Spaces

Do and don't rules for your AI agent when it talks to software users

This guide gives plain do and don't rules for an AI agent that answers users of a software product. It covers honesty about bugs, pricing, trials and when to bring in a person.

Runs in your browser. Nothing you enter is sent anywhere or saved.

Before you start

Your agent will speak for your company in hundreds of small moments. Most go well. The ones that go badly tend to be the same few: a promised feature date, an invented answer about pricing, a cheerful reply to someone who is furious. Rules prevent these.

Treat the list below as a starting point for your own instructions. Read each line, change the wording to match your product and add any rules your team has learned the hard way. Keep it short enough that people actually read it.

Tick each one as you go. 0 of 10 done.

[ ] Do answer how-to questions with the exact menu path and one help link.
[ ] Don't promise that a bug will be fixed or a feature will ship by a date.
[ ] Do say plainly when you do not know and hand the chat to a person.
[ ] Don't make up prices, limits or discounts that are not in your notes.
[ ] Do tell trial users the exact end date when they ask about it.
[ ] Don't extend a trial or offer a free month without a person agreeing.
[ ] Do acknowledge frustration in one sentence before solving the problem.
[ ] Don't ask for passwords, card numbers or API keys in the chat.
[ ] Do say you are an AI agent when a user asks who they are talking to.
[ ] Don't argue with a user about whether something is a bug.

Rules that earn their place

Good rules are specific and short. 'Be helpful' tells an agent nothing. 'Link to the exact help page and give the steps in order' tells it exactly what good looks like. Whenever you see a bad reply in a weekly review, turn it into a new do or don't.

Keep a copy with your hand-over rules, since the two work together. The do and don't list covers how the agent behaves, and the hand-over list covers when it stops.

FAQ

How many rules should we give the agent?

Ten to fifteen clear rules work better than fifty vague ones. Start with the mistakes you most want to avoid, and add a new rule only when a real chat shows the need.

Should the agent apologise for bugs?

Yes, briefly and sincerely, without taking blame for things it cannot know. Something like 'Sorry about that, I know it is in your way' is enough before you move on to help.

Can we change the rules later?

Yes, and you should. Your product, plans and policies change, so review the list when they do. Test the agent with a few example messages after each change.

More free tools for software companies

Read the guide: Answer your users' messages with an AI agent.