Business case and benefits
- CCIO
- A · Accountable
- Product owner
- R · Responsible
- Solutions architect
- C · Consulted
- Technical PM
- C · Consulted
- Engineers
- I · Informed
- Clinical safety officer
- C · Consulted
- IG & DPO
- I · Informed
- Security
- I · Informed
The consultation is captured, the note and the letter are drafted before the patient leaves, and the clinician checks, edits and signs every word.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
In a busy outpatient clinic, the clinician has two jobs at once: listening to the patient and writing it all down. Notes get typed with one eye on the screen, or saved up for the end of clinic. Letters are dictated, typed by a medical secretary, sent back for checking and signed days later. In a typical trust, the letter to the patient and the GP goes out a week or more after the appointment, and the GP practice acts on it later still.
Ambient voice technology (AVT) changes the order of work. With the patient's knowledge, it listens to the consultation, transcribes it, and drafts the clinical note, the clinic letter and suggested codes in the specialty's own format. The clinician reads, corrects and signs before the next patient comes in. NHS England is funding a national rollout that starts in outpatient settings.
The tools are easy to demo and hard to run well. A draft that reads fluently can still be wrong: a drug name misheard, a side mentioned in passing turned into a diagnosis, a plan said aloud that never reaches the EPR. Accents, interpreters, background noise and three people talking at once all test it. The real work is the design around the tool: consent, checking, integration with the EPR, clinical safety and evidence that it helps.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Clinician time spent writing notes and letters, per clinic, measured before and after.
ExampleDown by a third or more
Days from the appointment to the signed letter reaching the patient and the GP.
ExampleFrom 7–10 days to 1–2
Clinically significant errors in signed notes, from a weekly sample read against the audio or transcript.
ExampleNone unresolved, and falling
How much clinicians change the drafts. Very high means it isn't helping; near zero can mean drafts are being signed unread.
ExampleStable, and reviewed weekly
How many patients say no, and why, split by age, language and specialty.
ExampleTracked, with no group avoiding care
Transcription quality for different accents, interpreted consultations and noisy rooms.
ExampleNo group materially worse
Cross-sector rules are on the UK regulation map.
Organisations stay accountable for safe use. The guidance expects a DPIA, a clinical safety officer and DCB0160 documentation, integration with the EPR, clear contracts on liability and model updates, monitoring independent of the supplier, and staff trained to review every output. Pilots are time-limited to four months and must not bypass these steps.
Products that only transcribe, summarise, draft correspondence or suggest codes for clinician review are generally not medical devices. Products intended to support diagnosis or treatment can be. Write the intended purpose down, keep the tool to documentation, and re-check its status whenever the supplier adds a feature.
The supplier provides its DCB0129 safety case; the trust, as deployer, needs its own DCB0160 safety case, hazard log and monitoring. Hazards include misheard medicines, invented findings, missed actions and the wrong patient's record. Both standards are under review, so expect revised versions.
Consultation audio is special category health data. Update privacy notices before go-live, tell patients what is recorded and how long it is kept, and give them a real chance to object. Keep audio only as long as needed to produce and check the note, under a retention period set in the DPIA.
The guidance expects both the trust and the supplier to meet the DSPT and hold Cyber Essentials. The new service, its supplier and its data flows go into the trust's CAF-aligned DSPT submission and information asset register.
The clinician who signs the note is responsible for it, whoever or whatever drafted it. Training and the sign-off step make that explicit, and the audit trail records who checked what.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Build or buyMost trusts will buy. The registry is self-certified, not an endorsement, so test the evidence like any other supplier's. The architecture below applies either way: it's what you integrate, assure and monitor. | A product from NHS England's self-certified AVT supplier registry or a national framework, or your own build on AWS | A product from the registry or a national framework, or your own build on Azure | A product that can be hosted privately, or your own build on open-weight models |
| Capture: app, audio, filesCapture on trust-managed devices only. Audio is encrypted in transit and at rest, and deleted on a short, agreed schedule. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Speech to textAdd drug names, local place names and specialty terms. Test on your own patients' accents, interpreted consultations, children's clinics and noisy rooms, not a supplier's benchmark. | Amazon Transcribe (en-GB) with a custom vocabulary | Azure Speech in Foundry Tools, with custom speech | Open-weight speech models (e.g. Whisper) on your own GPUs |
| Drafting and reasoningBoth clouds offer a choice of models, including Claude. Templates per specialty, agreed with clinicians. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| Clinical codesSuggested codes are validated against the current SNOMED CT release, and OPCS-4 where the trust codes outpatient procedures. The clinician or coder confirms each one. | NHS England Terminology Server via FHIR | NHS England Terminology Server via FHIR | NHS England Terminology Server, or a local FHIR terminology server |
| Workflow | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to the EPRThis decides the design. Agree in discovery whether drafts land in the EPR for signing, or sit beside it to copy from. In general practice, the route is IM1. | Through your EPR supplier's published interfaces (FHIR, HL7 or their API) | Through your EPR supplier's published interfaces (FHIR, HL7 or their API) | Through your EPR supplier's published interfaces (FHIR, HL7 or their API) |
| GuardrailsBlocks the draft from adding diagnoses, medicines or doses that aren't in the transcript. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionClinicians sign in with the identity they already use. Smartcard access stays with the EPR. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every draft, every edit and every sign-off, with the model and template version. Keep your own measures, independent of the supplier. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Many trusts' IG policies will require UK processing for consultation audio. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The EPR, the PAS and the letter route to patients and GPs stay where they are, with your EPR supplier. The signed note lives in the EPR, never in the AI service, and smartcard access is unchanged.
