Back to Blog

Write the Brief Before You Build the Agent: Context Design Skills for Business Professionals

By | Published | Updated | 8 min read

Person facing four holographic work-loop screens linked by glowing arrows on a dark reflective platform

Open almost any team's shared drive and you will find an SOP that nobody quite follows. It was written two years ago, updated in a hurry once, and now carries a few notes in red that say "check with Finance first" or "ignore step 4 for overseas customers". The people who do the work don't need it. They learned the job by sitting next to someone.

An AI agent can't sit next to anyone. It may draw on general knowledge, its own default behaviour and whatever tools or files it can reach, but none of that tells it how your team does this job. What you write down is the part you control. So when someone says "let's get an agent to handle this", the first useful piece of work is not choosing a tool. It is writing a brief the agent can act on.

That is what we mean by context design. It sounds technical, but it is a skill most business professionals already half-have: explaining a job clearly to someone new. This post shows how to turn a messy SOP into a working brief, with an illustrative example you can copy.

Why the SOP isn't enough

An SOP tells an experienced person what order to do things in. It assumes they already know why the work matters, where the information lives, which customers are sensitive and when to stop and ask.

An agent has none of that. Give it the SOP as it stands and it may read a step too literally, misread an ambiguous one, fill the gaps with its own assumptions, or carry on past the point where a person would have picked up the phone. A clear brief reduces those risks. It doesn't remove them, which is why supervision still matters.

The fix is to make the hidden parts of the job visible. Six headings cover most of it.

The six parts of an agent brief

  1. Goal. One or two sentences on the outcome, written as a result rather than a list of steps. "The customer gets a correct, polite decision on their refund request within the agreed time" is a goal. "Process refund requests" is a task label.
  2. Inputs. What the agent receives, and where it is allowed to look. Name the actual sources: the request form, the order system, the refund policy page. If a source is out of bounds, say so.
  3. Rules. The decisions that are already settled. Thresholds, eligibility, tone, what counts as proof. These are the things a new hire would be told on day one.
  4. Human approval gates. The points where the agent stops and a named person decides. Be specific about who approves and what they see when they do.
  5. Exceptions. The cases that break the normal flow and what to do with each. Most real work lives here, and most SOPs skip it.
  6. Definition of done. How anyone can tell the job is finished and done properly. What gets updated, what gets recorded, and what the customer or colleague receives.

If you can fill in all six on one page, you understand the work well enough to hand part of it to an agent. If you can't, that gap is useful information too. It usually means the process lives in someone's head.

A worked example (illustrative)

This example is made up to show the method. It isn't a client story, and the numbers are placeholders you would replace with your own policy.

Here is the kind of SOP many customer service teams have:

Refund requests

1. Check the email.

2. Look up the order.

3. If it's within 30 days, refund it.

4. Big ones go to Sarah.

5. Reply to the customer.

Note: damaged items, ask for photos. Don't refund marketplace orders here.

It works for the two people who wrote it. Now look at what is missing. What counts as "big"? Who covers when Sarah is on leave? What happens if the photos never arrive? Is "30 days" from the order date or the delivery date? Where does the refund get recorded?

Here is the same job written as a brief.

Goal. Every refund request sent to the support inbox gets a correct decision and a clear reply within two working days, and every refund issued is recorded against the order.

Inputs. The customer's email and any attachments. The order record in the order system (read only). The current refund policy page. The agent may not use payment card details or anything outside these sources.

Rules.

  • The refund window is 30 days from the delivery date shown on the order record.
  • Requests inside the window, for orders up to the team's set limit, with no exception flags, can be prepared for refund.
  • Replies use the approved refund templates and the customer's first name, and never promise a date the payment provider controls.

Human approval gates.

  • The agent never issues money itself. It prepares the refund and a short summary, and the duty customer service lead approves or rejects it.
  • Any refund above the set limit goes to the finance lead, not the duty lead.
  • If the named approver is away, the request waits in the approval queue. It doesn't move to someone else by default.

Exceptions.

  • Damaged item: ask for photos using the photo template. If none arrive within five working days, send one reminder, then pass the case to the duty lead.
  • Marketplace order: don't process. Reply with the marketplace's own refund route and log the case.
  • Customer is angry, mentions legal action, or has written in three times: stop and hand the whole conversation to the duty lead.
  • Anything the agent can't match to a rule: stop, record what it found, and ask.

Definition of done. The decision is approved by the right person, the customer has the reply, the refund (if any) is recorded on the order with the approver's name, and the case is closed in the inbox with a one-line reason.

Notice what changed. The steps barely moved. What we added was the judgement that used to sit with Sarah: the limit, the timing rule, who decides, and when to stop. That is the part that makes an agent safe to supervise.

How to write your first brief

You don't need a technical team to start. You need one real SOP and an hour with the person who does the work.

  • Pick one workflow that repeats. Weekly or daily work gives you enough examples to check the brief against.
  • Interview the doer, not the manager. Ask "what do you do when this goes wrong?" The answers become your exceptions list.
  • Write the goal last. Once you've listed inputs, rules and exceptions, the real outcome is easier to see.
  • Make every approval gate name a role. "A manager checks it" isn't a gate. "The duty lead approves before any money moves" is.
  • Test the brief on old cases. Take ten past requests and walk each one through the brief on paper. Any case that makes you say "it depends" needs a new rule or a new exception.

That last step matters more than it looks. A case that trips up the brief on paper would probably have tripped it up in front of a customer. Passing your past cases is a useful check, but it doesn't prove the agent is safe or ready; new cases will turn up that your history didn't include. The next step is a controlled trial: a narrow scope, a person supervising the output, and explicit approval before anything reaches a customer or moves money.

Where this fits with problem framing

If you read our earlier post on problem framing for AI agents, this is the next step. Problem framing helps you decide which piece of work is worth handing over. The brief is how you hand it over without losing the judgement your team built up.

Both are business skills first. The people best placed to write a good agent brief are usually the ones who know the work, the exceptions and the customers. They don't need to write code. They need to write down what they already know, clearly enough that someone new could act on it and a supervisor could check the result.

Start with one SOP this week. Rewrite it under the six headings. You'll find out quickly whether your process is ready for an agent, and you'll probably end up with a better SOP for your people too.

Sources

About the Trainer

Melverick Ng is Founder of Nexius Labs and Master Trainer at Nexius Academy. He has trained business teams and non-technical professionals to design practical AI workflows for sales, operations, and customer support.

Talk to a Course Advisor