Most jobs in a business were never written down. People learned them by doing them, from whoever did them before. The rules live in habits: which customer always gets a call instead of an email, which supplier's invoices need a second look, which request goes straight to the managing director. A new colleague picks these up in a few weeks by asking and watching.
An AI agent can't learn that way. An AI agent is software that does one defined job inside the business, under rules the business sets. It does what it was told and nothing more. So the first piece of work, before anything is built, is to write the job down. Doing that also shows whether the business understands the job well enough to hand it over.
This piece is for the COO who has chosen a job for AI agents and now has to describe it.
The one-page job brief
Write seven lines. If a line won't fit, the job isn't ready.
- The job in one sentence. Give a verb, what it acts on and who the work is for. For example: "Log every enquiry that reaches the sales inbox and send it to the right salesperson the same working day." If you need "and also", you have two jobs.
- What starts it and what finishes it. Name the event that starts the job, and describe "done" in a way someone can check. "Done" is never "handled". It is a record that exists, a person who has been told, or a status that has changed.
- What it works from. List the sources it reads: the inbox, the customer list, the price book. Say which source wins when two disagree. Many errors start where two lists say different things.
- Three lists: alone, ask, never. Write down what the agent may do on its own, what it must ask a person about first, and what it never does, however sure it is. Start with a short "alone" list and a clear "never" list.
- One owner. Name one person who writes the lists and is the only one who changes them. Not a department, and not a committee.
- What it leaves behind. The agent keeps a record of each thing it did, the rule it followed and who it passed the work to. Write it so the owner can read it in minutes. If you can't see what the agent did, you can't trust it with more.
- How you'll know it works. Pick two or three measures and count them before the agent starts: how long the job takes now, how often it goes wrong, and how much work ends up with nobody. Without the "before", there is no "after".
A brief, filled in
Here is the brief for one job a mid-sized business might hand to AI agents. It shows how the method works. It is not taken from a client.
- Job: Log every enquiry that reaches the sales inbox and send it to the right salesperson the same working day.
- Starts and finishes: It starts when an email arrives. It is done when the enquiry is logged with the company, what they asked for and the salesperson who owns it, and that salesperson has been told.
- Works from: The sales inbox, the customer list and the territory list. If an email signature and the customer list disagree, the customer list wins and the agent notes the difference.
- Alone: Log the enquiry, tag what it is about, and assign it to the salesperson for that territory.
- Ask first: Any email that mentions a price, a discount, a complaint or a contract. Any sender it can't match to a company or a territory.
- Never: Reply to the sender, quote, promise a date or delete an email.
- Owner: The sales manager.
- Leaves behind: One line per email: what it decided, which rule it used and who received it.
- Measures: The time from arrival to a named salesperson, and the number of enquiries still unassigned after one working day. Count both for the month before the agent goes live.
Notice what is missing: the agent sends nothing to a customer. That isn't a limit on what AI agents can do. It is where this business chose to start.
Where the method breaks
The brief describes the job as it should be. A brief written at a desk misses the exceptions. Write it with the person who does the job today, and test it against a week of real cases before anything is built.
The "ask" list grows until the agent asks about everything. Then the owner drowns in requests and starts approving without reading. An approval nobody reads is worse than none. Keep "ask" for what matters, and move routine cases to "alone" or "never".
The unwritten rules stay unwritten. Habits are hard to see from the inside. One question helps: what would get a new colleague quietly corrected in their first month? Each answer becomes a line on one of the three lists.
The brief goes stale. Jobs change: a new product, a new territory, a new system. The owner reviews the brief whenever the job changes, not once a year.
What stays with people
The owner writes the lists and is the only one who changes them. A line moves from "ask" to "alone" only when the record shows the agent has handled that kind of case well, many times. That is earned autonomy: trust that grows from evidence, one line at a time. Some lines never move. Anything that commits the business to a price, a date or a promise stays with a person.
A short brief means a clear job
A brief that fits on one page shows the business understands the job. A brief that won't fit shows where the work is still unclear, and that is worth knowing before anything is built. Write the page first. Everything the agents do later is measured against it.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets. If you have chosen a job for AI agents and want a second view on its brief, start a conversation.
