Why it matters
A new employee doesn't get to pay invoices on day one. An agent shouldn't either. Agents are fast and tireless, which is the point, but it also means a mistake can be repeated 200 times before anyone looks.
Zero trust means the agent gets the least access the task needs, and the important decisions stay with someone who is accountable for them.
How it works
Every action an agent can take falls into one of three groups. You can change them per agent.
Reading and drafting
Reading invoices, tickets and documents. Drafting replies, reports and bookings.
Anything with impact
Emailing customers, posting in channels, booking, paying, or changing records. Deleting always asks.
Off-limits
Reaching sites off the list, seeing your real keys, or changing its own rules.
There's one more rule that doesn't depend on settings. If an agent has read something from outside, has access to private data, and now wants to send something out, it always stops and asks. That combination is how data gets stolen.
An example
Nine payment reminders
The finance agent finds 11 overdue invoices, skips 2 that are in dispute, and drafts 9 reminders. It checks every amount against Exact Online. Then it stops: "Send 9 payment reminders?" You see the 9 emails. Approve and they go. Deny and nothing is sent or changed.
What you control
- Which actions each agent may take alone, which need your OK, and which are blocked.
- Who on your team may approve what.
- A budget per agent. When it's reached, the agent pauses and tells you.
For your security team
- Approval requests show the reason and the exact content, not a summary written by the agent.
- A denied action leaves no side effects. The agent records the answer and moves on or stops.
- Agents can't change their own rules, budgets or allowed sites.
The code is open source, so your team can read how each of these works. Questions? Email security@upmo.ai.