To brief an AI agent on sales work — an account research pass or weekly signal sweep — you need something the chat window never asked for: an instruction that survives your absence. A prompt assumes you are there to redirect it. A brief assumes nobody is. Here is a four-part structure, two prompts for ChatGPT, Claude or Copilot, and a way to judge output you did not watch.
Microsoft announced a rebuilt Copilot with Home, Code and Autopilot on 25 September 2026, and Autopilot's premise is work that carries on while you are away from your computer (our write-up). The tooling is arriving faster than the habit.
A brief is not a longer prompt
Most sales prompts are written for a conversation: you ask, read, correct, ask again. That loop is where you supply the judgement the prompt left out. Take it away and the missing judgement becomes the output. An agent told to "research this account" will research something; whether it stopped at the right depth or invented a funding round to fill a gap, you find out while reading 900 words you did not commission. If you have read what agents actually do and decided where delegation helps and where it creates risk, this is the next piece: the writing.
The four parts of a working brief
1. Scope, stated as sources. Not "research Acme" but "read Acme's careers page and their last two quarterly releases." Naming sources stops an agent wandering into a press round-up and treating it as fact.
2. The artefact, described as a shape. "Six bullets, newest first, one sentence each, each with the URL it came from" is checkable. "A summary" is not.
3. A stop rule. Say what to do when evidence is thin, because the default is to fill the space. "If you cannot find a dated source, write NOT FOUND instead of estimating" turns a silent failure into a visible one.
4. The acceptance test. One line naming what would make you reject the output, and what you read against tomorrow morning.
Prompt 1: the standing account watch
You are preparing a weekly account watch for a B2B seller. Run this unattended. SCOPE - look only at these, for each account listed below: - the company's own newsroom or press page - the company's careers page, roles in [DEPARTMENT] PRODUCE - one block per account, in this exact shape: ACCOUNT: [name] SIGNAL: one sentence, what changed DATE: the date on the source, not today's date SOURCE: the full URL SO WHAT: one sentence, what it changes for a seller here CONFIDENCE: HIGH if the source is the company's own page STOP RULE - if an account has no dated change in the last 7 days, write "ACCOUNT: [name] - NO CHANGE" and move on. Do not substitute older news, industry trends, or anything you inferred. Do not estimate a date. ACCEPTANCE - reject the whole output if any SOURCE URL does not resolve, any DATE is missing, or any SIGNAL is about the industry rather than the named company. ACCOUNTS: [list 10] DEPARTMENT: [e.g. Revenue Operations]
That ratio of refusals to asks is what separates a brief from a prompt, and it is the discipline behind reusable sales skills.
Prompt 2: the review pass
Below is output produced by an AI agent working unattended against a brief. I did not watch it run. Audit it against the brief, not your own view of the topic. Label every claim as one of: SOURCED - a URL is given and the claim matches it UNSOURCED - a specific number, date, name or role, no URL OUT OF SCOPE - outside the sources the brief named Then answer in one line each: 1. Which instruction in the brief was ignored? 2. Which single claim would embarrass me most if it were wrong, and why that one? 3. What would this output have looked like if the agent had found nothing at all? Does it look different from that? BRIEF: [paste the brief] OUTPUT: [paste the output]
Question three earns its place: unattended work fails quietly, and an agent that found nothing produces something that reads like an agent that found plenty.
Where to run this today
You do not need Autopilot to start. ChatGPT runs scheduled tasks now, with active-task limits by plan: 3 on Free and Go, 5 on Plus, 10 on Business and Edu, 15 on Pro and Enterprise, and paid plans can set hourly schedules and exact delivery times (help.openai.com, checked 29 September 2026). Ten accounts, one weekly run, one review pass is a complete loop, and Promptifi's scheduled prompts run on that cadence.
Common mistakes
Writing it as a conversation. If it contains "then ask me" or "let me know if", nobody is there to answer.
Omitting the stop rule. Without it an agent fills gaps with plausible material and you carry it into a call.
Reviewing inside the conversation that produced the work. The model defends what it just wrote. Open a new one and paste both in.
Scheduling before verifying once. A recurring task with a flaw produces a flawed artefact every week until someone reads one.
Frequently asked questions
What is the difference between a prompt and an agent brief?
A prompt is written for a conversation you are present for, so it leaves judgement to the back-and-forth. A brief is written for work that happens without you, so it states the sources, the shape of the output, what to do when evidence runs out, and the test you will judge it against.
How long should an AI agent brief be?
Usually 150 to 300 words for a sales task — long enough to name its sources, output shape, stop rule and acceptance test. The test is whether a competent stranger could execute it and produce something you would accept.
How do I check work an AI agent did while I was away?
Audit it in a fresh conversation against the brief, not your own view of the topic. Label every claim as sourced, unsourced or out of scope, identify which instruction was ignored, and ask what the output would have looked like if the agent had found nothing — then check whether what you have looks different.
Start with one brief
Pick the task you would least like to do on a Monday morning, write the four parts, and run it by hand once before it goes near a schedule. For a starting point, get a free sales prompt built for your role and rewrite it as a brief.