Verifiable, never marketing
Anyone can write that they know AI. This page is the opposite of that: how the agent you just talked to is built, what we decided and why, and what it costs to run in dollars. If you are going to hand us a process from your business, this is what you should get to look at first.
These figures come from the same database the agent uses, not from a slide.
There are not enough conversations yet for the figures to mean anything. When there are, they will show up here untouched. Publishing averages over four conversations would not be transparency, it would be noise.
All the content on this site — 30 pages in two languages — is about 13,000 tokens. The model takes 200,000. It fits whole in every request, so there is nothing to chunk, index or retrieve. That removes the classic failure mode of these agents: retrieval pulls the wrong fragment and the agent confidently tells you something untrue. Here it always has everything in front of it. Measuring before building saved an entire piece of infrastructure.
That content travels identically in every request, so it gets cached: each conversation re-reads about 10,000 tokens that cost a tenth of their normal price and do not count against the per-minute rate limit. Without it, an agent that reads the whole site on every message would be expensive. With it, it costs cents.
Besides answering, the agent can do exactly one thing: record your details if you tell it you want us to write to you. Nothing else. It does not delete, charge, promise or touch any system. Anyone who opens this site writes this model's input, so giving it broad capabilities would hand out attack surface for nothing in return.
When the agent records a contact, those fields do not go into the database as they come: they go through the same validation as a public form, because in practice that is what they are. A malformed email or an oversized field is rejected there. Seven automated tests cover that boundary specifically.
If the database does not answer, the agent still replies and points you to the form; the failure is logged on our side. Storing analytics must never break someone's page. It is the same rule we apply on client projects: the incidental does not take down the essential.
The server authenticates against the model with its own infrastructure's managed identity. There is no secret to store, rotate or leak. One credential fewer is one failure mode fewer, and this very project had already lost deployments to expired credentials.
A loose sales agent invents prices and promises deadlines. These constraints are written into its instructions, and you can check them by asking it:
Open the chat on this page and try to make it break one of those rules. It is the fastest way to check whether any of the above is true.