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

Most AI agents don't fail because the technology is broken. They fail because nobody told them what "good" looks like.

You've probably had this experience. You set up an AI tool to handle replies, qualify leads, or draft emails. The first few outputs look impressive. Then a client gets a reply that sounds nothing like you. Or the agent confidently tells someone the wrong price. Or it answers a complaint with three paragraphs of corporate apology when one line would do.

The instinct is to blame the model. The real problem is almost always the brief. A new hire who gets one paragraph of context and no examples would produce the same mess. AI agents are the same. They need a proper brief, and most people are writing the equivalent of a two-line WhatsApp message and hoping for the best.

Stop writing prompts. Start writing job descriptions.

A prompt is a sentence. A brief is a document. If you want an agent to do real work in your business, you need the second one.

Think about what you'd give a new staff member on day one. Not just "answer customer questions." You'd tell them who the customers are, what products you sell, what the prices are, what tone to use, what's off-limits, who to escalate to, and what a good reply looks like versus a bad one. You'd probably show them five or six examples of past conversations.

That's the brief. Strip out the human pleasantries and you have exactly what an AI agent needs. The agents that work well in production almost always have a system prompt that runs to one or two pages. The agents that embarrass their owners have three sentences.

The five things every brief needs

After enough deployments you start to see the same gaps. Here's what a working brief contains.

1. Identity and scope. Who is the agent? What business does it represent? What is it allowed to do, and what is it not allowed to do? If you skip this, the agent will happily answer questions about your competitors, give legal advice, or promise discounts you never authorised.

2. Tone and language rules. Do you use Nigerian English or American English? Do you address customers by first name or as "sir/ma"? Do you use "Naira" or "₦"? Is it okay to switch to Pidgin if the customer does? Write these rules down. Don't assume the model will guess.

3. The actual facts of your business. Prices, opening hours, delivery timelines, return policies, payment methods. If the agent doesn't have these, it will invent them. This is the single biggest source of public AI embarrassment.

4. Decision rules. What does the agent do when a customer asks for a discount? When someone wants to speak to a human? When a complaint comes in? When the question is outside scope? Be explicit. "If X, do Y. If you're not sure, do Z."

5. Examples. Three to ten worked examples of input and the ideal output. This does more work than any other part of the brief. Models learn from patterns, and examples are patterns. Telling an agent to "be friendly but professional" is vague. Showing it five replies you'd actually send is precise.

The test that matters

Once you've written a brief, run it through a simple test. Take twenty real messages from your inbox or WhatsApp. Past messages, where you already know what the right reply was. Feed each one to the agent and read the output.

Score each reply: would you have sent this? Yes, no, or "yes after small edits." If your hit rate is below 80%, the brief is not ready. Don't put it live. Go back, find the patterns in the failures, and add rules or examples to cover them.

A typical brief goes through four or five rounds before it's good enough to handle real customers. That's normal. It's also why most DIY agent projects stall — people stop iterating after the first version looks "okay" and then wonder why it embarrasses them in week two.

What to do when the agent goes wrong

It will go wrong. The question is how you respond.

When you see a bad output, don't just edit that one reply and move on. Ask why the brief allowed it. Then add a rule, an example, or a constraint that would have prevented it. Treat the brief as a living document. Every bad reply is a free piece of training data.

The businesses getting real value from AI agents are the ones doing this every week. The ones who set it up once and never touched it again are usually the ones telling you AI doesn't work.

The bottom line

AI agents do exactly what you brief them to do. If the output is vague, the brief was vague. If the tone is off, the tone wasn't specified. If the agent invents prices, you didn't give it the real ones.

Treat your agent like a new hire who reads fast, never sleeps, and forgets nothing — but only knows what you told it. Write the brief properly and the rest of the work gets a lot easier.

---

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.