- Home
- Industries
- Health and life sciences
AI agents for healthcare and life sciences
Health data is special-category data under GDPR, and medical software has its own regulation. We work on the administrative and documentation side, where agents help most and the risk is easiest to manage, and we say clearly when something would be a medical device.
The problems we hear first
Referrals that arrive incomplete
A missing result or an unclear question means a call back to the referring practice before anything can be booked.
Letters written twice
Clinicians write follow-up letters and summaries from notes they already wrote once.
Regulatory and quality documents
Submissions, SOPs and deviation reports need every statement traced to an approved source.
Patient questions by phone
Preparation for a procedure, opening hours, what a leaflet says: answers that exist, asked again and again.
Each one prepares. A person decides.
Examples of what we build in this sector, not client work. For each one: where a person approves, and what it never does.
Referral intake assistant
Reads incoming referrals, extracts what the team needs, flags what is missing and prepares the record.
- A person approves
- A coordinator or clinician reviews every record.
- It never
- Triage patients or set clinical priority.
Clinical letter drafter
Drafts the letter from the visit note and your template, in plain words, each sentence traced to the note.
- A person approves
- The treating doctor signs or discards every letter.
- It never
- Add a diagnosis or treatment advice that isn’t in the note.
Regulatory drafting assistant
Drafts sections of submissions and quality documents from your approved sources, with the source of every statement.
- A person approves
- Your regulatory lead approves every section.
- It never
- Sign or submit a document.
Patient information answers
Answers from your approved leaflets only, says it is an AI, and hands over to staff.
- A person approves
- Anything clinical goes to your staff.
- It never
- Give a diagnosis, triage or medical advice.
One job, step by step, with the person inside the flow.
Referral intake
A referral for an outpatient cardiology appointment arrives by secure mail. The agent prepares the intake record; a coordinator reviews it.
- Sector
- Health
- Risk class
- Administrative: no triage, no clinical priority
- Who decides
- An intake coordinator
- Referral receivedA letter from a GP practice with two attachments (sample).Done
- Only what intake needsPatient identifiers, the reason for referral, the attached results and the service asked for; nothing from other records.GDPR Art. 9Art. 5(1)(c)Done
- Record preparedYour intake template filled in, each field with the page it came from.Drafted
- Missing resultThe letter mentions an ECG that isn’t attached. A request to the practice is drafted.Flagged
- Waiting for the coordinatorThe agent sets no clinical priority.Accept the recordRequest the ECGSend to a clinicianAI Act Art. 14Waiting for a person
- Booking request createdOnly after the coordinator accepts.Locked
- Run recordedThe referral, the record, the request and who decided.AI Act Art. 12On record
A sample, scripted on invented data: no model is called and nothing here is client work.
Where the law draws the line for you
Each rule with the day we read it. An engineer’s reading for your counsel to confirm, not legal advice. How we read the AI Act.
| Rule | What it says (our summary; quotes are verbatim) | What it means for an agent |
|---|---|---|
| GDPR Art. 9Read 12 November 2026 | Health data is a special category. Processing it is prohibited unless an Art. 9(2) condition applies, such as (h) for medical diagnosis, care and treatment and the management of health systems, under professional secrecy (Art. 9(3)). | Each agent has its condition named, a DPIA and the narrowest data path. |
| GDPR Art. 35Read 9 November 2026 | A data protection impact assessment where processing is “likely to result in a high risk”. | We supply the DPIA inputs: the data path, the sub-processors and the model route. |
| AI Act Art. 6(1), Annex IRead 12 November 2026 | AI that is a medical device, or a safety component of one, follows the Annex I route; those high-risk rules apply from 2 August 2028 (Art. 113). | We don’t build agents with a medical purpose, and we say when a use would cross into it. |
| AI Act Annex III 5(d)Read 12 November 2026 | Lists AI used to evaluate and classify emergency calls and to dispatch or set priority for emergency first response. | Our agents don’t triage calls or patients. |
| NIS2 Annex IRead 12 November 2026 | Healthcare providers are in a sector of high criticality, with supply-chain security duties (Art. 21(2)(d)). | We answer your supplier security review layer by layer, including what is still partial. |
| AI Act Art. 50(1)Read 12 November 2026 | Providers make sure people are told they are interacting with an AI system, unless that is obvious. Applies from 2 August 2026. | Customer-facing agents open with an AI notice in their first message. |
Nine layers around every agent, each with what is partial.
- Access and identityGoogle sign-in, revocable keys, an agent’s rights narrowed by its spec.Partial No MFA of our own and no enterprise SSO; open sign-up is on.
- Separation between clientsEvery record and search index filtered by the signed-in identity; a live agent gets its own containers and database.Partial Logical separation in one shared database, not yet tested independently.
- Hosting, network, encryptionHetzner, Nuremberg; HTTPS with HSTS; no inbound web ports; credentials sealed.Partial Stored data in general is not encrypted at rest.
- Only the data the task needsThe agent reaches only the systems its approved spec declares; secrets are scrubbed.Partial Personal data is not masked before the model.
- The model stepProvider set per deployment; each request admitted and recorded before it is sent.Partial The region is named in your contract, not promised EU-only.
- Untrusted content and testingDocuments are data, never instructions; a safety pack that must all pass.Partial No independent penetration test yet.
- A person decidesWrites wait for a person; a person approves every release; Stop and spend caps.Partial A live agent’s held actions are approved by our operator, not on your screen.
- Every run on recordEvery run recorded step by step, with approvals and who decided.Partial Append-only by convention, not tamper-evident; no client export screen.
- Deletion, export, backupsErasure with a receipt per step; a backup before every release.Partial Export and account deletion by hand, on request; backups not encrypted.
A timeline up front, a fixed price after one call.
Most teams start with a readiness review of the process they have in mind, or go straight to a pilot when the use case is clearly outside the high-risk list. We don’t publish prices: the proposal after a short scoping call gives a fixed price for your scope, in EUR, excluding VAT.
All engagements and what moves the priceBook a callPrice on request
AI Act and GDPR readiness for one use case.
What’s in itBook a callPrice on request
One agent on one process, tested and approved.
What’s in itQuestions from health and life sciences
Is an AI agent that drafts patient letters a medical device?
Not when it drafts from the clinician’s own note, adds no diagnosis or treatment advice and a doctor signs every letter. Whether software is a medical device depends on its intended purpose under the Medical Device Regulation; if a use would add clinical judgement, we stop and say so.
Can health data go to a language model?
Only with a lawful condition under GDPR Art. 9, a DPIA, a DPA and the model route named in your contract. Personal data is not masked before the model today, so the data path is drawn to leave out anything the task doesn’t need.
Do you triage patients?
No. Annex III 5(d) lists the evaluation of emergency calls and the dispatch and priority of emergency response as high-risk, and clinical triage is clinical judgement. Our agents prepare records; clinicians set priority.
Who signs off the agent?
Your medical or quality lead approves the release on its review card, after the tests on your cases. Each new version needs that approval again.
Related pages
Posts on these rules
Bring one process. Leave with a risk class, a pass bar and a plan.
A 45-minute scoping call with the people who would build it. You get a first view of the AI Act risk class, the controls it needs, and whether an agent is the right tool at all.