Industry · Healthcare & Life Sciences

AI in healthcare. Responsibly built, clinically useful.

We build AI applications for hospitals, mental health organisations, health insurers and health-tech startups. No black-box demos, but workflow AI, administration AI and EHR integrations that respect strict healthcare compliance and hold up in practice.

SettingHospitals & university medical centres
SettingMental health & long-term care
SettingPrimary care
SettingHealth-tech & insurers

The Dutch healthcare sector in numbers.

~70
General and academic hospitals
~1,5M
People working in healthcare and welfare
~13%
Healthcare's share of GDP
High-risk
Medical AI under EU AI Act Annex III

Source: CBS Healthcare in figures, NZa monitor, EU AI Act Annex III.

AI in healthcare is no longer pilot territory.

Many healthcare organisations are stuck between two extremes. On one side, software suppliers who put "AI" on their roadmap without anything clinically useful coming out of it. On the other, standalone AI start-ups with impressive demos that never get past information security, the EHR team or the scientific board.

In practice, you need something different: AI that runs inside an existing EHR flow, complies with GDPR, NEN 7510, the MDR and the AI Act, and can be explained to regulators and the medical staff. No black box. No diagnostic claim that won't get through MDR certification.

We build on the non-diagnostic side of the healthcare AI chain: administrative AI, workflow AI, customer service AI and predictive analytics on operational data. We integrate with EHRs such as HiX, ChipSoft, Epic and Cerner. For medical device AI, we work with NEN 7510 partners who handle the MDR route.

Target groups differ greatly in their AI maturity. University medical centres and STZ hospitals often have their own data science department and are looking for a partner to make research models production-ready. General hospitals and independent treatment centres mainly want workflow gains without having to work out how GDPR, NEN 7510 and the MDR interact. Mental healthcare providers and long-term care organisations are looking at reporting AI and client communication. Out-of-hours GP services and pharmacists face triage bots and medication monitoring. Health insurers and diagnostic centres already have their own customer service AI and claims AI projects under way. We match the team's level of maturity: not everything needs to be AI, and not all AI needs to be built by us.

The caveat behind all these use cases: for medical diagnostics or treatment AI, MDR Class IIa or higher is not optional but mandatory. We do not build that device layer under our own name. What we do build is the administrative, workflow and non-diagnostic AI layer that delivers the vast majority of the time savings for healthcare professionals, often the difference between a doctor who spends evenings catching up on paperwork and one who closes their records within their shift.

Use cases we typically build.

Each type of AI application has its own risk classification, its own integration chain and its own compliance track. Four main categories that Appfront works on.

Workflow and administrative AI

This is where healthcare gains the most time in the short term, with the lowest MDR risk. Documentation assistants for doctors, DBC coding, claims checking, extraction of nursing reports, and intake bots for out-of-hours GP services. Non-diagnostic, so no Class IIa MDR track, but with full GDPR, NEN 7510 and AI Act transparency.

Comparable commercial tools are Nuance DAX, Suki AI and Abridge for ambient clinical documentation. We build custom when deep integration with HiX or ChipSoft, a Dutch-language healthcare vocabulary or a specific protocol template is not covered by a SaaS product.

  • Ambient documentationAn LLM that records the consultation and fills in the EHR fields according to protocol, with the doctor remaining ultimately responsible.
  • DBC coding & claimsSuggested codes drawn from the record, with checks before claims go to the health insurer.
  • First-line triage botsSymptom checker for out-of-hours GP services within NHG guidelines, with escalation to a human.
  • Report extractionConverting nursing notes into structured EHR fields in line with Nictiz profiles.
  • Patient flow optimisationScheduling, theatre capacity and readmission risk based on operational data, with no patient-level diagnostics.
  • Adverse event detectionMedication interaction checks using pattern detection on prescription and lab data.

Built on the strictest compliance framework in the Netherlands.

The healthcare AI stack involves more overlapping legislation than any other sector. From the first sprint, we work within the full framework, with a DPIA and risk analysis as deliverables.

