Built for: commandmodule.com — the agent in the corner of this page
The Situation
Every business wants an AI agent on their website. Most of what actually ships is one of two failures: a canned FAQ bot that frustrates everyone, or a raw language model with a logo on it — one that will cheerfully invent pricing, promise things the business never offered, and recite its own instructions to anyone who asks nicely.
The Problem Behind the Problem
The hard part of a business agent isn’t answering questions. It’s boundaries. An agent that speaks for a company needs to know what it is not allowed to say: no made-up prices, no claimed clients that don’t exist, no leaking its instructions, no drifting out of the company’s voice. Those guardrails are the product. The conversation is just the interface.
What We Built
The agent answering questions on this site right now. It’s powered by Claude and runs on an edge worker — serverless, fast, and costing pennies. It’s locked to this domain, its prompt is cached to keep per-conversation cost near zero, conversations are capped and time-limited, and its instructions encode the boundaries: it won’t quote a price we haven’t set, won’t invent a client list, and deflects attempts to fish out its instructions — in character, without breaking stride.
It also knows the business — the products, the process, the founder’s actual background — because an agent that has to say “I don’t know” about its own company is worse than no agent at all.
The Result
You can run this case study yourself, right now. Click “Start a conversation” and interrogate it. Ask it what an engagement costs. Try to get it to promise you something. That’s the point — the best case study is one you can test live.