Skip to article

Setting Up an AI Agent for Your Team: What to Write Down Before It Starts

A written brief comes first. The platform is where the brief runs.

By Jacqueline V. TwillieNovember 20, 20264 min readGuide
Quick answer

Write four things before anyone builds: the AI agent's job, the background files it can read, the tools it can use, and the points where it must stop for a person's sign-off. It drafts, you decide what goes out. Test it on real scenarios before it touches real work, and name its owner.

A team's AI agent behaves the way its brief reads. If the brief is a single sentence, the output looks like the work of a system that was never told the job. The writing happens before the build.

The brief comes before the build

BNEDai's build method starts with a piece of writing. Every setting on the platform's screen is a decision that has to come from somewhere, and the brief is where it comes from. The platform is where the brief runs. Whichever platform your team uses, an AI agent runs on the same parts: a clear job, the right background files, access to specific tools, and rules on when a person signs off.

The brief has four parts

The job. What this AI agent does, in order, and why each step is there. Write it as one job with one outcome you can check. "Prepare the brief before every external meeting" is a job. "Help with meetings" is a wish. Write into it who the AI agent is, what it does, what it reads, what it never does, and how it takes a correction.

The background files. Exactly what it can read, and what it can never touch. Name each file and each connection.

The tools. The specific tools it can use, and the step-by-step workflows it follows, which BNEDai's published pages call skills. An AI agent that only drafts doesn't need permission to send, so scope each tool to the job and nothing wider. Where the tool lets you separate read access from write and delete access, decide each one on purpose. Each workflow should be something you can teach and something you can test.

The stopping points. Where it must stop for a person's sign-off. The rule is short: it drafts, you decide what goes out. Add a never-do list, with each line next to the job it protects. A never line that carries its reason, such as "never guess a number, because a wrong figure reaches a customer", is easier for the AI agent to apply to a case you didn't list.

A shared AI agent needs three more entries

A shared AI agent needs three entries that a personal one skips.

Name an owner. If the answer to who owns this AI agent is a team or a shrug, nobody is reading its output.

Name a person to ask. When output looks wrong, the team needs a name written where the work happens. The Index asks the same question: is there a written who-to-ask path when AI output looks wrong?

Give it a second person who can edit it. An AI agent's triggers run on the connected accounts of the person who created them and stop when that person loses access, so before the builder moves on, a second person should hold edit and access rights. Keep an AI Agent Running After the Person Who Built It Leaves covers the steps.

Test it before it touches real work

Run it on a real scenario you already know the answer to, and read the whole output. Then tell the AI agent what you think and why, instead of saying try again. How Do You Test an AI Agent and Give It Feedback? goes through that step.

Bring it to autonomy in order. Test by hand, then run it watched with a person reviewing every result, then schedule it for the cadence the watched runs justified, then book a date to re-read its output. Most teams skip the watched stage, and it is the one that catches edge cases while they are cheap.

Put the next reading of the brief on the calendar

On a calendar reminder. If the AI agent can update its own instructions, someone has to read them. The review is 15 minutes: the instructions, each skill, the connections and the tool settings, with three questions: what changed since last time, what can it still read, and what can it send. What Do I Read in the 15-Minute Review of My AI Agent? has the order.

Check where your team stands against these rules with the free AI Readiness Index. Its questions cover the same ground: a written goal, a yes before anything sends, scoped access, a test before real data and a named person to ask.

Frequently asked questions

The brief has four parts

A clear description of its job, the background files it can read, the specific tools it can use, and the points where it must stop for a person's sign-off. Then you test the AI agent with real scenarios before it touches real work.

Who should own a team's AI agent?

A named person, with a second person who can edit it, so that its triggers and access don't stop when the builder leaves.

How often should the instructions be read again?

Put a recurring reminder on the calendar. BNEDai's example is every three weeks, and every two weeks for an AI agent used daily.

Reviewed on November 20, 2026.

More on Ownership for Non-Technical Managers