Case Study · Generative AI

Ambient Clinical Documentation and Coding Assistant for a Multi-Specialty Hospital Group

An ambient AI scribe that listens to the consultation, drafts the clinical note into the EHR, suggests the codes the encounter supports and flags the documentation gaps that cause claim denials, with PHI never leaving the hospital's own environment.

← All case studies

Client

Multi-specialty hospital group

Industry

Healthcare

Delivers

Notes drafted during the visit, not after hours

The problem

Clinicians spend a large share of every working day on documentation rather than patients. Notes are written between consultations or late at night from memory, which makes them thinner and later than they should be. Thin notes then cost the hospital twice: the coding team has to chase clinicians for missing detail, and payers deny or downcode claims that the record does not fully support.

The obvious fix, dictation, only moves the typing. It does not structure the note, does not know the patient's history, and does not tell the clinician what is missing while the patient is still in the room.

Why healthcare makes this hard

Protected health information: consultation audio and notes are among the most sensitive data a hospital holds, so nothing can be sent to a public API and every access has to be auditable.
Clinical accuracy: a plausible-sounding but wrong note is worse than no note. A hallucinated medication, dose or negation is a patient-safety event, not a formatting defect.
Messy audio: real consultations have two or three speakers, interruptions, ambient noise, code-switching between languages and heavy clinical shorthand.
EHR reality: the note has to land in the right encounter, in the right section, in the format the EHR and the specialty expect, or clinicians will not use it.
Coding and compliance: codes must be supported by what is actually documented. Suggesting codes the record does not justify is upcoding, so the system has to cite its evidence.

What we build

Consented capture

Ambient capture starts only after the patient consents, with a visible indicator and a one-tap stop. Audio is encrypted in transit and at rest and is deleted on a short, configurable retention clock once the note is signed.

Private-by-design deployment

Speech recognition and the language model run inside the hospital's own VPC or on-premise GPUs. PHI never leaves the client's environment, and every read and write is written to an immutable audit log.

Clinical ASR and diarisation

Speech recognition is adapted to clinical vocabulary, local drug brand names and accented, code-switched speech, with speaker diarisation to separate clinician, patient and relative.

Structured note generation

The transcript is turned into a specialty-specific structured note (SOAP, admission, operative or discharge), pulling relevant history, allergies, active medication and recent results from the EHR over HL7 v2 or FHIR so the note is written in context.

Evidence linking

Every generated statement is linked back to the transcript span or EHR record it came from. The clinician can click any line and see exactly what it was based on, which is what makes review fast enough to be worth doing.

Coding and gap detection

A retrieval layer over ICD-10, CPT and the payer's own policy documents proposes codes with the supporting evidence attached, and flags documentation gaps, for example a diagnosis stated without the specificity the payer requires, while the patient is still in the room.

Clinician sign-off

Nothing enters the legal record unsigned. The clinician edits and signs; the edits are captured as supervised training signal to improve the model on that specialty's own style.

Safety guardrails

Extraction over invention: medications, doses, allergies and numeric results are copied from structured EHR fields or explicit transcript spans, never generated freely.
Negation and uncertainty are preserved explicitly, so "no chest pain" can never surface as "chest pain".
Low-confidence passages are left visibly blank for the clinician rather than smoothed over with fluent text.
Clinical review board: a sample of notes is audited weekly by senior clinicians against the source audio, and model changes ship only after that review.
Role-based access, full audit trails, BAA-backed infrastructure and de-identification for any data used in evaluation or fine-tuning.

Why it matters

The impact is not that a note gets typed faster. It is that documentation moves back into the consultation, where the clinician can still check a detail with the patient; that the record is complete enough to defend a claim the first time; and that the hours clinicians currently lose to catch-up documentation return to patients or to rest. That is the difference between a transcription tool and a clinical system, and it is built almost entirely out of the unglamorous parts: privacy architecture, retrieval, evidence linking and review workflow.