Case study

A clinical documentation platform

A digital health company building an ambient clinical scribe that listens to patient consultations and drafts structured clinical notes for clinician review. The product is used across primary care and outpatient settings, where documentation accuracy directly affects patient safety and clinician workload.

01 — In their words

“Clinicians didn't want a scribe that was confident all the time. They wanted one that knew when to say 'check this'. Once the drafts started flagging their own weak points, the conversation in our user interviews changed completely.”

Recorded verbatim · Attribution withheld

02 — The work

Challenge

The scribe produced fluent, well-structured notes — which was precisely the danger. A fabricated dosage or an inferred symptom reads exactly like a correct one, and tired clinicians reviewing at the end of a session are poorly placed to spot confident errors. Early pilots showed that clinicians either distrusted every draft and re-wrote from scratch, eliminating the time savings, or trusted drafts too readily. Neither failure mode was acceptable in a clinical setting, and adoption stalled.

Solution

The team deployed Hyperpriors guardrail policies over every generated note: medication names and dosages are checked against the consultation transcript, clinical claims without transcript support are flagged rather than silently included, and structured fields are validated against the practice's formulary data. A Hyperpriors harness routes flagged content into an explicit escalation path — uncertain passages are highlighted inline for clinician confirmation, and notes exceeding a risk threshold are held for full review rather than presented as ready.

03 — The system

Fig. 1 — System architectureHealthcare technology · Illustrative topology

04 — Results

The platform runs Hyperpriors guardrail policies and a human-escalation harness across its ambient scribe product, with every generated note passing through automated checks before a clinician sees it.

01
Unsupported clinical claims flagged before review in every draft, with fabricated-detail incidents in review sessions reduced to near zero
02
Structured review workflow cut average note-review time by roughly a third, because clinicians check flagged passages instead of re-reading everything
03
Clinician adoption in pilot practices more than doubled within a quarter of the escalation harness shipping

05 — The record

Industry: Healthcare technology
Location: Northern Europe
Model providers: One hosted provider plus a fine-tuned open-weight model
Workloads: Ambient clinical scribing, structured note generation

06 — The full account

The situation

The scribe worked well in the demo. It listened to a consultation, produced a clean structured note, and saved clinicians the documentation hour that usually follows a clinic session. The difficulty began where demos end.

Language models fail fluently. A hallucinated dosage does not look like an error; it looks like a dosage. In early pilots, clinicians split into two camps: those who re-verified every line, recovering none of the promised time, and those who signed off drafts quickly — which the company’s own safety review flagged as the more dangerous outcome. Trust was miscalibrated in both directions, and a scribe that clinicians cannot calibrate against is a liability, not a product.

The team concluded that the model’s accuracy was not the core problem. The core problem was that the product gave clinicians no signal about where to look.

What changed

The platform integrated Hyperpriors guardrails as a policy layer between generation and presentation. Every draft note is checked before a clinician sees it: medications and dosages are verified against the consultation transcript, clinical statements without transcript support are marked as unsupported rather than passed through, and structured fields are validated against reference data.

Crucially, the guardrails do not silently delete questionable content — they surface it. A Hyperpriors harness routes each flag into a defined escalation path: low-risk uncertainty is highlighted inline for one-click confirmation or correction, while notes crossing a risk threshold are held in a review queue rather than presented as complete. The clinician’s attention is spent where the system is least certain, which is the only place it was ever needed.

Where it landed

Fabricated details reaching the review stage dropped to near zero, because unsupported claims now arrive pre-flagged. Review time fell by roughly a third — clinicians confirm highlighted passages instead of re-reading entire notes defensively.

The number the company cares most about is adoption. Within a quarter of the escalation harness shipping, active clinician usage in pilot practices more than doubled. The lesson the team draws is a quiet one: clinicians never asked for a model that was always right. They asked for a system honest about when it might be wrong.

07 — Begin

Make uncertainty visible before it reaches the record.

Request access All case studies