anthropic says owners can share projects with view-only or edit access and can switch project sharing off. If a document would cause harm if it leaked, it does not belong in a shared project.
How many UK small businesses are already using AI tools like this?
The Office for National Statistics published figures on 20 July 2026 showing that 28% of businesses with 0 to 9 employees reported using at least one AI technology in June 2026. Among firms with 10 or more employees, large language models such as Claude were the most widely used type of AI, at 18%. The same ONS release found that only 10% of AI adopters with 10 or more employees describe their use as extensive. That gap is the point of a knowledge base. Most firms are trying a chatbot occasionally, while very few have given it their own documents to answer from, which is where the practical value sits.
Should a knowledge base answer customers directly or only staff?
Start with staff. An internal knowledge base lets the team ask Claude about returns, pricing rules or booking procedures and check the answer against the source document before passing it on. Mistakes get caught by a person, and you learn which documents are unclear or out of date. Once answers have been reliable for a few weeks, you can consider a customer-facing version, but that usually means a separate build with a chat widget, logging, a handover to a human and clear limits on what it can say. Putting a raw Claude project in front of customers on day one is how wrong refund promises get made.
