An AI knowledge base is an internal assistant that answers staff questions from your own company documents, such as policies, procedures, contracts and past projects, and points to the file each answer came from. For a small business, the safe way to build one is on a business-grade tool that respects the file permissions you already have, starting from a cleaned-up set of documents. As of October 2026 that means one of four routes: a ChatGPT knowledge base built with Projects, the Company Knowledge plugin and workspace agents on ChatGPT Business or Enterprise; Claude Projects with the Microsoft 365 connector on Claude Team or Enterprise; Microsoft 365 Copilot over SharePoint and OneDrive; or a custom retrieval build when the answers live in systems none of those reach. Below: how it works, the options and what they cost, how to build one in ChatGPT and in Claude, a setup checklist, and the questions small business owners ask most.
How does an AI knowledge base work?
Most use a technique called retrieval-augmented generation, or RAG: when someone asks a question, the system first searches your approved documents, then has an AI model write an answer from what it found, with links back to the sources. The model is not retrained on your files; it reads the relevant ones at the moment of the question. That is why answers stay current when a document changes, and why the quality of your documents sets the quality of the answers.
It will still get things wrong sometimes. NIST's Generative AI Profile names the risk "confabulation": confidently stated but false content. Require citations, teach staff to open the source before acting on anything important, and keep a person in the loop for HR, legal and financial questions.
Which AI knowledge base option fits a small business?
| Option | Fits when | Check before you buy |
|---|---|---|
| ChatGPT Business or Enterprise: Projects, the Company Knowledge plugin and workspace agents | Your team already works in ChatGPT and the sources are Google Drive, SharePoint, Teams, Outlook, Slack, Box or Notion | ChatGPT Business (formerly ChatGPT Team) is 2 to 200 seats at $20 per user per month billed annually or $25 monthly as of October 2026; a project holds 40 files; apps are enabled by default, so decide which ones stay on; no BAA on Business |
| Claude Team or Enterprise: Projects plus the Microsoft 365 connector | You want knowledge organized by project with written instructions, and the documents live in SharePoint, OneDrive, Outlook or Teams | Claude Team is 2 to 150 seats at $20 per seat per month billed annually or $25 monthly as of October 2026; Enterprise is $20 a seat annually with a 20-seat minimum and usage billed at API rates; the connector needs Microsoft Entra admin consent and is read-only by default |
| Microsoft 365 Copilot | Your documents already live in SharePoint, OneDrive, Teams and Outlook | Permissions and oversharing first; as of October 2026 Copilot Business lists at $21 per user per month on annual billing, with an $18 first-year offer on annual subscriptions started by December 31, 2026 |
| An agent built in Copilot Studio over a chosen set of documents | One team needs answers from one body of content, such as an HR handbook or a procedures library | Who keeps the source documents current, and who can use the agent |
| A custom retrieval build (RAG development) | Answers have to come from systems outside the SaaS apps above, such as a line-of-business database, help desk or file server, or inside your own application | Connectors, whether file permissions carry over, where data is stored, who maintains it, and whether your content is used to train models |
How do you build a knowledge base with ChatGPT?
On ChatGPT Business or Enterprise, in this order. Personal plans (Free, Go, Plus and Pro) have Projects too, but no workspace controls, fewer files, and no way to build or publish GPTs, so a company knowledge base belongs in a workspace.
- Start with one Project per topic. A project keeps related chats, files and instructions together; on Business and Enterprise it holds 40 files (10 can be uploaded at once), project instructions override a user's general custom instructions, and Google Drive files or folders and Slack channels can be added as sources by link, per OpenAI's Projects article. Shared projects take up to 100 collaborators in a workspace and are always project-only memory, so they cannot draw on a member's outside memories.
- Connect the apps for live sources. ChatGPT's apps (what OpenAI used to call connectors) reach Google Drive, SharePoint, Microsoft Teams, Outlook, Slack, GitHub, Box and Notion in real time, each user signs in to the source, and OpenAI states that ChatGPT respects your organization's existing permissions and does not train on data accessed through apps by default. In Business, apps are enabled by default and admins set workspace-wide availability, and the Microsoft apps need a Microsoft Entra administrator to grant consent. Review the list before you point staff at it.
- Turn on Company Knowledge for cross-source answers. Company Knowledge is a plugin, available on ChatGPT Business, Enterprise and Edu, that answers from the connected sources and respects their permissions, so a member retrieves only what they are already allowed to open; it also works with a custom MCP server that offers search and fetch, per OpenAI's Company Knowledge article. This is the piece most owners mean when they ask for "a ChatGPT knowledge base."
- Package a curated assistant as a plugin or a workspace agent. For a fixed, expert-curated body of content such as a benefits handbook, build a plugin with reference files, or a workspace agent, which on Business and Enterprise combines a chosen model with tools, apps, custom MCP servers, skills and files (512 MB per file, 10 GB per agent), can run on a schedule, and has write actions set to "Always ask" by default. Our guide to building a company assistant in ChatGPT walks through it, and ChatGPT Projects vs custom GPTs covers which container fits which job.
