Ambient clinical documentation, and the discipline of refusing

The shape of the problem

Clinicians spend a large part of every day on documentation, most of it after the last patient. In staff surveys it reliably ranks among the top reasons for leaving — often above pay.

Ambient scribe pilots address exactly this, and a good number of them are abandoned at the clinical safety review. Not because the transcription is poor. Because a proportion of generated lines assert something the encounter did not support, and in a clinical setting that is not a quality problem. It is a safety problem, and it ends a pilot.

Why the obvious fix does not work

The failure has little to do with model quality. The system has no way to refuse.

A model asked to write a clinical note will write a clinical note. Given a partial transcript and a template with a "past medical history" heading, it fills that heading — from the most plausible completion available, which is not the same thing as the patient's actual history.

The failure mode is not hallucination in the dramatic sense. It is smoothing: a note that reads correctly, is structurally complete, and contains lines nobody said. A better model smooths more convincingly.

The design

Grounding as a hard gate, not a prompt instruction

Every generated span has to be traceable to one of two sources: the encounter audio, or a retrieved element of the patient's governed record. Spans that cannot be traced are not written. The note comes back with the section empty and a marker saying the system had nothing to support it.

This produces notes that are visibly incomplete, and that is the hardest thing to sell internally. An empty "review of systems" heading looks like a broken product. It is the product working — the alternative is a filled-in heading that is sometimes wrong.

A deterministic lane over medications and allergies

Medications, allergies and dosages never pass through a generative path at all. They are retrieved from the record, checked against the pharmacy system, and rendered. If the two disagree, the note flags the disagreement rather than resolving it.

Evaluation that can block a release

The clinical safety group owns an evaluation suite built from annotated encounters, and the bar is set on unsupported-claim rate rather than on any fluency measure. No model version reaches a clinician without clearing it. A suite that has never rejected a candidate version is not a gate.

What to measure afterwards

• Unsupported-claim rate, measured continuously against a rolling audit sample rather than reported once at pilot.
• Note acceptance: accepted unchanged, accepted with cosmetic edits, materially rewritten.
• Documentation minutes per clinician per day, against a pre-deployment baseline.
• Rate of flagged medication disagreements, and how many were real.

Where it runs

Inside the provider's tenancy, in-country, on dedicated endpoints with no data leaving the estate. Information governance has read access to every trace, and the evaluation suite is the provider's asset rather than ours.

The uncomfortable property of this design is that the win comes from making the system do less. A pilot generates more. A production system generates less and is trusted, and being trusted is what makes any time saving real.