AI and automation
What to prepare before an AI assistant answers customers: content, boundaries, escalation, privacy, cost control, and a way to measure whether it helps.
An AI assistant is only as useful as the material it is allowed to use and the boundaries it is given. A chatbot that invents an opening time or a price is worse than a contact form, because it sounds confident while being wrong.
This checklist is for a service business preparing its first customer-facing assistant. It covers the work that happens before launch, not the choice of model.
Write one sentence that describes the job. Good examples are answering questions about services and hours, helping a visitor choose the right service, and handing a ready buyer to a booking page. A bad example is answering anything a customer asks.
Decide what the assistant must not do. It should not diagnose, give legal or medical advice, quote a fixed price for custom work, promise a delivery date, or collect payment details. Those boundaries belong in the system prompt and in the content the assistant is allowed to retrieve.
The assistant can only be accurate if the source material is accurate. Collect the current service descriptions, fees or fee ranges, opening hours, locations, cancellation terms, and the answers to the questions the front desk hears every week. Remove old pages and duplicate versions before importing anything.
Write answers in short paragraphs. A page written for search engines often contains several ideas in one block, which makes retrieval less precise. A question-and-answer format, with one topic per entry, works better for both customers and retrieval.
The most important answer is the one the assistant gives when it does not know. It should say so plainly, offer the contact route, and not fill the gap with a guess. Test this by asking questions that are deliberately outside the knowledge base.
Define how a conversation reaches a person. For a clinic, that may be a phone number during opening hours. For a software studio, it may be a form or a booking link. The handover should be visible in the conversation, not buried in a footer.
A chat conversation can contain personal information even when the business did not ask for it. Decide what is stored, where it is stored, who can read it, and how long it is kept. If the assistant offers to send a transcript or follow up by email, the consent line must be explicit and unticked by default.
Check the provider's data terms as well. Some providers retain prompts for training unless configured otherwise. The business remains responsible for what customers are told, regardless of which model answers.
Set a budget per conversation or per month, and decide what happens when it is reached. A hard stop with an honest message is better than a silent failure or an unexpected bill. Test with the questions real customers ask, including the awkward ones.
Run a small pilot with staff before the public sees it. Ask them to try to break it, and record every wrong answer. A launch with twenty known failures is easier to improve than a launch with no review at all.
Written and maintained by Bitecodes. If something here is wrong or out of date, tell us and we will correct it.
A developer-friendly checklist for finding and fixing the common causes of a slow business website, without chasing a single score.
Questions that reveal how a studio scopes, estimates, documents, and hands over work, with notes on what a good answer sounds like.