Finn, my AI sales assistant

The problem

I work alone, so I can’t answer every visitor the moment they arrive. Most first conversations also go the same way: what do you want automated, roughly what’s the budget, who decides, and when do you want to start.

I wanted an assistant that could have that conversation properly, answer questions about how I work without making things up, and put the right people straight on my calendar.

What I built

Finn is a chat widget built as a WordPress plugin, running on Claude Haiku 5.5 by default. The rules for who gets a meeting live in plain code, not in the prompt. Finn books a call once it has a name, an email the visitor typed and one line on what they want. It asks once about the company, budget, timeline and who decides, and saves the answers for me, but they never hold up a booking. Finn gathers the answers, but the code decides, so nobody can talk it into a booking.

When someone qualifies, Finn offers open times from my Google Calendar and books the one they pick, with a Meet link. Every conversation is saved as a lead in WordPress, and I get an email when someone leaves their address or books.

For questions about pricing, timelines and how I work, Finn answers from my own company documents. I built five ways of giving it those documents (none, keyword and vector search, a knowledge graph, a hybrid of the two, and simply including all of them) and measured each one. A second, cheaper model then checks every reply against the sources and trims anything they don’t support before it goes out.

How it turned out

I tested Finn with 42 questions, from simple facts to multi-step questions and attempts to get it off track, and had each answer graded against the documents it actually had. It handled every off-topic and trick question. With my documents this small, giving it all of them beat search: that setup gave complete answers to 94% of factual questions with the top model.

The fact-check pass is what makes the cheap model usable. On Haiku it lifted fully supported answers from 79% to 90%, at an average of about $0.002 per answer. A monthly spend cap keeps the bill predictable.

Want one like it for your business? Here’s how I build a lead qualification agent for other companies.