What never goes in a small-business chatbot

A chatbot on the shop site is not a receptionist. It is a public speaker with a memory you do not fully control. If you paste live prices, license numbers, a second Maps listing, or customer ticket IDs into the training notes, that text can come back in a chat with a stranger. Ethical AI is in Frisco, Texas. We install listings, on-page SEO, and review workflows for trades and other local shops. We ask clients to finish a one-page shop card before they open a bot. The longer field guide is the pillar: AI for small business — how to use AI sustainably in 2026.

The method is the same on ChatGPT, Claude, Gemini, or a site widget. The model drafts. A named person sends. If a fact is not on the shop card, the bot writes that the office will call — it does not invent a street or a number.

Pricing stays off the bot

HVAC and trade prices move by job: after-hours, attic access, permit fees, parts that changed this week. A chatbot that quotes a number becomes the shop’s public price list. When the real invoice is different, you own the argument. Keep ranges off the widget. The bot can collect the city, the system type, and a callback number. A named person sends the estimate.

Do not invent a dollar amount in the prompt “so the model has something.” If a fact is missing, say the office will call. Ethical AI · Frisco, TX · 469-669-2728.

Permits and licenses stay off the bot

License numbers, insurance policy IDs, and permit status change. A stale number on a chat transcript is a complaint waiting for a screenshot. Put the legal name and the city on the shop card. Put the current license on the invoice and on the listing that already ranks. If Google and Yelp already disagree on the phone, fix that before you add a widget.

Angry callers do not belong in the script

A 1-star caller is not a FAQ. The bot should offer a human handoff and stop. Do not let it apologize for a job it cannot see. Do not let it offer a coupon. Review replies are a different workflow: the model drafts, a named person sends. That is covered in how to request reviews without faking them. Same rule for chat: the model does not close the ticket.

Training-data off switch

Every vendor that stores chats will train, log, or both unless you turn that off in writing. For a five-person shop the rule is simple:

• No customer names, ticket IDs, or photos in the system prompt
• No live price sheet
• No second Google listing “just for ads”
• Opt out of training where the vendor allows it
• Export the prompt and the FAQ so you can fire the vendor without losing the copy

If you cannot explain the monthly bill and the off switch in one paragraph, you do not have sustainable AI. You have a leak with a chat bubble.

Need the next step? Start at get started. Do not buy a chatbot until name, address, and phone match on the directories that already send you calls.