How to Brief an AI Agent So It Actually Does What You Want

If your AI agent is producing garbage, the problem is almost never the model. It's your brief.

We've deployed agents for law firms, logistics operators, e-commerce brands, and consultancies across Lagos, Abuja, and Nairobi. In every single case where an agent underperformed, we traced it back to the same thing: the person setting it up wrote a brief they would never accept from a human hire.

You wouldn't hand a new employee a Post-it note that says "handle customer emails" and walk away. But that's exactly how most people brief their AI agents. Then they blame the AI when it goes sideways.

The mental model most people get wrong

People treat AI agents like search engines. You type a question, you get an answer. So they write briefs the same way: short, vague, and full of assumptions the agent has no way to verify.

An AI agent is not a search engine. It's a junior staff member with perfect memory, zero context about your business, and no ability to ask for clarification unless you tell it to. It will do exactly what you say, at speed, including the wrong things.

The shift you need to make: stop writing prompts. Start writing job descriptions.

The five things every agent brief needs

A brief that actually works has five parts. Miss any of them and the output degrades in predictable ways.

1. Role and scope. Who is the agent, and what is it responsible for? "You are the first-response agent for inbound WhatsApp enquiries at a real estate agency in Lekki. You handle enquiries about listings, viewings, and pricing. You do not handle legal questions, negotiations, or payment issues — those get escalated."

2. The user it's serving. "Most enquiries come from Nigerians aged 28-45 looking for 2 or 3-bedroom apartments to rent between ₦4M and ₦12M annually. Many are relocating from Abuja or abroad. They usually ask about location, security, and payment terms first."

3. What good output looks like. Show it. Paste in three or four examples of ideal responses. Not descriptions of ideal responses — actual, copy-pasted good replies. This one thing will improve output quality more than anything else you do.

4. Hard rules. What it must never do. "Never quote a price that isn't in the listings database. Never promise a viewing time without checking the calendar. Never claim a property is available without confirming. If unsure, say so and escalate to a human."

5. Escalation triggers. When should it stop and hand off? Be specific. "Escalate to a human when: the client mentions legal issues, mentions a competitor by name, uses aggressive language, asks for a discount above 5%, or when the conversation goes past 6 messages without a clear next step."

Skip any of these and you'll get an agent that hallucinates prices, promises things you can't deliver, or dumps every conversation on your team anyway.

The examples problem

Most briefs fail on point three. People describe what they want in adjectives — "professional but friendly", "concise", "helpful" — and expect the agent to figure it out.

Adjectives are useless. They mean different things to different people, and the agent will pick an interpretation that may not match yours.

Show, don't describe. If you want responses that sound a certain way, paste in five real examples of that style. If you want responses to follow a specific structure, show the structure. If there's a phrase you always use to open messages, put it in the brief verbatim.

One of our clients cut their agent's error rate by roughly 70% just by replacing "respond in a professional tone" with six pasted examples of their best sales rep's actual replies. Nothing else changed.

The context most people forget

An AI agent doesn't know your business. It doesn't know that your delivery cutoff is 3pm, that you don't service Ajah on Sundays, that your MD hates the word "kindly", or that returns require a receipt within 14 days.

You have to tell it. All of it. Every rule you've built up in your head over years of running the business needs to end up in the brief, or it won't be applied.

A good exercise: shadow yourself for a day. Every time you make a decision about how to respond to a customer, write down the rule you just applied. By the end of the week you'll have 30 to 50 rules that were living in your head. That list is the missing half of your brief.

Test it like you'd test a hire

Nobody hires a person and lets them talk to real customers on day one. But that's what people do with agents. They set them up, connect them to WhatsApp or email, and turn them loose.

Instead: write 20 test scenarios before you go live. Include the easy ones, the edge cases, the customers who send one-word messages, the ones who write essays, the ones who are angry, the ones who are trying to negotiate. Run each scenario through the agent and check the output.

You'll find gaps. Fix them in the brief. Run again. Do this for two or three rounds before a single real customer ever talks to it.

This process takes maybe six hours. It saves you the far larger cost of an agent embarrassing your brand for two weeks before you notice.

The bottom line

AI agents are only as good as the briefs behind them. If you treat the brief like a quick prompt, you'll get quick-prompt results. If you treat it like the job description for a new hire — role, scope, examples, rules, escalation, context — you'll get an agent that actually earns its place in your operation.

The good news: this is a skill, not a talent. Once you've written one proper agent brief, the next ten get easier. The businesses pulling ahead right now aren't the ones with better AI. They're the ones with better briefs.

---

If you want to see how this applies to your business, book a 30-minute audit with LVD Labs. We'll look at your current setup and tell you exactly where AI and automation can make a measurable difference.