1. Its job, in one sentence.
"Help the operations team" is not a job. "Read incoming supplier confirmations, match them to open orders and flag the ones that don't match" is. A narrow job can be tested, measured and handed to a named person. A broad one can only be demonstrated.
Why it matters: every other line on the page depends on this one. If the job is vague, so are the limits.
2. What it may read, and what it may change.
Reading a record and changing a record call for different levels of trust. Write down which systems the agent can see, which it can write to, and which are off limits.
Why it matters: most of the risk sits in the list of things it can change. Keep that list short at the start. It can grow later.
3. Where a person says yes.
Name the decisions that always need a person: anything that commits money, changes what a customer has been promised, or can't easily be undone. The agent prepares the work. A person approves it.
Why it matters: people trust an agent sooner when they know where their own judgement still decides.
4. What it does when it isn't sure.
Every agent meets cases its rules don't cover. Decide in advance what happens then: it stops, hands the case to a named person and says why. Write down, too, who can pause the agent altogether.
Why it matters: an agent that guesses quietly is more dangerous than one that asks often.
5. What it records.
For every action, record what it was asked, what it read, what it did and who approved it, in words a person can follow.
Why it matters: when something goes wrong, the record lets you fix the cause instead of switching everything off.
6. Who owns it, and how it earns more room.
Name the person in the business who answers for the agent's work, and set a regular review. At each review, read the record: where the agent was right, where it was corrected, where it stopped. Widen what it may do only where the evidence supports it.
Why it matters: autonomy should be earned against evidence, not granted on day one or withheld forever.
Why write it down
None of the six is technical. They are what you would agree with a new member of staff: the job, the access, the sign-offs, what to do when stuck, what to keep on record and who they report to.
Think of the page as a short constitution for each agent. The principles stay fixed. The room the agent is given grows with its record. Write the page before the build starts, and the conversation about trust becomes a conversation about evidence.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets.
