What the assistant can access
Exactly what you can access, and nothing more. When the assistant looks up data or performs an action, it does so as you. Clarus checks every request against your own account’s permissions, just as it does when you click through the interface. The assistant can’t grant itself extra access, and a request for anything outside your permissions is refused.Your sign-in credentials are never shared with the AI model. The assistant works within your permissions, but the model itself never sees or holds your login — so nothing it’s asked to do can make it act as another user or reach another customer’s data.
Actions the assistant takes
When the assistant does something — placing stock on hold, updating an order — it creates a normal Clarus transaction, with the same validation and the same full audit record as the equivalent action from the standard interface. There is no separate, less-audited path for changes made through the assistant. Every action is attributed:- to you, when it comes from your session, or
- to the AI agent account, when it comes from an automation (see below).
- Enforced controls — the assistant cannot perform an operation you don’t have permission for. This is a hard boundary, checked by Clarus on every request.
- Guidance controls — you can give the assistant custom instructions for your subdomain, such as asking it to confirm before certain operations. This shapes how it behaves in normal use.
Automations and the AI agent account
Some AI actions are triggered by automations rather than a person — so there’s no signed-in user behind them. These run under a dedicated AI agent account, whose permissions you control exactly like any other user account. Whatever you allow that account to do is the limit of what your automations can do. Two things worth knowing:- The agent account is shared across all AI automations on your subdomain, so automated actions are attributed to that account rather than to an individual person.
- Automations run unattended, with no person reviewing each action as it happens.
How your data stays protected when the assistant reads text
To answer questions the assistant sometimes reads free-text that other people or systems have written into your records — order notes, inbound documents, reference fields. Several controls keep that safe:- Permissions are enforced by Clarus, not by the assistant. Whatever is written in a record, the assistant still can only reach data and actions your account is allowed — so nothing in a record can widen its access.
- Your credentials are never in reach of the model, so nothing can make the assistant authenticate as someone else or cross into another customer’s data.
- Every action is audited and reversible through normal Clarus transaction history, so anything unexpected is traceable and can be undone.
- Least privilege on automation accounts keeps the reach of unattended actions to the minimum you’ve allowed.