A bad AI assistant is worse than none: it frustrates customers and makes your business look careless. The difference between one that helps and one that embarrasses you is almost always the brief, not the model.
The brief that fails
- "Answer customer questions about our business."
- Pointed at the website and left to work it out.
- No rule against quoting a price.
- No handover, so it argues with people instead.
The brief that works
- "Answer questions about delivery, returns and stock. Book a callback for anything else."
- Grounded in the returns policy, the shipping page and the live stock feed.
- Never quotes a price. Sends pricing questions to sales.
- Two failed attempts and it offers a human, with the transcript attached.
Same technology in both columns. The one on the right just knows what its job is.
Get these five things right first
- One clear job. "Answer product questions and book demos" beats "do everything." A focused assistant is a good assistant.
- Its source of truth. Point it at your real content: docs, policies, product info. An assistant is only as accurate as what it is grounded in.
- What it must never do. Draw the lines: no medical or legal advice, no guessing prices, no promises you cannot keep.
- The handoff. Decide exactly when it passes a conversation to a human, and make that path smooth.
- How you will measure it. Agree what success looks like before launch: questions resolved, response time, enquiries captured.
