Service · AI development

AI intake integration.

Replace traditional intake forms with a conversational AI flow that understands context, asks follow-up questions at the right moment, and delivers structured data for your EHR, CRM, HRIS or case management system. For healthcare, legal, HR, insurers, government and B2B sales. Not a standalone chatbot experiment — a clearly scoped intake layer, in your brand, with GDPR, NEN 7510 and the AI Act in order from the start.

Conversational intakeEHR/CRM/HRIS integrationJSON schema outputAI Act compliantIndustry fine-tuning
Fabian van Dijk · Business developer · fabian.vandijk@appfront.nl

A good intake is not a form — it's a conversation.

Most intake flows in Dutch organisations still look the same: a long questionnaire, locked down with dropdowns and mandatory fields, forcing the user to squeeze into the boxes of the underlying system. A patient with a complaint, a client registering a case, a new employee: all pushed through the same rigid funnel. The result: half-completed forms, incorrect classifications, a second-line employee who still has to phone, and data that has ended up in the wrong field.

An AI intake does it differently. Users describe in their own words what is going on, by typing or speaking, and a language model translates that into the structure the underlying system needs. Is information missing? The AI asks a follow-up. Does something not add up? It asks for confirmation. Once the conversation is finished, a clean JSON payload comes out with exactly the fields that the EHR, CRM, HRIS or case management system expects. That is the core of what we build every day in AI development.

What makes it work: we don't build a general-purpose chatbot that you then try to trim down. We build a clearly bounded intake layer based on the JSON schema needed at the back end. The AI knows which questions to ask to get that schema filled, and stops as soon as everything is in order. No open ends, no improvising with advice the model is not allowed to give. Industry knowledge, such as anamnesis protocols, complaint classifications and types of legal question, is added through a targeted RAG layer. Fast, cheap to run, and explainable to the regulator.

Three typical forms of AI intake.

In our experience, AI intake integrations fall roughly into three categories, each with a different emphasis on compliance, integration depth and user experience. Which form suits your organisation depends on the kind of decision made after the intake and the sensitivity of the data collected.

Embedded · in your portal or app

Embedded intake widget

A conversational intake component that you embed in an existing website, customer portal or mobile app. Users stay in your branding and are never sent to a third-party domain. Suitable for insurers who want to take in damage claims, municipalities with a reporting flow, or B2B sales teams who want to automate an initial needs assessment. The output lands directly in your CRM, ticketing system or case management system.

JS widgetSSOCRM integrationYour own branding
Voice intake · phone or in-app

Voice-driven intake

Users call or speak in an app, the AI listens in and structures the conversation into the fields of your intake schema, and on the other end the employee receives a clean summary instead of a raw call recording. For healthcare triage, municipal 14-numbers, claims lines and support call centres where volumes are high. We connect a speech recognition engine to an LLM layer that turns the transcript into a structured intake record.

Twilio / GenesysReal-timeDiarisationSentiment
Domain-specific · healthcare, legal, HR

Deeply integrated domain intake

An intake layer built entirely for a single domain: a GP anamnesis app that asks questions according to the NHG guideline, a law firm intake that handles legal area classification and conflict checks up front, or an HR onboarding flow that lands directly in the HRIS. For healthcare contexts, we work within the framework of our appointment portal for clients with NEN 7510 compliance from the first sprint.

EHR integrationRAG on guidelinesNEN 7510Conflict check

What you have at the end of the project.

A production-ready AI intake integration that fits your people's workflow and your IT architecture, plus everything needed to manage, audit and extend it. The code, prompts, schemas and models are yours. We only handle ongoing management if you want us to; if it suits you better, you can carry on elsewhere tomorrow.

  • Conversational intake layerA chat or voice flow that guides the end user through the intake, in your house style, with a tone and language level suited to your audience.
  • JSON schema-driven outputA formal JSON schema describing the structure of a valid intake, with structured outputs / function calling that guarantees the output conforms to it.
  • RAG layer on your guidelinesRetrieval-augmented generation on your organisation's protocols, formularies or HR handbooks, so the model falls back on established sources.
  • Downstream integrationIntegration with your EHR (HiX, Epic, ChipSoft, CGM), HRIS (AFAS, Workday, BambooHR), CRM (Salesforce, HubSpot, Dynamics) or case management system.
  • Human-in-the-loop layerFor intakes that prepare a risk decision, such as medical triage or legal classification, an explicit human review step, compliant with the AI Act.
  • Full audit trailEvery intake becomes a traceable record: which questions, which answers, which model and version, which RAG sources, and which human actions.
  • EU residency and privacy by designLLM inference via Azure OpenAI in Western Europe, Claude via AWS Bedrock in Frankfurt, or a self-hosted model in a Dutch cloud.
  • DPIA, AI Act classification and TIAThe compliance documentation required by law, ready for your data protection officer and for any audit by the supervisory authority.
  • Your own prompts and version controlPrompts, system messages, few-shot examples and evaluation cases in your own git repository, with code review and automated tests.
  • Management contract (optional)Monitoring of intake quality, model updates, prompt evaluations against a growing test set, and further development based on patterns seen in practice.

Who we build AI intake integrations for.

Eight patterns where AI intake adds value in practice. If you recognise your organisation or use case in one of these profiles, we'd be happy to talk further, even if you're still unsure whether your own integration is the right choice instead of an off-the-shelf package.

