AI for teams

AI team permissions: who controls what your AI can do

AI team permissions decide who can use an AI tool, what it may do for each team, and when it has to stop and ask a person. They matter more once AI can act, because an assistant that runs code or sends email can do more than give a wrong answer.

The four questions to settle

QuestionWhat it controls
Who gets in?Accounts, invitations and administrators
What may the AI do for this team?Which tools it may use: documents, code, email and so on
When must it ask first?Which actions wait for a person's approval
Who is answerable?Who can approve, and how you see what happened

Many products focus on the first. Controlling who can reach an AI is a separate matter from controlling what it can do, so check which of these a product really covers.

Why user roles are not enough

Traditional software does what the person using it could do anyway. An AI in a team chat is different: several people can ask it for things, each with different responsibilities. It also reads documents and web pages, and text in those can try to steer it. So the AI needs limits of its own, not just those of whoever is asking.

How to set them up

  1. List what the AI can reach, tool by tool, for each team.
  2. Turn off what a team does not need. It is easier to open up later than to clean up afterwards.
  3. Sort actions by how easily they can be undone. Reading is low risk. Sending something outside the company cannot be taken back.
  4. Make hard-to-undo actions wait for a person, with no way around it.
  5. Keep a record of what the AI did and who changed its permissions.

How SyncMind does it

These controls come from SyncMind's API edition.

What SyncMind may do for each team

Administrators have one switch per team for each of these. Every switch starts on until an administrator changes it, so review them when you set up a team.

A change applies on every computer at once, and shows who made it and when. In a chat shared by several teams, SyncMind may use a tool only if every one of them allows it. Web search has no switch, but the levels below control when it asks first.

When it asks first

LevelWhat SyncMind does
SupervisedShows its plan and waits for approval, then asks again before running code or using the web
Balanced (default)Searches and reads freely; asks before running code or using the web
IndependentUses every tool that is switched on without asking, apart from the actions below

A team's manager sets the level for its chats, and each person sets their own for personal chats. A chat held by several people or teams follows the strictest of them.

What always waits for a person

Sending email, deleting email and changing code on GitHub ask for approval at every level, including Independent, and can never be set to always allowed. A sent email cannot be taken back, and an email can try to talk an AI into sending one.

Who approves, and how you can check

Whoever asked can approve. In a team chat, so can the team's manager. The Agent Log shows every step as it happens.

This is part of how SyncMind does multiplayer AI. If nothing can leave your building, the on-premise AI assistant has an administrator account and invite codes, with fewer of these controls listed today. The pricing page compares them.

Questions

What are AI team permissions?

They are the rules for who can use an AI tool, what it may do for each team, and when it has to ask a person first.

Are user roles enough to control AI?

Usually not. Roles control what people can do. An AI that acts in a shared chat also needs its own limits, and rules for when it must wait for approval.

Can SyncMind send an email without asking?

No. Sending and deleting email always wait for approval, at every level, and can never be set to always allowed.

Which SyncMind edition has these controls?

Tool settings for each team are listed for the API edition. Both editions have an administrator account and invite codes.

Last updated