How to Set Up a Business Knowledge Base Your AI Can Actually Read

A step-by-step guide for small businesses and startups: how to organize your company's files so Claude can find them, read them and cite them.

By Alex Guldbeck, AI consultant for small businesses and startups. Last checked against Anthropic's documentation on 23 October 2026.

Short answer: Put your company's important documents in one cloud location, organize them into a small number of clearly named folders, and add a plain-text README that tells the AI what is where. Then connect that location to Claude, either by adding files to a Claude Project or through a connector such as Google Drive, and write short project instructions telling it to answer from your files and cite them. Keep passwords, personal data and anything you wouldn't hand a new employee out of it, and give one person the job of keeping it current.

Most "AI for business" advice stops at "turn on the Google Drive connector." That part takes two minutes. The part that decides whether the answers are any good is how your files are organized before the AI ever sees them. That is what this guide covers.

Step 1: Decide what goes in

Start with the questions your team asks every week, not with the folders you happen to have. Good candidates:

• Standard operating procedures and how-to documents
• Pricing, proposals and the terms you actually agreed
• Client and project histories, including what was promised and when
• Product and service descriptions
• Vendor records and what went wrong with each one
• Brand, tone and marketing guidelines
• Decisions and the reasons for them

A useful test is this: if the person who knows something quit tomorrow, would you need it? If yes, it belongs in the knowledge base.

Step 2: Pick one home

Choose one place where the knowledge base lives: Google Drive, Notion or Dropbox. Use whichever your team already works in. The worst setup is the common one, where half the documents live in one cloud, half in another, and the rest on someone's laptop.

One home does not mean moving every file you own. It means the documents that matter have one authoritative copy, and everyone knows where it is.

Step 3: Use a shallow, numbered folder structure

Keep it to about seven to ten top-level folders and no more than three levels deep. Numbers keep the order stable and make the folders easy to refer to in instructions.

Company Knowledge Base/
  00-README                (the index; see Step 5)
  01-Company/              mission, team, org chart, policies
  02-Products-Services/    descriptions, specs, pricing
  03-Clients/              one folder per client
  04-Projects/             one folder per project
  05-Operations-SOPs/      how we do things, step by step
  06-Sales-Marketing/      proposals, case studies, brand guide
  07-Finance-Legal/        contracts and terms (restricted; see Step 9)
  08-Decisions-Log/        what we decided, when, and why
  99-Archive/              superseded material, kept but out of the way

Notion users can build the same thing as top-level pages with sub-pages. The principle is the same: few top-level areas, predictable names, and nothing important buried six clicks deep.

Step 4: Name files so a stranger could find them

The AI reads file names when it searches. So does every new hire. Use one pattern everywhere:

YYYY-MM-DD_client-or-area_document-type_v1

For example: 2026-03-14_acme_proposal_v2, or 2026-01-05_operations_onboarding-sop_v3.

• Put the date first, in year-month-day order, so files sort correctly.
• Say what the document is: proposal, contract, SOP, invoice, notes.
• Use version numbers, never "final," "final-2" or "USE THIS ONE."
• When a document is replaced, move the old one to the archive instead of leaving two versions side by side. Two conflicting versions are the most common reason an AI gives a confident wrong answer.

Step 5: Write a README at the top

The single most useful file in the knowledge base is a one- or two-page index at the top level. It tells the AI, and any person, what the company is and where to look. Keep it in plain text or a Google Doc.

A README should cover:

• Who we are: one paragraph on what the company does and for whom.
• Map: one line per top-level folder saying what is in it.
• Sources of truth: which document wins when two disagree (for example, "current prices live in 02-Products-Services/price-list; ignore prices in old proposals").
• Vocabulary: product names, abbreviations and internal terms a new person wouldn't know.
• Owner and last update: who maintains the knowledge base, and when the README was last checked.

Step 6: Use formats the AI can read

Claude reads text. Anthropic's help pages say that files added to a Project are processed by text extraction only, except for PDFs, which it can also read visually, and that each project file can be up to 30 MB. When Claude reads a file through its Google Drive connector, it extracts the text only, and images embedded in documents are not processed.

In practice:

• Prefer Google Docs, Word files, text-based PDFs, spreadsheets and plain text.
• Run text recognition (OCR) on scanned documents, or the text inside may not be readable.
• If a diagram or screenshot carries important information, write that information out in words next to it.
• Split very large documents into sensible parts with clear names.

Step 7: Connect it to Claude

There are two common routes, and many businesses use both.

Route 1: A Claude Project. A Project is a workspace with its own uploaded files ("project knowledge") and its own instructions. Anthropic says Projects are available to all users, including free accounts, which can create up to five. On paid plans, when a Project's knowledge approaches the limit of what Claude can hold at once, it automatically switches to retrieval (RAG), searching your files for the relevant passages, which Anthropic says expands capacity "by up to 10x." On Team and Enterprise plans, Projects can be shared with colleagues with "Can view" or "Can edit" permissions.