EU AI Act — Annex III

Medical AI = high-risk

AI in medical decision-making and triage falls under Annex III. A conformity assessment, a quality management system and post-market monitoring are mandatory. Our pipeline and documentation are set up for these requirements — see also our page on enterprise AI implementation.

MDR — EU 2017/745

Medical Device Regulation

AI that directly influences diagnosis or treatment is a Class IIa, IIb or III medical device. We do not build that device layer ourselves — we work in a partner flow with MDR-certified suppliers and deliver the integration and workflow layer around it.

GDPR + UAVG

Special category personal data

Health data falls under the Article 9 special categories. That means an explicit legal basis, data minimisation and pseudonymisation wherever possible. Our AI architecture isolates PII on a separate layer — where possible, we train models on pseudonymised data.

NEN 7510 + NEN 7512/13

Information security in healthcare

Our development and hosting environment complies with NEN 7510. Access control, logging, encryption at rest and in transit, BIA and risk analysis are standard. See our page on NEN 7510-compliant software for the details.

WGBO + Wkkgz

Treatment agreement and quality

The WGBO governs record-keeping, information and consent. The Wkkgz requires incident registration and complaint handling. AI features that touch on this, such as AI-generated record notes, get an audit log that a regulator can review.

MedMij + HL7 FHIR + Nictiz

Interoperability

Data exchange with patient PHRs and between healthcare providers follows the MedMij framework and Nictiz HL7 FHIR profiles. We build our integrations to comply with them — no proprietary export formats that break the chain.

Algorithm register

Public transparency

Health insurers and public healthcare providers publish AI applications that affect citizens in the Algorithm Register. Our model cards and governance documentation are written so that this registration can be completed.

Cyber Resilience Act

Security obligations across the product lifecycle

The CRA requires security updates, vulnerability reporting and an SBOM. Our releases ship with an SBOM, dependency scanning and a vulnerability disclosure procedure that supports the CRA deadlines.

Seamlessly connected to the healthcare IT ecosystem.

We integrate with the systems already running in your organisation. No rip-and-replace — our AI layer complements them, uses existing authorisation and respects record permissions.

HiX
ChipSoft EHR
Epic
EHR platform
Cerner
Oracle Health EHR
Nictiz
FHIR profiles
MedMij
PHR exchange
DigiD
Patient identification
Azure / GCP
EU-region hosting
OpenAI / Anthropic
LLM providers

Off-the-shelf vendors where relevant, custom where needed.

Strong commercial solutions exist for imaging AI — Aidoc, Lunit, RetinAI, PathAI, Paige.AI. For clinical documentation there are Nuance DAX, Suki and Abridge. For triage there are Babylon, Mediktor and Ada Health. Dutch players such as Pacmed and Healthplus.ai serve specific niche markets. We recommend exploring that market first before considering a custom solution.

Custom development through Appfront is appropriate for niche applications without a commercial product, deep EHR integration where SaaS integrations fall short, multi-system orchestration (EHR + lab + radiology + patient portal), or a custom LLM integration on internal clinical data in a research context. For health insurers, we build white-label customer service AI with explainability layers that support the Algoritmeregister (Dutch algorithm register) format; for independent treatment centres (ZBCs), we build patient portals in which an AI assistant supports the appointment flow without giving medical advice itself.

We make the choice between SaaS and custom in the validation phase. Sometimes our advice is to run an existing tool for six months first before custom AI makes sense — only once the actual workflow and data quality are clear is it wise to build your own model around it. We don't always sell our own build; sometimes integrating an existing vendor into your EHR is the most sensible first path. For broad AI adoption across the whole organisation, we refer clients to AI literacy training and enterprise AI implementation.

From use-case validation to go-live.

An AI project in healthcare has its own rhythm: more compliance work up front, more clinical validation afterwards. Five phases that we go through for every project.

01 · Use case

Validation and risk classification

A workshop with IT, medical staff, the DPO and the privacy officer. Outcome: a chosen use case, an AI Act risk classification and a first draft DPIA assessment.

