Every AI agent demo shows the happy path. A customer asks a clear question, the agent answers correctly, everyone nods. The interesting question is the other one: what does the agent do when it does not know?
Left to its own devices, a language model answers anyway. It is very good at producing text that sounds exactly like a correct answer. To a customer reading it on WhatsApp at nine in the evening, there is no visible difference between «our standard installation is four weeks» retrieved from your price list and the same sentence invented on the spot.
Why this is not a small bug
A wrong answer from a human employee is a mistake. A wrong answer from your automated channel is a commitment, made in your company's voice, at scale, with a timestamp. We have seen quoted prices that did not exist, availability confirmed for a slot that was taken, and — the one that keeps us cautious — a clinical suggestion offered by a booking bot that had no business offering one.
The rule
Every STEMD agent is built around one constraint that is not negotiable and not configurable: prices, terms, availability, and clinical or legal statements come only from approved company sources. Not from the model's training. Not from a plausible inference. From your price list, your calendar, your approved documents.
And when the source has no answer, the agent does the thing that feels least impressive in a demo and matters most in production: it says it does not know, and hands the conversation to a person — with the full context attached, so the human does not start from zero.
How we measure it
An escalation is not a failure of the agent. It is the agent working. So every agent ships with an escalation rate as one of its numbers, alongside response time and qualification rate. An agent that never escalates is not a good agent — it is an agent that has stopped admitting uncertainty, and that is precisely when it starts inventing.
This costs something. A more cautious agent resolves a smaller share of conversations without a human. We think that trade is obviously correct: the value of an automated channel is that customers can trust what it says. Spend that trust once and the channel is worth less than the phone it replaced.