# Knowledge transfer when an employee leaves: get it out of private AI chats

When someone leaves, their AI chats often hold the context nobody else has. A checklist for capturing decisions, files and open work before the last day.

## The short answer

Start two weeks before the last day. Ask the person leaving to go through their AI chats, pull out the decisions, the current files and the open work, and hand each open item to a named colleague with a deadline. Archive what is finished. Do not count on the chats staying readable after the account is removed.

The rest of this article explains why AI chats have become a blind spot in offboarding, gives you a two-week checklist, and shows what to hand off and what to archive.

## Why AI chats are the new knowledge silo

A few years ago, a leaver's knowledge sat in their inbox, their files and their head. Email could be forwarded and files could be moved. Today a lot of the thinking happens in a chat with Claude or ChatGPT: the draft that went through six versions, the reason a price changed, the summary of a client call.

That context is usually private. In Claude, each chat stays with the person who ran it. Even in a project shared with the whole organisation, a new member gets the project's documents and instructions, but not other people's conversations.

It gets harder when the person leaves. According to one analysis published in August 2026, when a member is removed from a Claude organisation, the rest of the team loses access to that person's chats, and old snapshot links return “Conversation not found”. A project they shared with the whole organisation stays in the Team tab. Check the current behaviour for your own plan before you rely on it; products change.

Source: [Intelligence by Intent: In a shared ChatGPT project, everyone can read everyone’s chats. In Claude, nobody does](https://intelligencebyintent.com/insights/in-a-shared-chatgpt-project-everyone-can-read-everyones-chats-in-claude-nobody-does), last verified 2026-10-15.

Sharing a chat before the last day helps less than it seems. Claude's Share button creates a snapshot up to the moment you share. Attached files are not included, and later messages stay private. A snapshot shows what was said. It does not say who owns the work now, what the deadline is or which file is the final one.

Source: [Claude Help Center: sharing chats](https://support.claude.com/en/articles/10593882), last verified 2026-10-15.

So the goal is not to save the chats. It is to get the useful parts out of them, into a place the team can find, with an owner for everything that is still open.

## The two-week offboarding checklist

Two weeks gives the person leaving time to sort their chats, and gives the people taking over time to ask questions while the answers are still a message away. Use the first week to map and capture, and the second to hand off and close.

- [ ] List every client, project and recurring task the person owns, with the colleague who will take each one.
- [ ] Ask them to go through their AI chats from the last three months and mark the ones tied to open work.
- [ ] For each open item, write down the decisions already made and why, in two or three sentences.
- [ ] Record the facts that matter: deadlines, amounts, contract numbers, who signs off at the client.
- [ ] Find the final version of every working file and note where it lives. Drafts in chats do not count.
- [ ] Collect the prompts and project instructions they use often, so the next person can reuse them.

- [ ] Hand each open item to its new owner, with the summary, the facts and the files together.
- [ ] Have each new owner accept the work and ask their questions while the leaver is still there.
- [ ] Introduce the new owner to the client or the internal contact, in writing.
- [ ] Move finished work to the team archive, with a one-line description people can search by.
- [ ] Check that nothing is still sent only to the leaver: shared links, notifications, recurring requests.
- [ ] On the last day, review the list together: every item has an owner, a status and a deadline.

A short chat with the assistant can do a lot of the first week's work. The person leaving can ask their own Claude or ChatGPT: “Go through this chat and list the decisions, open questions and next steps for the person taking over.” They should check the result before anyone relies on it.

## What to hand off and what to archive

Not everything in a leaver's chats deserves a handover. Most of it is history. A simple rule: if someone will act on it, hand it off; if someone might need to look it up, archive it; if it is personal or out of date, leave it.

| Item | Hand off | Archive |
| --- | --- | --- |
| Open client work and its next step | Yes, to a named owner | |
| Decisions still in force, and why | Yes, as facts with the work | |
| Deadlines, amounts, contract numbers | Yes, as facts | |
| Final versions of working files | Yes, attached to the work | |
| Finished projects | | Yes, with a short summary |
| Superseded drafts | | Only if the history matters |
| Reusable prompts and instructions | Share with the team | |
| Personal notes and unrelated chats | | No: leave them out |

Two mistakes are common. The first is exporting everything into one big folder that nobody opens. The second is handing over files without the reasons behind them, so the new owner repeats a discussion that already happened. A good handoff is short: what the work is, what has been decided, what comes next and by when.

For a fuller look at what to ask the person leaving, see [how to continue a colleague's Claude work](/blog/continue-a-colleagues-claude-work/).

## Knowledge stays with the company

Colle is built for exactly this moment, and for every smaller handover before it. The person leaving tells their assistant who takes what, in plain words:

> Send James my call notes, he takes over the Nordlit account

James gets a notification with who sent it, what it is and by when. The work arrives as a case: the summary, the facts and the files travel together and stay in the case. James asks his own Claude or ChatGPT “what did I get?” and carries on with the full context.

This helps with offboarding in a few concrete ways:

- **Facts keep their history.** Deadlines, amounts, decisions and owners are versioned, and conflicting changes are flagged. The next person sees the current value and how it got there.
- **Work has an owner and a status.** A handoff is sent, accepted, done, returned or overdue. On the last day, you can see what is still open.
- **History travels with the work.** A forwarded handoff keeps its full history, so the third owner knows what the first one decided.
- **Teams, not just people.** Work can go to a space, where all members are notified and one picks it up. Useful when the replacement is not hired yet.
- **It can be found later.** People search the cases they can access, by keyword or by meaning, including files by their description.

Because the work lives in cases rather than in one person's chats, the next handover is a sentence, not a project. For project teams, see [project handovers](/use-cases/projects/). To see the three steps in full, read [how Colle works](/how-it-works/), and for costs, see [pricing](/pricing/).

## FAQ

### Can we just export the leaver's AI chats?

You can try, but an export is a pile of conversations, not a handover. It has no owners, no deadlines and no clear final version. Pull out the decisions, facts and files, and hand each open item to a named person.

### How early should knowledge transfer start?

Two weeks is a good minimum for most roles. It leaves one week to capture and one week for the new owners to ask questions while the person leaving can still answer.

### What if the person has already left?

Talk to whoever worked closest with them, check shared projects and drives, and contact clients for the latest agreed terms. Then start handing work off properly, so the next departure is easier.

[Join the Colle pilot](https://www.trycolle.com/#pilot)

---

Canonical page: https://www.trycolle.com/blog/knowledge-transfer-when-an-employee-leaves/ · Last updated: 2026-10-15