DEV Community

Rick Leffke
Rick Leffke

Posted on

Give your AI agent a written list of things it must never say

Most teams that put an AI agent in front of customers spend a lot of time on what it should say: the tone, the greeting, the product facts, the escalation path. Far fewer write down what it must never say.

That gap matters because the most damaging brand mistakes are usually short sentences. A promise you can't keep. A comment about a competitor. A guess about a legal or medical question. An apology that admits fault you haven't agreed to admit. None of these need a complicated failure. They just need an agent that was never told "not that."

Why "be professional" isn't enough

Instructions like "stay on brand" or "be professional" leave the line up to the model. It will fill the gap with its own idea of what sounds reasonable, and that idea can shift when the model, the prompt, or the conversation changes.

A written list removes the guessing. It turns a vague expectation into something you can check.

Start with three

You don't need a long policy. Start with three things your agent must never say for your brand. Some examples of the kind of items teams pick:

  • Promises outside policy. "I'll refund that in full" when refunds need approval, or "it will arrive tomorrow" when delivery isn't guaranteed.
  • Statements about other companies. Comparisons, criticism, or claims about a competitor's product.
  • Advice it isn't qualified to give. Legal, medical, tax, or safety guidance that should go to a person.

Your three will depend on your business. The useful question is: which single sentence from this agent would cause the most trouble if a customer screenshotted it?

Put them in word for word

Once you have the list, add it to the agent's instructions exactly as written, not paraphrased. Vague wording gets vague results. Spell out what to do instead, too: decline politely, offer to connect the customer with a person, or point to the written policy.

Then test it. Ask the agent questions that invite each forbidden answer and see what comes back. Rerun those same questions any time you change the model or the instructions, since a change can quietly undo what used to work.

A small first step

Write your three items down today, in one sentence each, and paste them into your agent's instructions. If you can't agree on three, that disagreement is worth sorting out before a customer finds the answer for you.

Disclosure

I run AgentRisk, a risk check for businesses putting AI agents in front of customers, so I think about lists like this a lot. More at https://rleffke.com/agentrisk.

How do you decide what your agent should never say? I'd like to hear what's on other teams' lists.

Top comments (0)