Who should see your team's AI chats? Permissions for shared AI workspaces

A shared AI workspace is useful because the AI remembers what the team tells it. That's also why permissions matter more than they did in a chat app. The AI's memory, the files it reads and the summaries it writes all carry what was said - so "who can see this channel" quietly becomes "who can see everything the AI learned in it".

Here's what a shared AI workspace should give you, and the questions to ask before your team moves in.

Roles that match how teams work

Most teams need four:

• Owner: runs the workspace and its billing.
• Admin: manages people, settings and connected tools day to day.
• Member: does the work - talks to the agents, uploads files, starts tasks.
• Viewer: reads, never writes. Useful for clients, mentors, judges, managers or stakeholders who need to follow along without accidentally changing anything.

A viewer that can still trigger the AI is not really read-only. Check that viewers can't start tasks, approve actions or delete messages - and that the server refuses those, not just the interface.

Private channels - including what the AI remembers

A private channel should be private all the way down:

• its messages,
• its files,
• the AI's memory and summaries from that channel,
• background tasks and their results,
• search results and notifications.

If the AI's memory from a private channel can surface in an open one, the channel isn't private. Ask how the tool scopes memory, not just messages.

Approval before the AI changes anything

When an AI agent can act in your tools - create issues, edit files, update a spreadsheet - the safe default is that a person approves every change before it happens. Reading can be automatic; writing shouldn't be.

An audit log

Someone will eventually ask "who removed them?" or "when did this channel go private?". A good workspace keeps an activity log for owners and admins: people added and removed, role changes, channel access, shared keys, exports. It should be written by the server, so nobody can edit it from their browser.

Data you can take with you

• Export: owners or admins should be able to download the whole project - messages, memory, files.
• Encryption: API keys and AI memory should be encrypted at rest.
• No training on your data: check the policy of both the workspace and every AI provider it sends your messages to.

A five-minute checklist

• Is there a true read-only role, enforced on the server?
• Does a private channel hide the AI's memory and files too?
• Does the AI ask before changing anything in connected tools?
• Is there an activity log owners and admins can read?
• Can you export everything and leave?

How Teamski handles it

Teamski has owners, admins, members and viewers. Viewers can read every open channel and any private channel they're added to, but can't post, ask the agent, upload, start tasks or approve anything - and the database refuses those even if someone tries. A private channel's messages, files, memory, agents and tasks are visible only to the people on it, plus owners and admins. Agents ask before changing anything in a connected tool, owners and admins see an activity log in Settings, and a project can be exported at any time. Saved keys and the agent's memory are encrypted at rest. See also AI agents for teams.