On this page
An agent that only answers questions can get away with being wrong now and then. An agent that books invoices, updates records or replies to customers can't. Once software does the work start to finish, the question changes from "is it clever?" to "who's watching, and what happens when it's unsure?"
Decide what it may never do
Start with the limits, not the abilities. Write down what the agent may read, what it may change, and what it must never touch. Some lines are absolute: moving money, signing contracts, sending anything outside the company without a person's say-so.
Approvals where the risk is
Not every step needs a person. Approvals belong where a mistake is costly or hard to reverse: payments, messages to customers, changes to contracts or pay. Everything else can run, as long as it's recorded.
A good rule: the newer the agent, the more it asks. Like any new hire, it starts on probation and earns more room as the team sees its work.
Handovers that make sense to people
When the agent can't finish something, like a document that doesn't match or a request outside the rules, it should hand over, not guess. A good handover says what came in, what was tried, and exactly what the person needs to decide. Quick to read, one click to act.
An audit trail you'd want to read
- Every action logged: what changed, where, and why.
- Every change reversible, so reviewing beats redoing.
- A short daily summary: what was done, and what's waiting for a person.
- A named person accountable for the agent's work.
Supervision gets lighter, not lost
In the first weeks, people review almost everything. As the agent proves itself on real work, review moves to samples and exceptions. The record stays complete; the effort to read it shrinks.
It's how we run agents in production, including on our own back office with OFFHOURS.

