Guide / AI Systems

How to write an AI agent brief that actually works.

A vague brief produces a vague agent. Here's the seven-part structure we ask for before scoping any custom AI agent project — use it whether you're briefing us or your own team.

The template

Seven things a real brief answers.

01

The specific job, not the general idea

"Help with customer support" isn't a brief — it's a category. "Answer pre-sales questions about pricing tiers and features, using our current docs" is. Name the exact job, not the department it lives in.

02

The inputs it will actually see

What triggers it, and what does the raw input look like — an email, a form submission, a webhook payload? Include a few real examples, especially the messy ones, not just the clean happy-path case.

03

The actions it's allowed to take

List the specific, bounded actions — send a reply, update a record, tag a ticket — not "handle the request." An agent needs a defined set of moves, not an open-ended mandate.

04

What "done" looks like

A concrete, checkable definition of success for a single run. If you can't describe what a correct outcome looks like, the brief isn't ready yet — that's a sign the process itself needs to settle first.

05

Where it should hand off, and to whom

The specific conditions that should trigger a handoff to a person, who that person is, and what context they need to pick up the case without re-investigating it from scratch.

06

What it should never do

The explicit boundaries — actions that are always off-limits regardless of how confident the agent is. This matters more than the list of what it can do.

07

How you'll know it's working

The specific signal you'll check after launch — error rate against real inputs, time saved, escalation volume — not a vague sense that it "seems fine."

Example

What this looks like filled in.

An illustrative example, not a real engagement — showing the level of specificity a working brief actually needs.

Job: Triage inbound support tickets tagged "billing" and resolve the ones matching a defined set of known issue types.

Inputs: New Zendesk ticket, subject + body text, customer account ID.

Actions allowed: Look up account status via internal API, reply with a templated resolution for known issue types, apply a resolved tag.

Done looks like: Ticket resolved with an accurate reply, or escalated with a clear reason if it doesn't match a known type.

Escalate to: Billing support queue, with the ticket, account lookup result and the reason it didn't match, attached.

Never: Issue a refund or credit — that stays a human decision regardless of confidence.

Success signal: Resolution accuracy against a sample of real historical tickets, reviewed weekly for the first month.

Before you write one

Two things worth checking first.

Not sure if your process is ready for this level of detail yet? Read how to tell if your workflow is ready first — a brief is only useful once the process underneath it has settled.

Bring us the draft

Even a rough version is a good starting point.

You don't need every section filled in perfectly before reaching out — we'll help sharpen it in the first conversation.

One good conversation

Bring the messy version. We’ll find the signal.

Tell us where work gets stuck. We’ll come back with a sharper view of what to automate, what to keep human, and what to leave alone.