Guide
How do permissions work in an AI workspace?
Published September 2026. About a 5 minute read.
A permission model for an AI at work has to decide six things: who's allowed to talk to it, what company data it can see, what it can do on its own versus what needs a person's approval, who gets to read what it answered, whether someone outside the company can use it safely, and how any of that gets checked after the fact. Most products answer some of these and stay quiet on the rest. Here's how Glitch Team Zone answers every one of them, with the mechanism behind each answer, not just the promise.
Glitch is live in the chat on our home page. Say what your business needs and it starts working, no signup.
Talk to GlitchWho can talk to the AI in a workspace?
Every seat in a company isn't the same seat, and a permission model for an AI at work should say so plainly: who can ask it things, who can change what it's connected to, and who's just reading along. A product that treats every person the same is really saying it hasn't thought about this yet.
Glitch Team Zone has five roles: owner, admin, operator, viewer, and guest. Owners and admins can approve actions, connect accounts, and set the AI's own spend cap. Operators do the day-to-day asking and running of tools. Viewers can read and ask questions but can't change credentials or connections. Guests are scoped even tighter, covered next.
What data can the AI see?
The bigger risk with a shared company AI usually isn't a stranger breaking in, it's the AI itself quietly mixing up whose data is whose, especially once a product serves more than one customer on the same system.
Every table in Glitch Team Zone carries an organization id, and Postgres row-level security enforces it on every query, gated by a session variable each request has to set. Every query helper requires an org context and refuses to run without one. That's enforced at the database level, not by application code that has to remember to filter, which is the difference between a rule and a habit.
Can a client or contractor use the AI without seeing our data?
Bringing an outside person into a work chat is usually all or nothing: leave them out and updates arrive secondhand, or add them to a real channel and they can now scroll back through everything that channel has ever carried.
Glitch Team Zone's guest channels are restricted rooms. A guest sees only the room they were added to and nothing else in the workspace, no other channels, no direct messages to individual teammates. The AI answering inside that room does so without the company's shared memory or snapshot at all, so even a well-phrased question can't surface something the room was never meant to expose. That's two separate locks, the room boundary and the AI's own context, and both hold independently.
Try it on your business
Paste your website address. Glitch reads it and shows you where it'd start, in about a minute, before you sign up for anything.
No signup, no card. Nothing gets connected until you say so.
Does the AI act on its own?
Not everything an AI does deserves the same level of trust. Reading and drafting are low stakes. Anything that spends money, changes a record, or goes out the door under the company's name is a different category, and a permission model should treat it that way.
Reads, drafts, and other low-stakes work run without waiting. Writes that matter go through an approvals tray instead: the AI shows its plan, and an owner or admin decides before anything happens. Once something finishes, it lands on a receipts board with the proof attached, the link, the email, the invoice, so checking what actually happened takes a minute instead of a morning.
Who can see what the AI answered, and is it recorded?
An answer that only the person who asked can see doesn't teach the rest of the team anything, and an action nobody can trace back to a cause is hard to trust even when it was the right call.
In Glitch Team Zone, answers land in the channel where they were asked, visible to everyone in the room, so a question asked once teaches the whole team, and a correction sticks for the next person too. Separately, the audit log records the consequential actions across the company, by whom and when, including changes the AI itself made. It's a real record, not a summary, and admins can read it in full.
What keeps the AI's own spending under control?
An AI that can run tools and take actions on a company's behalf is also an AI that can run up a bill, and a permission model should cap that the same way it caps everything else the AI can do.
Every plan includes a monthly amount of Glitch's work, and anything past that passes through at exactly what it costs us, under a monthly cap an owner or admin sets, nobody else. A team that would rather hold that relationship directly can connect its own Anthropic, OpenAI, or Gemini key instead, which removes the limit entirely. A connected key is encrypted at rest with a key derived separately for that org, and it never appears in logs, error messages, or audit entries. Only a fingerprint of its last few characters shows up anywhere, for safe reference.
Common questions
Can I control which data the AI can see?
Yes, at two levels. Every company's data is isolated from every other company's at the database level through row-level security, so there's no cross-company visibility to control in the first place. Within your own org, guest channels control it further: a guest room's AI answers with no company data at all, only what's written in that room.
Can a client or contractor use the AI without seeing our data?
Yes. Invite them into a guest channel, a restricted room where they see only that one room and nothing else in the workspace. The AI answering inside it doesn't draw on the company's shared memory or snapshot, so even a well-phrased question from a guest can't surface something from outside the room.
Does the AI act on its own?
For low-stakes work like reading and drafting, yes, without waiting. For anything that spends money or changes something that matters, no, it goes into an approvals tray first, and an owner or admin decides before it happens.
Who can see what the AI answered?
Whoever is in the channel where it answered. Answers land in the shared room, not a private history, so the whole team sees the same answer and the same later correction.
Can another company's AI see our data?
No. Every domain table carries an organization id and Postgres row-level security enforces it on every query, gated by a session variable scoped to one org at a time. There's no cross-org view in the product, and every query helper requires an org context to run at all.
Who can change permissions?
Owners and admins. They're the roles that can approve actions, connect accounts, invite guests into restricted rooms, and set the AI's own monthly spend cap. Operators and viewers can use the AI within the access they're given, but can't change what it's connected to or who else can reach it.
Keep reading
See it working, not described
Try it with your own website
Paste your website address. Glitch reads it and shows you where it'd start, in about a minute, before you sign up for anything.
No signup, no card. Nothing gets connected until you say so.