AI agent permissions: what should an agent be allowed to touch?
AI agent permissions decide what an agent may see and what it may do. An AI agent is a software helper that reads your data and carries out small jobs. A safe rule to start with: an agent never gets more access than the person it works for, and less wherever the job allows.
Think of the keys you give a new employee
A new shop assistant gets the key to the front door. They do not get the key to the safe on day one. Nobody finds that unfriendly. It is simply how trust is built.
An agent deserves the same thinking. It is fast and tireless, so a wrong action is also repeated fast. Clear limits mean a mistake stays small.
The simplest rule: the agent borrows your access
In the safest setup, the agent acts for whoever asked and sees exactly what that person sees. A salesperson's agent sees that salesperson's customers. It does not see the payroll. No new rules have to be invented for AI.
Some agents run by themselves, for example every night. Nobody watches each run, so their rights should be written down as narrowly as those of an outside contractor.
Four levels of action
- Read. The agent can look things up.
- Suggest. The agent prepares a change and a person confirms it. This is the right default for anything that alters your data.
- Change. The agent alters data by itself. Keep this for small edits that are easy to undo.
- Act outside. Sending email, paying, deleting. These cannot be taken back, so they should stay with a person.
A permission sheet you can fill in
Before an agent starts, fill in one line per kind of data. This example is for an agent that helps a sales team.
| Data | Read | Change | Delete or send |
|---|---|---|---|
| Customers and deals | Yes, those of the person asking | Suggest only | No |
| Notes on calls | Yes | May add new notes | No |
| Invoices | Amount and status only | No | No |
| Salaries and staff files | No | No | No |
| Passwords | Never | Never | Never |
If a cell is hard to fill in, the agent's job is still too vague. Sharpen the job first.
What an agent should never be able to change
- Its own rights. An agent that can edit access rules can give itself anything.
- Passwords. They are not for an AI to read.
- The record of what it did. You need to be able to check afterwards.
Check again from time to time
Jobs grow quietly. Put a reminder in the calendar every few months. Is the agent still used? Who looks after it? Has its job grown beyond the sheet? Remove what is no longer needed.
How Worknano handles this
Worknano has one set of permissions for people and agents. Roles set which modules, actions and fields a person may use, and further rules narrow which records they see. An agent reads only what the signed-in user may read. Changes go through a preview that a person confirms, and each agent's list of tools can be shortened in its settings. The Permission Manager agent can draft the access rules for you. You say who may create, read, update and delete a kind of record, and it suggests rules that give no more access than needed, using your existing roles. Nothing is saved until someone with the right to configure the workspace confirms it.
- Permissions: roles, access, and who can see what
- Permission Manager: sets who can create, read, update and delete
Create a free workspace and set up roles that your agents follow from the first day.