How to Write AI Chatbot Instructions That Actually Work

Telling an AI chatbot to “be helpful and professional” is not an instruction, it is a wish. A real brief tells the bot four specific things: what to do when it does not know an answer, when to hand the conversation to a person, how long an answer should run, and what it can never say. Skip any of those four and the bot fills the gap itself, the same way a new hire does when nobody trained them.

Why “be helpful” does not actually work

A large language model answers from whatever it is given. It does not know your policies unless someone wrote them down for it (see what an LLM actually is for why that matters). Tell it to “be helpful” and it decides on its own what helpful means for every question, including the ones you never thought to cover. That is not the bot filling in a gap smartly, that is the bot guessing, and it sounds confident either way.

A real case shows what that guessing costs. In 2024 a passenger asked Air Canada’s website chatbot about bereavement fares. The bot told him he could book a full price ticket and apply for a reduced bereavement rate within 90 days, which was not the airline’s actual policy. Air Canada refused to honor it, and argued in front of Canada’s BC Civil Resolution Tribunal that the chatbot was “a separate legal entity” responsible for its own words.

“This is a remarkable submission. It should be obvious to Air Canada that it is responsible for all the information on its website.” — BC Civil Resolution Tribunal, Moffatt v. Air Canada

The tribunal ordered Air Canada to pay the passenger $812.02. The company did not lose because the AI wrote badly. It lost because nobody had written down what the bot could never say, so it invented something that sounded like real policy.

The four things a working brief actually covers

Every one of these has to be an explicit line in the instructions, not something you assume the model will figure out:

• What it does when it does not know. Say so plainly and offer the next step, never guess at a policy or a price.

• When it hands off to a person. A short, fixed list of things it will not try to handle itself: a complaint, a refund, anything with money attached. This is the same line we draw in how much autonomy to give AI in your business.

• How long an answer runs. Two or three sentences for most questions, not a paragraph. A long answer reads like a lecture, not a reply.

• What it never says. The actual banned list: prices you have not published, promises about timing, any policy you have not written down somewhere real.

Every one of those is testable. Ask the bot a question it should refuse, and check what it actually does.

What a vague brief turns into, next to a real one

Vague: “You are a helpful assistant for [Business]. Be friendly and professional and answer customer questions.”

That line gives the model a personality and nothing else. It never says what happens when a customer asks something outside the FAQ, so the model invents an answer that sounds right, the same failure that cost Air Canada its case.

A working version for the same business names all four rules directly: what to do when the answer is not in the material it was given, which topics get handed off immediately, how long a normal answer should run, and the specific claims it is never allowed to make on its own. Same model, same product, completely different behavior.

Who actually writes this

The brief is real work either way, hired out or done yourself. Somebody has to sit down with your actual FAQs, your actual policies, and your actual refusals, and turn them into rules the bot follows every time, not just the first time. When we build one of these for a client, that interview is most of the first week: reading their real support messages, finding the same handful of questions repeated, and writing down what they actually tell a customer when the answer is not simple.

The bot is only as good as that brief. A strong model with a lazy brief still guesses, the same way a smart new hire still improvises without real training.

---

Common questions

Do I need to write the chatbot's instructions myself?

No. Someone has to gather your real FAQs and policies and turn them into the four rules above, but that is normally the builder's job, not yours. Your job is answering honestly when they ask what you actually tell a customer.

What happens if I skip one of the four rules?

The bot fills the gap itself. Skip “what it never says” and it will eventually invent a price or a policy, the same mistake that cost Air Canada its case. Skip “when it hands off” and it will keep trying to answer things that actually need a person.

Can I change the instructions after the bot is already live?

Yes, and you should. A brief written once and never touched goes stale the same way a knowledge base does. Update it whenever a policy changes, or you notice the bot handling a question the wrong way.

Does a longer set of instructions work better than a short one?

Not automatically. Length is not the point, coverage is. A short brief that states all four rules beats a long one that describes your brand's tone for three paragraphs and never says what the bot should do when it does not know something.

Is this the same as training the AI on my content?

No, they are two different jobs. Training feeds the bot your actual material, FAQs, docs, past support answers, so it has something real to answer from. The brief is the separate set of rules for how it behaves once that material is in front of it. You need both.

---

We write this brief from scratch for every client, built from their real FAQs and policies, not a template. If you want to see what one of these looks like running on your own content, the free demo at demo.ajmarketingresults.com builds a working version from your last five videos in about a minute.