02 · Architecture

Data flow and integrations

EHR integrations via FHIR, a pseudonymisation layer, model selection (in-house training or an LLM via an EU region). Security design review with information security.

03 · Build

Sprints with clinical tests

A working increment every sprint. Clinical validation with a small group of doctors and nurses. For every model: a model card, an evaluation set and a fairness check.

04 · Pilot

Small-scale clinical pilot

One ward or outpatient clinic runs live. Complete audit log and an incident procedure in place. Results reported to medical management.

05 · Maintenance

Post-market monitoring

Model performance measured continuously, drift detection active, annual re-evaluation. Updates follow the CRA procedure with SBOM and CVE monitoring.

In every phase there is a fixed touchpoint with information security, the medical staff and, where applicable, the scientific council. Compliance is not a report that appears at the end, but a continuous trail of DPIA updates, model evaluations and governance decisions that grows with you. When we hand over to your own team, we deliver the documentation, runbooks and monitoring dashboards, so your information security officer and your operations team can carry on independently without any remaining dependency on Appfront.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

What we typically build in healthcare AI.

Workflow · Hospital

EHR documentation assistant

An LLM records the consultation, fills in record fields according to protocol and respects the WGBO consent rules. The doctor remains ultimately responsible and signs off.

HiX / Epic
EHR integration
EU region
model hosting
Patient flow · University medical centre

Predictive readmission risk

A prediction model on anonymised record data that helps theatre planners with triage and readmission risk. With fairness monitoring per patient group.

FHIR feed
data source
Algorithm register
publication-ready
Insurer & primary care

Customer service and claims AI

White-label customer service AI for a health insurer, AI checks on DBC claims, and a triage bot for a GP out-of-hours service within the NHG framework.

GDPR art. 9
compliance layer
NEN 7510
infosec baseline

Research, education and the university medical centre context.

At university medical centres, RIVM and ZonMW-funded research projects, AI projects have a dual dynamic. The research side supplies the idea, the model and the evaluation set. The production side has to bring that model into an EHR or customer system, with the chain of evidence that regulators and the medical staff need. Many promising research models fail at exactly that gap: the step from a Jupyter notebook with good AUC scores to a production-ready system that complies with GDPR, NEN 7510 and the AI Act.

We bridge that gap. For research institutions, we build the production layer around a model, including pseudonymisation, audit log, model card, evaluation pipeline and governance documentation. For publication-oriented research, we provide reproducibility tooling and data management plans that support the FAIR principles. For education and internal knowledge transfer, we work with your own data science department: no lock-in, and no vendor handover that cripples your research team.

Drug discovery and genomics are a separate branch with their own toolchain. We don't build our own platform there, but we do build custom data pipelines, LLM integrations on public knowledge sources and integrations with laboratory instruments. For precision medicine applications in a university medical centre, we usually work with a biotech or pharma partner who provides the scientific validation.

A recurring theme at research institutions is FAIR data discipline: findable, accessible, interoperable and reusable. AI projects that don't meet it run into trouble later at publication, peer review or follow-on funding. Our pipeline records metadata, model versions and evaluation runs so that an external party, such as a scientific committee, a regulator or the grant provider, can reconstruct the chain. For health-tech startups heading towards a Series A, this is the difference between a credible due diligence and a red flag in the tech report.

How AI in healthcare is evolving.

Nictiz — Strategic Agenda 2025

"Deploying AI in healthcare requires early assessment against the GDPR, the MDR and the AI Act. Interoperability via HL7 FHIR is a prerequisite for scalability."

Federation of Medical Specialists — AI Vision

"AI applications that support decision-making require clinical validation, transparent model documentation and a clear division of responsibility between supplier and healthcare professional."

Dutch Data Protection Authority — regulator

"Processing health data for AI applications requires a DPIA and strict compliance with the principles of lawfulness, purpose limitation and data minimisation."

Answers for healthcare executives exploring AI.

Questions we often hear from CMIOs, CIOs, privacy officers and business developers in healthcare.