- If you have custom GPTs, plan their migration now. OpenAI has scheduled custom GPTs to retire on December 11, 2026 (February 11, 2027 for Enterprise workspaces with an approved deferral), new GPT creation on Enterprise is planned to end on October 26, 2026, and personal plans can no longer create them. Knowledge files move to plugin reference files; GPT Actions do not transfer. Treat GPTs as the retiring path, not a place to start.
Limits worth knowing before you upload: a document can be up to 512 MB and text files are capped at about 2 million tokens, storage is 25 GB per user and 100 GB per organization, and retrieval of images and charts inside PDFs is an Enterprise feature; other plans use text extraction. Your ChatGPT and SharePoint setup decides what the knowledge base can see, so fix permissions there first.
What is the Claude equivalent of a ChatGPT knowledge base?
Claude Projects. Anthropic describes them as self-contained workspaces with their own chat histories and knowledge bases: each project has written instructions and uploaded knowledge, and on paid plans a retrieval mode expands the knowledge capacity by up to 10x as the project nears its context limits, per Anthropic's Projects article. Free accounts are limited to five projects, and Anthropic does not publish a file-count or size limit on that page. On Claude Team and Enterprise, a project can be shared by email, by list or organization-wide with "Can view" or "Can edit" roles, and owners can turn sharing off. Skills sit alongside: a skill is a packaged procedure Claude applies when it recognizes the job, and owners can provision skills organization-wide on Team and Enterprise. Our guide to Claude Projects vs Skills explains how to split knowledge from procedure.
For live documents, the Microsoft 365 connector searches SharePoint, OneDrive, Outlook and Teams. On Team and Enterprise only owners can enable it, it requires a Microsoft Entra tenant on a Microsoft business plan and a one-time Global Administrator consent, it uses delegated permissions so each person sees only what they can already open, it is read-only by default, and Anthropic says it does not cache file content. The permissions review before rollout is in Claude Microsoft 365 connectors. Enterprise search is included in Team seats, and Claude Enterprise adds audit logs, custom data retention and a HIPAA-ready configuration with a BAA. Our Claude consulting and implementation service covers the setup.
When do you need RAG development instead of a built-in knowledge base?
When the answers live somewhere the built-in options cannot reach, or the knowledge base has to live inside something you own. Typical triggers: the source is a line-of-business database, a help desk, a practice management system or an on-premises file server; the assistant has to sit inside your own application or customer portal; you need your own permission model, logging and retention rather than the SaaS vendor's; or the volume makes per-seat tools the wrong shape. A custom build pairs a retrieval layer over your documents with a model called through the OpenAI or Claude API, and increasingly an MCP server over the system of record, which Claude and ChatGPT Business and Enterprise can both query with permissions enforced on your side. RAG vs fine-tuning explains why retrieval, not training, is the right tool for company knowledge; custom AI applications, OpenAI integration services and MCP server development are the NetSys services for that route.
A checklist before you switch it on
- Audit permissions. Find folders shared with everyone or by anonymous link, and fix them before the AI can surface them.
- Clean the content. Archive outdated policies, duplicate drafts and old price lists, because the assistant will quote them.
- Label what is sensitive. Payroll, HR and client files should carry sensitivity labels or sit in restricted libraries.
- Write 20 test questions with known answers, and run them before and after launch.
- Start with a pilot group of five to ten people who will report bad answers.
- Set a feedback route so a wrong answer leads to a fixed document, not a shrug.
- Close the side door. Tell staff which AI tools are approved, so company data does not end up in free consumer apps.
Frequently asked questions
What is an AI knowledge base?
It is a private assistant that answers from your company's own content rather than the open internet. You point it at your documents, and employees ask questions like "what is our refund policy" or "which vendor did we use for the Smith job" and get a direct answer with a source. The technique behind it is called retrieval-augmented generation, or RAG, which means the AI looks up your real files before answering.
Can you use ChatGPT as a company knowledge base?
Yes, on ChatGPT Business or Enterprise. Put curated documents in Projects (40 files each), connect the apps your documents live in, turn on the Company Knowledge plugin so answers come from every connected source with permissions respected, and package any fixed assistant as a plugin with reference files or a workspace agent. A personal Plus or Pro account is not a company knowledge base: there are no admin controls, fewer project files, and no way to publish an assistant to the team.
What happens to custom GPT knowledge files when custom GPTs retire?
OpenAI's migration moves a GPT's knowledge files into reference files on a plugin, and the migrated plugin starts private; GPT Actions do not transfer and have to be rebuilt as apps or custom MCP servers. Existing conversations stay accessible after the December 11, 2026 retirement. If a GPT is doing real work for your team, inventory it now, decide whether a plugin, a workspace agent or a project replaces it, and test the replacement before the deadline.
Why would a small business want one?
