Document copilots, patient apps, transcription and triage support for providers and health-tech companies, designed around patient-data protection, clinical safety and a human reviewer at every decision that matters.
The quickest wins sit in the paperwork around care, not in the diagnosis. Each of these touches a system you already run.
Referral and letter triage
Classify incoming referrals and clinic letters, extract the key fields and suggest a priority for a clinician to confirm. Touches the document management system and the patient administration system. A typical outpatient department measures the win in administrator hours per week.
Consultation transcription and note drafting
Record with consent, transcribe, and draft a structured note the clinician edits and signs. Touches the electronic health record. The design question is where audio is processed and how long it is kept.
Patient-facing assistants in the app
Appointment changes, preparation instructions and answers from approved content, inside a patient app or portal. Touches the booking system and the content library. Anything resembling clinical advice routes to a person.
Policy and protocol search for staff
A grounded assistant over clinical guidelines, local protocols and HR policies, with citations to the source document and version. Touches the intranet and document store. Built on our internal copilot pattern.
Operational prediction
Did-not-attend risk, bed demand and discharge readiness forecasts from the data you already collect. Touches the data warehouse and the scheduling system. Used to prompt a call or reorder a list, never to deny care.
Remote monitoring and adherence apps
Mobile apps that collect readings, nudge medication adherence and summarise trends for the care team. Touches wearables, the patient record and notification services. See AI for mobile apps.
Compliance and data
Patient data and clinical safety shape the architecture
What we design for in healthcare
Health data is special-category personal data under GDPR and UK GDPR, so lawful basis, minimisation and retention are decided before any model sees a record. Clinical safety means a named reviewer for every AI output that could influence care, clear labelling of AI-drafted content and a documented process for errors. Where a feature could count as a medical device, we flag it early so your regulatory advisers can weigh in. This is engineering guidance, not legal advice.
In practice that means de-identification before indexing, role-based access enforced at retrieval time, an audit log that records what was retrieved and who approved what, and a deployment choice made per use case. Our AI governance service produces the policies and controls your information governance team will ask for.
Reference architecture
Flagship workflow: clinical document copilot
Documents enter from the existing document store. A de-identification and classification step runs inside your environment before anything is indexed. Retrieval respects the user's role, so a receptionist and a consultant see different answers to the same question.
The copilot answers with citations to the source paragraph and version. The clinician reviews, edits and signs off. Every step writes to an audit log that your governance team can query. The same index later powers transcription summaries and patient-app answers.
We usually build this on PostgreSQL with pgvector and either a hosted model under a data processing agreement or a privately deployed open-weight model. The secure RAG architecture page explains the controls.
Identifiable data never leaves the dashed boundary; the clinician remains the signatory.
Track record
What we have built that applies
Patient and clinician mobile apps
Since 2006 we have shipped native iOS and Android apps with secure authentication, offline data capture and device integrations, the foundation for monitoring and patient-engagement products.
The AI support bot in Eventify answers only from approved content and hands off cleanly to a person. That is exactly the behaviour a patient-facing assistant needs.
Sometimes, under the right contracts and with de-identification, but we never assume it. The default design keeps identifiable data inside your environment and sends the model only what it needs. Where that is not acceptable we deploy open-weight models privately. Our AI data privacy service works through the options with your information governance lead.
Is an AI assistant a medical device?
It can be, depending on what it claims to do. Software that summarises a letter for a clinician to read is a different proposition from software that recommends a dose. We design to keep the first category clearly in scope and flag early when a feature would move into regulated territory, so you can involve regulatory advisers before building.
How accurate does clinical transcription need to be?
Accurate enough that the clinician's review is quick rather than a rewrite. We measure word error rate and, more importantly, medication and number errors on your own recordings, and we keep the clinician as the signatory. See our speech-to-text service.
What is a sensible first project?
Administrative document load is usually the best start: referral triage support, letter summarisation or inbox classification. The value is measurable in staff hours, the risk is contained by human review, and the same infrastructure later supports clinical use cases.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.