Route 2: A connector. A connector lets Claude search your files where they already live. Anthropic's Google Drive connector can search and read your Drive and mirrors your existing Google permissions, so Claude can only open files your account can already see. Google Docs added to a Project sync from Drive, so Claude works from the latest version. Drive files can be added only to private projects, not shared ones. On Team and Enterprise plans, an Owner must enable a connector for the organization before anyone can use it, and each person still connects their own account. Notion's own documentation says that once connected, the AI "can read and update content that you can access" in the workspace you chose. For Dropbox and other tools, check the Connectors Directory in Claude's settings.

A simple rule: put stable reference material (SOPs, price lists, the README) in Project knowledge, and use a connector for material that changes daily.

Step 8: Write project instructions that do real work

Project instructions are the standing orders the AI follows in every conversation in that Project. Keep them short and specific. A good starting set:

You are the internal assistant for [Company], a [one-line description].
Answer from the company knowledge base first. Start with 00-README to find your way.
Always name the document you used, with its file name or link.
If the knowledge base does not contain the answer, say so plainly. Do not guess.
When documents disagree, follow the sources of truth listed in the README and point out the conflict.
Current prices come only from [file name]. Never quote prices from old proposals.
Write in our house style: [two or three plain rules].
Do not draft anything that commits the company to a price, deadline or legal term without saying it needs a human's approval.

What to leave out of instructions:

• Whole documents. Those belong in the knowledge base, not in the instructions.
• Facts that change often, such as this month's promotion. Put those in a file and point to it.
• Passwords, keys or account numbers, ever.
• Long lists of things to avoid. A few clear rules work better than thirty.

Step 9: Keep the wrong things out

Some material should never go into a knowledge base an AI can read, even a private one:

• Passwords, API keys, recovery codes and login details
• Payment card numbers and bank account numbers
• Social Security numbers and government ID scans
• Health or medical information about staff or clients
• Personnel files, salaries and performance reviews, unless they sit in a separate, restricted area
• Client material your contracts forbid you to share with third-party tools

If you're not sure, ask whether you would hand the file to a new employee on their first day. If not, keep it out, or keep it in a restricted folder that is not connected.

Step 10: Set basic data-safety rules

• Use a business plan. Anthropic's privacy pages say it "does not use the data you share when using our commercial products to train our models," unless you choose to join its Development Partner Program.
• Remember what connectors can do. Anthropic notes that connecting a service grants Claude permission to access, and potentially modify, data in it based on your account permissions. Connect only what the work needs.
• Use least access. Connect the knowledge-base folder, not the whole drive. Restricted folders stay unconnected.
• Write it down. A one-page AI use policy that says what goes in, what stays out and who to ask saves arguments later. Have your team read and sign it.

Step 11: Maintain it

A knowledge base that nobody maintains is out of date within a quarter. Keep it light:

• One owner. One named person is responsible for it.
• An inbox folder. New documents land in one "to file" folder and get sorted weekly.
• A decisions log. Every significant decision gets one dated line: what was decided and why. This is often the most valuable file in the whole system.
• Archive, don't delete. Superseded documents move to the archive, so the AI stops quoting them but nothing is lost.
• A quarterly check. Update the README, clear out duplicates, and re-check what is connected.

Step 12: Test it before you trust it

Ask it five questions you already know the answers to, such as what you quoted a particular client last year or which vendor caused the shipping problem. Check that it found the right document and named it. Where it fails, the fix is almost always in the files (a missing document, a duplicate or a vague name), not in the AI.

A real example

A startup founder's company records were spread across two computers, two cloud services and a USB drive. I merged about 2,000 files into one organized knowledge base, checked that nothing was left behind, and set up his own Claude to work from it with standing instructions.

When to get help

You can do all of this yourself, and this guide is meant to let you. If you'd rather have someone who has done it before set it up with you, the AI Setup Session is a 90-minute working session for $750. We look at where your knowledge lives, decide the structure and what stays out, and set up your Project and instructions together. The fee is credited toward any project. For a full build, the Business Knowledge Base service starts at $3,500.

About the author: Alex Guldbeck is an AI consultant and AI automation consultant for small businesses and startups. He has worked in and alongside operations and digital marketing departments for 25 years, and grew one client's organic traffic from about 450 to 7,500 visitors a month. He builds Business Knowledge Bases, connects Claude to the tools teams already use, and builds custom apps and dashboards.

Frequently asked questions

Sources: Anthropic, "What are projects?". Anthropic, "Upload files to Claude". Anthropic, "Use Google Workspace connectors". Anthropic, "Use connectors to extend Claude's capabilities". Anthropic Privacy Center, "Does Anthropic act as a data processor or controller?". Notion, "Connect to Notion MCP". All checked 23 October 2026.