Uses cloud AI
Transcription, drafting and the draft check run in a UK cloud region, inside the chosen product, or on the trust's own GPUs for the sovereign option. Audio and transcripts are kept only until the note is signed and checked, then deleted on the agreed schedule.
How they connect
Capture devices reach the service over the trust network or HSCN. Drafts go back through the EPR supplier's interfaces and an API gateway, over a private link such as AWS Direct Connect or Azure ExpressRoute, or inside the trust's data centre.
If the link drops
The clinic carries on. The clinician types the note or dictates the letter exactly as before the AI, and the capture app tells them clearly that nothing is being recorded. No consultation depends on the scribe working.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Note and letter drafter
Where a person signs offThe clinician reads, edits and signs every note and letter. Drafts stay marked as drafts until they do. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Draft checker
Where a person signs offFlags are shown to the clinician at the point of signing. A weekly audit sample is read by clinicians against the source, independent of the supplier's own figures.
The EPR stays the record. Drafts reach it only through the EPR supplier's published interfaces and the API layer, never by screen-scraping or direct database writes. If the supplier only allows a side panel, the clinician copies from it, and the design accepts that until a better route exists.
Events keep the AI decoupled. An ended consultation becomes a message on a queue, and the draft is written back when ready. If the service is slow, the clinician isn't blocked: they can sign later or write the note themselves. Every message carries the appointment ID, so a retry never files a duplicate.
The signed letter leaves through the trust's existing route to patients and GP practices, with its usual checks. The AI never sends anything directly, and coded data goes into the EPR the same way as codes entered by hand.
Map today's note and letter process, measure the baseline, agree the EPR integration route, shortlist products from the registry or a framework, and engage the ICB and regional team.
Gate · must be true to passBaseline agreed, EPR route confirmed, clinical safety officer named, DPIA and hazard log started, intended purpose written down.
Two or three specialties with willing clinicians. Drafts are generated, clinicians sign as usual, and quality is audited against the source every week.
Gate · must be true to passError rate and edit rate within agreed limits, no group of patients worse off, DCB0160 safety case and CCIO sign-off.
Go live in the pilot specialties with monitoring, support, contract terms for model updates, and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan, training and patient information in place.
Add specialties in waves, then emergency and inpatient settings, then community services.
Gate · must be true to passEach new specialty passes the same pilot gate, with its own template and accent and noise testing.
| Activity | CCIO | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Clinical safety officer | IG & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Informed |
| Product selection and EPR integration design | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted | Consulted |
| Clinical safety case and hazard log | Accountable | Consulted | Consulted | Informed | Consulted | Responsible | Informed | Consulted |
| DPIA, patient information and DSPT | Consulted | Consulted | Consulted | Informed | Informed | Consulted | Accountable and Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Consulted | Informed | Consulted |
| Pilot audit against the source | Accountable | Responsible | Consulted | Responsible | Consulted | Responsible | Informed | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Accountable | Responsible | Consulted | Consulted | Responsible | Consulted | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: product selection against your EPR, the integration route, audio retention, security, monitoring and the gates, working closely with your clinical safety officer and CCIO. I'd report to your programme board.
Option B
If your partner, your EPR supplier or your own team already has an architect, I'd take on the EPR integration, the event flow and the independent monitoring under their lead, and give them a second senior opinion on security, supplier evidence and the clinical safety case.
Goes wrong
The demo was one clear voice in a quiet room. Real clinics have accents, interpreters, children and corridor noise.
How the design avoids it
Test on your own clinics, every specialty and voice, and report accuracy by group, not as one average.
Goes wrong
Clinicians sign fluent drafts without reading them, and an invented finding enters the record.
How the design avoids it
A draft checker that flags unsupported statements, short drafts, an edit-rate measure and a weekly audit against the source.
Goes wrong
Drafts sit beside the EPR, so clinicians copy and paste and the time saved disappears.
How the design avoids it
Agree the EPR integration route in discovery, before choosing a product.
Goes wrong
A supplier update adds a 'suggested diagnosis' feature, and the tool becomes a medical device without anyone deciding it should.
How the design avoids it
A written intended purpose, contract terms on change notice, and a re-check of device status and the hazard log at every update.
Goes wrong
Patients aren't told properly, and trust is lost after one complaint.
How the design avoids it
Information before the appointment, a consent script, an easy no, and opt-outs tracked and listened to.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Talk to your regional digital transformation team early, and collect baseline measures now so you can show benefits in the national format.
If you buy through it, plan contract length and exit terms around its end date, and keep your integration independent of one product.
Keep the hazard log and safety case structured and current, so moving to the new versions is an update, not a rewrite.
Keep the intended purpose written down and narrow, and review it against the new rules as they land.
Make sure signed notes and codes go into the EPR as structured, coded data, so they can flow into the shared record when it arrives.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.