Where client files get mixed up in AI tools
A firm works per client: a file, the people assigned to it, and knowledge that belongs to that client (agreements, exceptions, how their books are kept). Most AI tools work per person: one account, one memory that comes along into every chat. That mismatch is the risk.
How the layers work in Claude, as of November 2026:
| Layer | Belongs to | Shows up in |
|---|---|---|
| Account memory | One user, from chats outside projects | Every chat, including projects, unless turned off per project |
| Project memory | The project, within the user’s memory | Chats in that project, and chats outside projects if memory is on |
| Project knowledge and instructions | The project, shared with everyone who has access | Only that project |
So a staff member who worked on client A outside a project can see that context come back inside client B’s project. Project A’s memory does not enter project B, but only with Use account memory off is a project closed in both directions.
In ChatGPT Business the default for shared projects is stricter: OpenAI states that shared projects automatically use project-only memory and cannot access an individual member’s memories or context from outside the project.
A setup per client file
- One project per client. Sharing projects requires Claude Team or Enterprise; in ChatGPT, Business, Enterprise and Edu users can invite individuals or workspace groups.
- Close the memory. In Claude, turn off Use account memory under Memory on the project page. In ChatGPT, keep client work in shared projects, where project-only memory is automatic.
- Share per file, not with everyone. Claude offers Can view and Can edit; ChatGPT offers chat and edit access. Only people on the engagement belong in the project.
- Put fixed knowledge in the project, not in chats. Claude notes that context is not shared across chats in a project unless it is in the project knowledge base.
- Set firm-wide policy at the top. On Claude Team and Enterprise, owners decide whether memory exists at all and admins can turn off project sharing. In ChatGPT Business, a workspace owner can turn memory off for the workspace, which deletes members’ existing saved memories.
The common misunderstanding
Many firms assume a shared project means shared memory. Half true. Project knowledge and instructions are the same for everyone in the project. But what a staff member learned in their own chats stays theirs: in ChatGPT Business, memories cannot be transferred to other users, even in the same workspace. What one person found out about a client reaches a colleague only once someone puts it in the project.
What this means for the firm
Two requirements seem to pull against each other. Knowledge about a client must be shared with everyone on the file, also when someone is out sick or leaves. And it must never surface in another client’s file. Personal memory does neither well. Memory per file, with access per file, comes closest to both, provided you also check that the separation holds. The Dutch version of this question, including the guidance of the Dutch accounting bodies, is on AI-geheugen per klantdossier; for choosing a tool, see which AI solution fits an accounting firm.
Where Bombos fits
Bombos is a plugin that gives every AI tool the firm uses, from Claude and ChatGPT to Copilot, one shared memory that belongs to the firm instead of the employee. Knowledge, learned procedures and what already sits in mail and documents can apply within their own context, such as one client file, and every answer shows what it rests on. It runs in our cloud or on your own server, and one staff member can try it first with their own agents.
Related questions
Can a partner see what a staff member's AI remembers?
In ChatGPT Business, memories are tied to each account and cannot be transferred to other users. In Claude Team and Enterprise, owners decide whether memory is available but members manage their own. Neither gives the firm a place to review memory per client.
Is an incognito chat safe for client data?
In Claude, an incognito chat is not used for memory or chat search, but on Team and Enterprise it is included in standard data exports and follows the organization's retention policy. It avoids memory; it does not keep data out of the system.
What happens to the file when the engagement ends?
Remove people from the project and delete it according to your retention policy. In Claude, archiving keeps all members and knowledge, and an archived project must be unarchived before it can be deleted.
Can we share one project with the whole firm?
Yes, for firm-wide material such as checklists and house style. In Claude, choose Share with your broader organization; in ChatGPT Business, share via a workspace link. Keep client files out of those.