Because answers are scattered across folders, inboxes and people's heads, and finding them costs time every day. An AI knowledge base cuts that hunt down to a question, gets new hires productive sooner, and captures know-how that would otherwise walk out the door when someone leaves.
Is our company data safe if we use AI this way?
It can be, if you use a business-grade tool with the right settings. Microsoft's Copilot privacy documentation, for example, says prompts, responses and the data Copilot reads through Microsoft Graph aren't used to train foundation models, and that Copilot only surfaces data each user already has at least view access to. ChatGPT Business and Enterprise and Claude Team and Enterprise do not train on workspace data by default either; our AI data privacy FAQ has the plan-by-plan table. The bigger risk is staff pasting sensitive information into free consumer AI tools, which our guide to shadow AI and data leaks covers. A sanctioned internal knowledge base is the safer alternative, not the danger.
Does the AI see documents an employee should not?
Only if your permissions already allow it. A properly configured knowledge base honors the same permissions your files have, so a staffer only gets answers from documents they are allowed to open. The catch is that many small businesses have overshared files without knowing it, and AI makes them easy to find. Fixing that first is one of the most important parts of the project; our guide to Copilot oversharing shows where to look.
Do we need Microsoft Copilot, or are there other options?
You have four. If you run Microsoft 365, Copilot answers from your SharePoint and OneDrive content and is a natural starting point. ChatGPT Business and Claude Team reach the same Microsoft 365 content through their apps and connectors, and ChatGPT also reaches Google Drive, Slack, Box and Notion. A custom build pulls from systems like your CRM, help desk or a file server. The right choice depends on where your documents live, which tool your team already uses, and how much you want to connect. Our Copilot FAQ covers the Microsoft path in detail.
How is this different from an AI chatbot on our website?
A website chatbot answers customers using public-facing information. An internal knowledge base answers your own staff using private company documents. The technology overlaps, but the audience, the content, and the security requirements are different. Some businesses run both. Our AI chatbot FAQ covers the customer-facing side.
What does it take to set one up?
Three things: clean documents, correct permissions, and a tool that fits your stack. The work is mostly preparation, organizing and tidying the content the AI will draw from, then connecting it and testing answers for accuracy. Feeding it messy or outdated files is the fastest way to get wrong answers, so a short cleanup pays off.
What is included in an AI knowledge base project?
A complete project covers choosing the tool, auditing and fixing permissions, cleaning and organizing the source content, connecting the sources, configuring who can use it, testing answers against known questions, a pilot with a small group, and short training on how to ask questions and check sources. Ongoing, someone keeps the content current and reviews reported bad answers.
How does onboarding work?
Start with one body of content and one group, such as office procedures for the operations team. Fix permissions and clean that content, connect it, and run your test questions. Give the pilot group two to four weeks, collect every wrong or missing answer, fix the documents behind them, then widen the content and the audience a step at a time.
What affects the cost of an AI knowledge base?
Licenses are the visible part: per-user subscriptions such as ChatGPT Business, Claude Team or Copilot Business, or a platform fee for a standalone tool. The less visible parts are the cleanup and permissions work, which depend on how messy your file shares are, connectors to systems outside your main suite, any custom retrieval build, and the ongoing time to keep content current. A small group with tidy SharePoint libraries costs far less to set up than a company with years of overshared folders. NetSys quotes the build as a fixed project after a scoping call and does not publish a price list.
Who handles support and escalations?
Split it by job. Each department owns its source documents and fixes them when an answer is wrong. IT, or your provider, owns permissions, connectors, licenses and platform problems. Questions with legal, HR or financial consequences should send people to a named person rather than the assistant, and staff should know who that is.
How is a knowledge base different from an AI agent?
A knowledge base answers questions. An AI agent takes actions, like drafting a reply or updating a record, often using that same knowledge. Many businesses start with a knowledge base because it is lower risk and immediately useful, then add agents later. Our AI agents FAQ explains where agents fit.
By Jamie Baum for The NetSys Group, which has delivered managed IT, cybersecurity and cloud services from Brooklyn since 1998. Our engineers hold degrees in electrical and computer engineering. We work on-site across New York City, Long Island, Westchester, the Hudson Valley, North Jersey and Fairfield County, CT, and remote-first everywhere else.
What NetSys does
NetSys picks the route with you, fixes permissions and cleans the source content, sets up Projects, Company Knowledge and workspace agents in ChatGPT, Claude Projects and connectors in Claude, or Copilot in Microsoft 365, builds the custom retrieval layer when the answers live elsewhere, and runs the pilot with your test questions. See custom GPT and ChatGPT assistant development for the ChatGPT route, Microsoft Copilot consulting for the permissions cleanup on the Microsoft 365 route, and AI services for the rest. Scope a company knowledge assistant and we will start with the one body of content your staff ask about most.
Discuss ai services for your business.
Tell us about your current systems, the result you need and your timeline. We will discuss the work, responsibilities and pricing before you decide on an engagement.