Why onboarding is slow, and what AI changes
A new hire spends the first weeks mostly on small questions. What do we call this client internally? Which version of the price list applies? Why do we do it this way? Repeated questions interrupt colleagues, and a new hire may hesitate to ask the same thing again.
AI can absorb those small questions, but only if the answers live somewhere the new hire’s AI can reach. That is the whole point. An AI that knows nothing about you gives a generic answer that is almost right, and during onboarding that is worse than no answer.
The misconception: “the AI knows us by now”
Teams that have used AI for a year often assume the AI knows their company by now. For the people already there, maybe, in their own memory and their own chats. But Anthropic says Claude’s memory is per person and not shared, and chats in a shared project stay private. The new hire does not automatically inherit colleagues’ personal AI memories; shared projects, connected sources and repository instructions can provide company context. What the team taught the AI over a year sits in other people’s accounts.
How to set it up
- Collect the way you work in a shared project. One Claude project per department or process, with instructions (how we write, who decides what, which terms we use) and knowledge (handbook, price list, examples of good work).
- Test before day one. Ask a few common questions and check that the new hire only sees sources they are allowed to see. Then give the new hire Can view on day one. They see the contents, knowledge and instructions and can chat in the project without changing anything.
- Put coding conventions in the repo. A CLAUDE.md in the project folder, or rules in .cursor/rules or an AGENTS.md, travel through version control. Commit the instructions and check which files and rules your tool actually loads; opening the repo alone does not guarantee it. See how to share Cursor context with a team.
- Have them ask the AI first during week one, and note where the answer was wrong or empty. Use the AI for routine questions, require a source for company-specific answers, and send unanswered or sensitive questions to a named colleague. That list is your best map of what has not been written down yet.
- Fix the gaps in the shared project, not in a private chat. Then it helps the next new hire too.
| What the new hire needs | Where it should live |
|---|---|
| How we write and communicate | Instructions of a shared project |
| Handbooks, price lists, examples | Knowledge of a shared project, or a connected document source |
| Coding conventions | CLAUDE.md, .cursor/rules or AGENTS.md in the repo |
| Who decides what and whom to call | A short document in the shared project |
| What colleagues already taught the AI | Nowhere, unless someone wrote it down |
What this means for a team
Onboarding with AI is mostly a test of how well your team knowledge is shared. Every question the AI can’t answer points to knowledge that lives only in a head or a personal account. It is the same gap you feel when someone leaves: see what happens when your AI champion leaves. For the documents side, see how to make company documents searchable for AI.
Where Bombos fits
The gap in the table, what colleagues already taught the AI, is exactly what Bombos is for. It is a plugin that gives every AI tool your team uses one shared memory, with team knowledge, the rules the team taught the AI, and what already sits in mail, documents and repos. A new hire then works from day one with an AI that knows how you work, whichever tool they start in.
Related questions
Will a new hire's AI already know how our team works?
Usually not. In Claude, memory is per person and not shared; only what sits in shared projects, instructions or files is reachable for the new hire.
What permission should a new hire get in a shared Claude project?
Start with Can view: they see the knowledge and instructions and can chat in the project, but can't change anything. Can edit is for whoever maintains the project.
Can AI onboard a new developer into our codebase?
AI can help explain the codebase when relevant conventions are available, for example in a CLAUDE.md, .cursor/rules or an AGENTS.md in the repo, but a teammate should still review the new developer's understanding and first changes.
