How to add an AI agent to a small business website
A useful website agent does a few specific jobs well: answer questions from your own approved content, qualify the enquiry, and hand off a booking or quote request to your real systems. Ground it in your content, give it clear limits and a human handoff, connect it to your tools through APIs or MCP, and review its conversations every week. Running costs are mostly usage-based fees paid to the AI model provider.
Chatbot or agent?
A basic chatbot answers questions. An agent can also take actions: check availability, create a booking, draft a quote or update your CRM. Most small businesses get the most value from something in between: an assistant that answers well and passes clean, complete enquiries to a person.
1. Pick the jobs, and the limits
Write down the three to five things the agent should do, and the things it must never do. For example:
- Do: answer questions about services, areas and hours; collect the job details and photos; offer booking times; take a quote request.
- Never: promise a final price, give advice outside your business, or invent an answer when it is not sure.
2. Ground it in your own content
Agents make things up when they answer from general knowledge. Give yours a set of approved sources instead: your services, pricing rules, service areas, policies and FAQs. Tell it to answer only from those sources and to say so when it does not know. Keep the sources up to date, because the agent is only as accurate as what you give it.
3. Add a human handoff
Every conversation should have a clear way to reach a person: a contact form, a call-back request or an email. Hand off automatically when the customer asks, when the question is outside the agent's limits, or when a decision needs your judgement, such as approving a quote.
4. Connect it to your tools
An agent becomes genuinely useful when it can act in the systems you already use. That happens through each tool's API, or through MCP (Model Context Protocol), an open standard introduced in late 2024 for connecting AI models to tools and data. Common connections are a calendar or booking tool, a CRM or job management system, email or SMS, and payments.
5. Handle privacy properly
- Tell visitors they are talking to an AI assistant.
- Collect only what the job needs, and say how it is used in your privacy policy.
- Keep API keys and customer data on the server side, never in the website's code.
6. Know the running costs
AI model providers charge by usage, based on how much text the model reads and writes, so cost scales with the number and length of conversations. Add any hosting and the subscriptions for the tools it connects to. Set a monthly usage limit with the model provider so a busy month cannot surprise you.
7. Review and improve
Read a sample of conversations every week for the first month. Look for questions it could not answer, answers that were wrong, and enquiries it should have handed off sooner, then update its sources and limits.
A real example
TradeCatcher, a platform launched by the founder of this studio, uses AI in this way: it collects job photos from customers by SMS, drafts a ballpark quote from the tradie's own rates, and leaves the final approval to the tradie. The TradeCatcher case study shows how it fits together.
If your agent runs on Workers AI
If your agent runs on Cloudflare Workers AI, check the model it calls: Cloudflare is retiring 24 Workers AI models on 20 November 2026.