Teach an AI agent your tone and rules, no code needed

When someone tells you an AI agent can sound "just like you" and stick to "your rules," it's easy to picture a developer typing away for weeks. The good news: shaping an agent isn't coding, it's writing. If you can explain to a new hire how you talk to customers and what they may and may not promise, you've already done most of the work. Let's turn what's in your head into instructions an agent can follow — without touching a single line of code.

"Tone" and "rules" are two different things

This is the most common mix-up, and it's worth clearing up before you start. Tone is how your business sounds: formal or casual, short lines or fuller explanations, emojis or none, how you open and how you sign off. Rules are something else entirely: what the agent is allowed to do and say, which facts are fixed, what it must never promise, and the exact moment it hands the conversation to a person.

You need both, and you need them separately. An agent with great tone but no rules sounds charming while it invents opening hours that don't exist. An agent with rigid rules but no tone is correct and cold, like a form. The trick is to give it both layers at once: how it speaks, and what it's allowed to say.

Your tone fits in five sentences

Almost nobody can describe their own tone, and "friendly but professional" tells an agent nothing. Instead of adjectives, give it concrete decisions. Answer these and your tone is defined:

• Formal or first-name basis? Match how you actually talk to customers.
• Short and to the point, or explanations with context?
• Emojis or not? If yes, which ones, and how many per message?
• How do you greet, and how do you sign off? Use your real wording — the one you type every day.
• Three words or phrases you use a lot, and three you'd never say.

From that you can build a short brief the agent follows to the letter:

First-name basis. Short sentences. One emoji at most, and only to close. Open with "Hi [name]!" and sign off with "Any time — we're here." Never say "Dear valued customer" or "at your earliest convenience."

That's far more useful than half an hour describing your "brand personality." These are instructions someone can actually carry out.

Rules are your business written as commands

There's nothing to guess here: these are the facts and decisions you already make every day, you've just never written them down. Go category by category. What doesn't change (hours, address, payment methods), what the agent may do (offer a slot from your calendar, share an order status), what it never does, and when it has to call a human.

Write them as plain commands, the way you'd brief someone starting tomorrow:

Hours: Mon-Fri 9:00-18:00. Closed in August.
Bookings: offer open calendar slots; only confirm if the customer picks one.
Prices: quote the catalogue; for custom jobs, collect details and flag a human.
Never: promise delivery dates or talk about other clients.
If unsure: say so and pass the conversation to a person. Don't make it up.

Pay attention to that last line, because it's the most important of all. Most AI scares don't come from the agent not knowing something — they come from it inventing an answer with total confidence. A single rule, "if you don't know, say so and escalate," heads off around 90% of those problems.

Teach with examples, not adjectives

This is the step nearly everyone skips, and it's the one that changes the result most. Take five or ten real conversations that went well and paste them in as-is. One good example teaches more than ten adjectives, because the agent learns how you actually resolve things, not an abstract description of it.

Show it the "not like this" too. If an auto-reply ever came out stiff, blunt, or too salesy, show it and mark why it doesn't fit. With a handful of "yes / no" cases, the agent calibrates your register far better than any list of brand values could. It's the difference between telling someone "be friendly" and showing them how you handle a customer who arrives annoyed.

When NOT to fine-tune this much

Over-tuning has a cost too, so it helps to know where to stop:

• Don't bake in facts that change weekly — sale prices, one-off stock, promotions. Let the agent look those up from the real source (your site, your catalogue, your system) rather than memorise them and go stale at the first change.
• Don't try to cover every edge case on day one. You'll get stuck and ship nothing. Cover the usual 80% — the stuff that repeats — and let a person handle the rest.
• If your business runs on personal relationships — a big sale to close, a sensitive query, a serious complaint — don't hand over the whole message. Use the agent for first contact and show up yourself where it counts.

Remember the principle: you approve, they execute. The goal isn't a perfect robot that does everything, it's a reliable assistant that knows exactly when to stay quiet and pass the baton.

Start with a single page

Don't open a twenty-page document or a process manual. Start with one page: half of it tone, half of it rules. Run it for a week on real conversations, watch where it slips, and adjust the two or three lines that failed. Within a month you'll have a "playbook" any assistant — human or not — could follow without asking you a thing.

If you want to see this in action, our receptionist agent is where it shows most: it handles first contact over WhatsApp and the web in your tone and by your rules, clears the repetitive stuff, and passes you only what genuinely needs your judgement. You bring the page; it brings the hours.