The first choice: who asks the questions
“AI chatbot with a knowledge base” covers two very different products. An internal assistant answers colleagues from handbooks, procedures, price lists and project notes. A customer-facing bot answers the public from approved help content, on your website or in your support inbox. They differ in who may see what, how wrong answers hurt, and which tools fit.
| Route | Who asks | Typical tools | Set-up time | Watch out for |
|---|---|---|---|---|
| 1. Shared project or assistant | Colleagues | Claude Projects, ChatGPT projects or custom GPTs, Copilot Studio agents | Hours | Copies of files that go stale; one setup per tool |
| 2. Support AI agent | Customers | Zendesk AI agents, Intercom Fin | Days | Only as good as your help center |
| 3. Your own RAG chatbot | Either | A vector store, an LLM API and your own code | Weeks | You own retrieval quality, security and upkeep |
Route 1: an internal chatbot in the tools you already pay for
- Pick the home. If the team works in Claude, create a project and add the documents under project knowledge. On Team and Enterprise plans you can share a project with specific people or the whole organization, as viewer or editor. In ChatGPT the equivalent is a custom GPT or a shared project. In Microsoft 365, build an agent in Copilot Studio.
- Add knowledge. Upload the stable documents. For sources that change, connect them instead: Copilot Studio can use SharePoint, Dataverse, uploaded files, public websites and Microsoft Search connectors as knowledge sources, and Claude Team and Enterprise owners can add connectors through MCP that each member then connects with their own account.
- Write instructions. Tell it to answer only from the sources, cite them, and say so when the answer is not there. Copilot Studio even has a setting, Allow ungrounded responses: with generative orchestration enabled, turning it off blocks responses in turns that used neither a knowledge source nor a tool.
- Test with real questions from the last month, including ones the documents cannot answer.
The differences between Claude Projects and custom GPTs are covered in Claude Projects vs ChatGPT Projects and custom GPTs.
Route 2: a customer-facing support agent
Here the knowledge base is your help center. Zendesk AI agents generate replies from a Zendesk help center and from external content brought in through a web crawler or knowledge connector, which usually syncs every 24 hours. Unauthenticated visitors only get answers based on public articles. Intercom Fin can use Intercom articles and snippets, synced public websites, PDF uploads and content from tools such as Zendesk, Guru, Notion and Confluence.
The work is mostly editorial: fill the gaps in your help center, remove contradicting articles, and review the questions the bot could not answer. Zendesk itself warns that connecting too many sources can reduce accuracy and slow replies.
Route 3: build your own
Building your own makes sense when your data sits in a system no tool connects to, when the chatbot must live inside your own product, or when the data may not leave your servers. You then choose how documents are split, embedded and searched, how permissions are checked, and how updates flow in. That is real engineering and ongoing maintenance, not a one-off project.
The misunderstanding: the chatbot is the hard part
Every route above gives you a working chatbot in a short time. What decides whether people keep using it is the knowledge behind it: one leading version per topic, someone who owns updates, and answers that cite where they come from. A chatbot on top of three conflicting price lists confidently picks one. For a comparison of dedicated software, see Which AI knowledge base software fits a team?, and for preparing the content, How do you build a knowledge base for AI agents?
What this means for a team
An internal chatbot in one tool helps the people who use that tool. The colleague in Cursor, the developer in Claude Code and the account manager in ChatGPT each still work from their own context. And corrections made in a conversation (“that rate changed in October”) stay in that conversation unless someone updates the source.
Where Bombos fits
If your team uses more than one AI tool, Bombos keeps one shared memory underneath all of them instead of a chatbot per tool. It is a plugin for Claude, ChatGPT, Copilot, Grok, Codex, Cursor and your own agents that learns how your team works, draws on what already sits in mail, documents and repos, and shows the sources of each answer. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.
Related questions
Do I need to train the chatbot on my documents?
No. Uploading documents does not retrain the model. Tools use them as context, directly or through retrieval (RAG). Claude Projects automatically enables RAG when project knowledge approaches context limits.
Will the chatbot show restricted documents to the wrong people?
That depends on the tool. Copilot Studio can use the asking user's Microsoft Entra ID sign-in so it only surfaces what that user may see, and Zendesk only uses restricted articles for authenticated customers. Files uploaded into a shared project or GPT are visible to everyone who can use it.
How fresh are the answers?
As fresh as the source. Uploaded files stay as uploaded until someone replaces them. Zendesk syncs external sources usually every 24 hours, Intercom syncs websites weekly, and live connectors read the source when asked.
Can one chatbot serve both staff and customers?
Rarely a good idea. Customers should only reach public, approved content; staff need internal notes, prices and procedures. Most teams run two separate setups.
Sources
- What are projects? (Claude Help Center)
- Get started with custom connectors using remote MCP (Claude Help Center)
- Knowledge sources summary (Microsoft Copilot Studio, Microsoft Learn)
- Connecting knowledge sources to power generative replies in AI agents (Zendesk)
- Add your support content for Fin (Intercom Help)
