An AI medical scribe the NHS can run on its own infrastructure.
Speech recognition tuned for clinical language, structured notes in the board’s own templates, and SNOMED CT coding suggestions the clinician controls. Built with an NHS health board in Scotland and delivered on schedule across four milestones.
Documentation was eating consultation time, and the off-the-shelf answer was not an option.
In primary care and outpatient clinics, clinicians record the consultation, assign clinical codes and produce structured letters, all while trying to keep their attention on the patient. Documentation is one of the most time-intensive parts of the working day.
Commercial AI scribes have started to appear in NHS settings, but they run on external infrastructure and do not map to NHS coding standards. For a health board with information-governance responsibilities, that is a hard stop: patient data cannot leave the organisation’s control, and a tool that cannot code to the board’s own standards creates downstream data-quality problems.
The board wanted to know whether a locally controlled alternative was viable: a scribe built on open foundations, running on NHS infrastructure, aligned with NHS governance from the start.
A scribe designed around how the board’s clinicians actually document.
Clinical speech recognition
A speech recognition model adapted to clinical language: drug names, symptom vocabulary and specialist terminology, with simulated local accents in the training mix. General-purpose recognition mishears exactly the words that matter most in a consultation.
Specialty-configurable note generation
Transcripts alone are not useful to a clinician. The system produces structured notes in standard SOAP and extended GP consultation formats, with profiles per specialty that set the note structure, background context and relevant terminology. Verbosity is configurable from brief summary to comprehensive record. Demonstrated across mental health, diabetes and learning disability services.
SNOMED CT code association
Clinical concepts in the generated note are identified and matched to SNOMED CT codes, with three tiers of highlighting: standard codes, the board’s recommended codes with SNOMED mapping, and specialty-specific codes. Clinicians review, add, change or remove any suggestion before the record is finalised.
Outpatient letter generation
Structured notes transform into outpatient clinic letters, removing a second pass of manual documentation after the consultation.
- Built-in user guide inside the application, so adoption does not depend on separate training material
- Fortnightly sessions between the technical team and NHS clinicians throughout the build
- All four milestones delivered on schedule, closing with a stakeholder presentation and live demonstration
Accuracy on clinical speech, without a single real patient record.
The governance constraint shaped the engineering. No real patient data could be used at any stage, so the speech model had to be adapted and evaluated entirely on de-identified synthetic clinical speech. That meant building a benchmark that reflected real consultation conditions, including accent variation and non-clinical conversation, before the first improvement could be measured.
The second challenge was structure. A clinician does not want a transcript; they want the right note in the right template, with the right codes, for their specialty. Getting that right required the specialty profile system rather than a single generic prompt, and it required the coding layer to align with the board’s own recommended code set rather than a generic terminology lookup.
The team also identified the next hard problem early: negation. “No chest pain” must not be coded as chest pain. SNOMED CT has explicit absence codes for some findings but not all, so this was scoped as a defined enhancement for the clinical-validation phase rather than left as a silent failure mode.
The clinician stays in control of every line.
A consultation is recorded and transcribed. The transcript becomes a structured note in the clinician’s chosen template, with sections they can enable or disable per consultation type. Suggested SNOMED CT codes appear inline with their tier highlighted. The clinician edits the note and the codes, then generates the clinic letter from the finished record.
Nothing is committed without review. The system is a drafting and coding assistant, not an autonomous author. That boundary was a design decision, not a limitation, and it is what makes the tool acceptable in a clinical-safety and governance review.
Measurable gains, honestly scoped.
Figures below are from the project’s outcomes report. The prototype was tested by a small group of NHS clinicians on synthetic data; clinical-environment validation is the next phase.
Built to the board’s rules, not around them.
Designed to run on NHS infrastructure. Transcription, note generation and coding operate inside the organisation’s control, with no dependency on an external AI scribe service.
Development on de-identified synthetic data only, with no integration to live patient record systems during this phase. Real-data handling and consent frameworks are defined for the validation phase.
SNOMED CT association aligned to the board’s own recommended code set with mapping, so coded output fits existing data-quality and reporting processes.
Clinician review of every note and code before anything is finalised. Comprehensive clinical-safety validation is a stated prerequisite before any deployment.
Where this expertise lives.
Need AI that runs inside your organisation’s walls?
Book a 30-minute consultation with a senior consultant. We will map where AI can compound in your operation and what the first 90 days could look like.
Book a Consultation →