How to Build a Useful Company Knowledge Base for AI

When people want AI that “knows the business,” they usually mean an assistant that answers from the company’s own policies, prices and procedures. In ChatGPT and Claude, the practical tool is a project: a workspace with standing instructions and a set of reference files. A connected source, such as a shared drive the assistant can search, is the other option.

Two clarifications help set expectations. This is not training your own model: uploading files and writing instructions gives the assistant material to read while it answers, and the underlying model stays the same. It is also not a custom database project: both products handle the files themselves, so the work is choosing, cleaning and owning the content, not engineering.

Start with fewer, better sources

Answers can only be as good as what you put in. An export of every file the company has ever produced hands the assistant old prices, abandoned drafts and contradictory email threads. A short list of current, approved documents leaves far less room for error.

Good candidates are the documents you would hand a new employee: the current price sheet, service policies, approved email wording, a frequently asked questions page and step-by-step procedures. Leave out drafts, superseded versions, personal notes and raw email threads. Name each file so it says what it is and when it was approved; Anthropic recommends clear, descriptive filenames for project files.

The platforms’ limits point the same way. A ChatGPT project on Business or Enterprise holds up to 40 files. Claude project files can be up to 30 MB each, and the total has to fit within Claude’s context window (the amount of text it can work with at once), though paid plans switch automatically to searching the files as a project nears that limit. When we set up a workspace, we start with up to 20 organized documents, about 100 pages in total, and add more only when a test shows a real gap.

Choose one source of truth for each topic

For each topic the assistant will answer, decide which single document is authoritative and who owns it:

If a topic has no owner or no current document, that is a gap to fix in the business first. The assistant can’t settle a question the company hasn’t settled.

Resolve conflicts before the assistant sees them

Conflicting material invites confident wrong answers. If last year’s policy and this year’s policy are both in the project, the assistant may quote either one, or blend them.

• Remove superseded versions rather than adding the new one alongside the old.
• Put the effective date in each file’s name and at the top of the document.
• Say which source wins in the project instructions, for example: “If the price sheet and any other document disagree, use the price sheet and mention the conflict.”
• Keep instructions and files consistent. If the instructions say one thing and a policy file says another, fix one of them. In ChatGPT, project instructions override a user’s own custom instructions inside that project, so they suit team-wide rules.

Name an owner and set an update cadence

A knowledge base without an owner goes stale quietly. Name one person, not “the team,” who approves what goes in, removes what is out of date, and re-runs a few test questions after each change. That person doesn’t have to write every document; they keep the set current.

Tie updates to events as well as the calendar. A price change, a new policy or a revised process should trigger an update that week; a monthly or quarterly review catches the rest. Coordinate removals too: in a shared ChatGPT project, deleting a file removes it for everyone.

Read permissions are not the same as project sharing

In your file system, folder permissions decide who can open a document. In a shared AI project, project membership decides who can see the files in it. OpenAI says all members of a project “can view and download files added to the project.” Anthropic says members with view access can see the project’s contents, knowledge and instructions, and a public Claude project is open to everyone in the organization. Neither vendor says an uploaded copy keeps the permissions of the folder it came from.

The consequence: a restricted file uploaded to a shared project becomes visible to every member of that project, even if the original sits in a locked folder. Salary information, HR matters and one client’s financials don’t belong in a project the whole team can open.

Chats need the same thought. Members of a shared ChatGPT project can view the project’s chats; in Claude, chats in a shared project stay private unless the person shares them. Workspace owners on both platforms can control project sharing, and in Claude Team and Enterprise the organization settings for sharing and public projects are on by default.

Uploaded copies versus connected sources

An uploaded file is what we call a snapshot: a fixed copy as it stood when uploaded. It doesn’t change when the original is edited, so the owner has to replace it. Snapshots are predictable and easy to control, which suits stable reference documents.

A connected source is read from where it lives. OpenAI says its company knowledge feature on Business, Enterprise and Edu plans “respects permissions in the connected source,” so a member retrieves only what they can already access. Anthropic says Claude’s connectors inherit each person’s permissions, and that Google Docs added through the Google Drive connector sync to the latest version.

Connected sources have limits too. In Claude Team and Enterprise, connectors are only available in private projects. In ChatGPT, the Google Drive app doesn’t sync content in advance when added within a project, though it can still search files. Anthropic also warns that a connector using one shared credential reaches whatever that credential can access. When we set up a workspace, we include up to one supported read-only connection where the plan and permissions allow, and otherwise document a simple approved-upload process.

Test it, especially when the answer isn’t there

Before the team relies on it, run the knowledge base against about 20 representative questions with known answers. The most revealing tests are the ones where the right response is “I don’t have that information.” An assistant that fills a gap with a plausible guess, such as a refund window your policy never mentions, is more dangerous than one that admits the gap.

Put the expected behavior in the project instructions, in plain words: “If the answer isn’t in the project files, say so, don’t guess, and tell the user to ask the operations manager.” Then check that it does.

OpenAI’s own guidance for company knowledge is to review the answer and any source links before relying on the information, and that habit applies on either platform. Re-run a handful of these tests whenever the owner changes the sources or the instructions.

For earlier steps, see what to prepare before AI setup. For how we approach this work, see company knowledge. If you are still choosing a platform, see ChatGPT vs. Claude for business.