Which healthcare AI use cases are realistic to build now?
In the short term, most of the value lies in administrative and workflow AI: documentation assistants for doctors, DBC coding and claims checking, report extraction for nurses, and triage bots in primary care. For medical imaging and clinical decision support, strong commercial products exist, so there we first recommend a market scan. For deep EHR integration and multi-system orchestration, we build custom solutions.
Is an MDR process always needed for healthcare AI?
No, not always. Only when the AI directly influences medical diagnostics or treatment does the product become a Class IIa, IIb or III medical device under the MDR. Workflow AI, administrative AI and non-diagnostic support usually fall outside it. For applications that do fall under the MDR, we work in a partner flow with MDR-certified suppliers; we do not build the device layer ourselves.
How do you handle the GDPR and special category personal data?
Health data falls under the Article 9 category of the GDPR. Our architecture isolates PII in a separate layer with strict access control. Where possible, we train models on pseudonymised or synthetic data. As standard, we deliver a DPIA, a data processing agreement and model cards for the DPO. LLM calls to external providers go through EU-region endpoints with training opt-out.
Does your development process comply with NEN 7510?
Yes. Our development and hosting environment operates within the NEN 7510 framework: access management, logging, encryption at rest and in transit, and periodic BIA and risk analysis. For clients who are certified themselves, we supply the evidence they need for their audit. See also our service page on NEN 7510-compliant software.
How do you integrate with HiX, ChipSoft, Epic and Cerner?
We prefer to work with the official FHIR APIs and the Nictiz profiles. For HiX/ChipSoft, we connect through their integration portfolio; for Epic, through App Orchard / Vendor Services programmes; for Cerner, through the Oracle Health developer environment. Where the standard APIs fall short, we build a middleware layer rather than direct database integrations that would break the EHR's lifecycle.
What determines the cost of a healthcare AI project?
Four factors: scope (a single use case versus multi-system), compliance weight (MDR route versus GDPR/NEN 7510 only), integration depth (a single EHR integration versus EHR + lab + radiology + portal) and model requirements (out-of-the-box LLM versus custom training on clinical data). We work with a fixed budget per sprint and, after the validation phase, give you a concrete price for the complete build, including the compliance work.
How does this fit within the AI Act's high-risk requirements?
AI in medical decision-making and triage falls under Annex III of the AI Act as high-risk. This requires a conformity assessment, a quality management system, transparent documentation and post-market monitoring. Our pipeline is set up for these requirements: model cards, evaluation sets, drift monitoring and governance documentation are standard deliverables. For the broader implementation approach, see enterprise AI implementation and AI literacy training for your team.
How does AI fit with existing hospital budgets and reimbursement?
Many care providers fund AI from a combination of the regular IT budget, ZonMW or Health~Holland grants and innovation funds from health insurers. We help you make scope choices so the project fits within a single financial year, and we provide a business case that can feed into a grant or innovation application. We don't charge a percentage of savings and we don't offer a "value-based" rate.
What do you and don't you do in healthcare AI?
What we do: workflow AI, administration AI, customer service AI, predictive analytics on operational data, EHR integrations via FHIR, patient portals with AI assistance, and LLM integrations on internal clinical data in a research context. What we don't do on our own: medical-device AI that directly influences diagnosis or treatment and therefore falls under MDR Class IIa/IIb/III. For that layer, we work as a partner of an MDR-certified supplier. For medical imaging, we advise you to first explore the existing market (Aidoc, Lunit, RetinAI, PathAI) before considering a custom build.
How do we arrange the handover to our own IT team?
From the start, we work alongside your information security officer, EHR management team and data stewards. At the end of the project, we deliver runbooks, monitoring dashboards, model cards and a maintenance procedure that fits within your existing ITIL or agile workflow. We don't create lock-in on proprietary tooling: models, data pipelines and infrastructure-as-code templates remain yours.

Talk to us about AI in your healthcare organisation?

A half-hour exploratory conversation with your CMIO, CIO or business developer. We listen to the use case, outline the compliance route and give you an initial direction. No obligation.

Edit content