The most valuable thing an AI agent does is refuse

An agent that answers every question is not confident. It is uncalibrated. The distinction matters enormously once real money or real patients are downstream of the output, and it is almost never what teams measure.

In the Helios Orbit deployment, the behaviour that convinced a compliance team wasn't the retrieval quality or the reasoning depth. It was the refusal path: when span-level confidence dropped below a threshold, the system escalated to a human instead of producing a fluent, plausible, wrong number.

The uncomfortable truth is that refusal is expensive to build. You need calibrated confidence, provenance for every claim, and a definition of 'unsupported' that survives an adversarial reading. None of it demos well. All of it is what keeps the system in production eighteen months later.

Our rule of thumb: if you cannot describe, in a sentence, the conditions under which your agent will decline to act — you do not have an agent. You have a very expensive autocomplete with a tool-calling API.