Healthcare

Medical history intake for GP practices and mental health care

Patients describe their complaint in their own words through a secure patient portal flow; the AI structures it into an anamnesis that follows Dutch GP (NHG) guidelines, with red flags flagged in advance. For the GP or mental health practitioner, the consultation is shorter, better prepared, and the SOAP "S" field in the EHR is already filled in. For sensitive groups, such as paediatrics, psychiatry and oncology, we explicitly discuss which questions an AI may ask and which must first be led by a person. Also read about our approach to an appointment portal for clients.

Legal

Case classification for law firms

A prospective client describes their case; the AI classifies the area of law (family law, employment law, corporate law, criminal law), assesses the urgency, runs an initial conflict check against the firm's system and determines which partner is best placed to take on the case. The client gets targeted feedback faster, and the firm loses less time on cases outside its practice. For the high-risk tier under the AI Act, we build in a human review step for every case that is actually accepted.

HR

Onboarding flow for new employees

A new employee goes through a conversational onboarding in which the AI asks the full set of questions HR would otherwise collect via Excel or emails: address details, BSN (citizen service number), holiday arrangements, hardware preferences, travel expense claims. The data lands directly in the HRIS, IT provisioning, the travel expenses tool and payroll. For the employee, it feels like a single conversation.

Sales

B2B customer onboarding and discovery

A business customer describes their need to the AI, which asks the discovery questions a sales engineer would otherwise ask (user numbers, integration requirements, decision-making process) and delivers a complete brief instead of a raw lead.

Insurance

Claims intake for insurers

A policyholder reporting a collision, claiming for fire damage or reporting a stolen phone describes the incident in an AI-led intake conversation. The AI asks follow-up questions on the facts, accepts photos via an upload step, classifies the type of damage, and raises an initial fraud signal where the factors warrant it. The claims handler picks it up with a largely completed file.

Government

Complaint and report intake for municipalities and utilities

Residents who report an issue, file a complaint or submit a request do so through a conversational flow that gives staff far better groundwork than the current form. The AI categorises, assigns to the right department, and asks follow-up questions.

Hospital

Pre-surgery patient intake

Patients awaiting surgery complete a digital intake beforehand that gathers the information the anaesthetist and theatre team would otherwise have to elicit verbally: medication, allergies, previous procedures, comorbidity, mobility. The AI adapts to each answer and delivers a structured pre-surgery checklist that lands directly in the EHR alongside the existing record management system.

Education / public sector

Application and enrolment intake

For universities of applied sciences, municipal services or education enrolments: an AI flow that guides the applicant through all the relevant questions and enters the application into the back-office system in a structured way. With attention to accessibility, B1 language level, and alternative ways to complete it for those who would rather speak to a person.

Which LLM and stack fit.

We don't pick a favourite model up front. We choose only once we know the use case, the data category, the languages and the privacy requirements. A short overview per candidate, as we would go through it with you in a first session. We work vendor-independently: you own the prompts and the schema, and we can swap the model tomorrow if the market or pricing gives reason to.

EU residency · enterprise · Azure

Azure OpenAI (GPT-4-class)

For healthcare, government and financial clients, often the practical choice: EU residency in Western Europe, integration with Azure AD, blocking of data use for training, and an existing Microsoft contract that simplifies procurement. Strong support for structured outputs and function calling.

Bedrock EU · Claude · strong with long context

Anthropic Claude via AWS Bedrock

For clients on AWS who want to run with strict privacy: Claude models via Bedrock in EU regions, with strong performance on Dutch-language context and a conservative response profile that works well in compliance-sensitive intakes.

Self-hosted · open-source · zero cloud

Mistral / Llama in your own environment

For clients who, for policy reasons, don't want a US cloud provider in the stack (academic hospitals, law firms, defence contexts), we run an open-source LLM in your own cloud or on-premise. Lower out-of-the-box performance, which can be compensated for with fine-tuning and RAG.

RAG layer · vector search · evaluation

RAG and evaluation pipeline

For industry knowledge, we use a vector database (pgvector, Qdrant or Weaviate) with the organisation's own documents as the source. On top of that, an evaluation pipeline that automatically tests the intake prompt against a growing set of realistic cases, so a prompt change doesn't quietly break one type of intake.

Voice layer · Twilio · WebRTC

Voice stack for phone and in-app

For voice intake, we connect a speech recognition engine (Deepgram, AssemblyAI or Azure Speech) to a telephony or WebRTC layer (Twilio, Genesys, Amazon Connect) and to the LLM. The AI listens in, structures the conversation, and can route to a person at the right moment.

Front end · widget · SSO

Embedded widget and authentication

We build the intake component as a lightweight JS widget that fits into your portal, with an SSO integration (Azure AD, DigiD, eHerkenning, Auth0). For mobile, we build a native React Native or Flutter component.

How an AI intake project works.

1

Introduction, use case and schema definition

A conversation in which we pin down exactly which intake flow this will be: who the end user is, what data must come out at the end, which back-end system it lands in, and which decisions follow the intake. Together we draw up the JSON schema that describes the output — often an eye-opener for organisations that have never seen their existing intake form laid out on a drawing board in this way. Sometimes a short advisory phase is the right first step; see our page on AI development.

2

DPIA, AI Act classification and data architecture

For intakes that involve personal data — which is by definition the case in healthcare, legal, HR and government — we start the DPIA in parallel with the build. We classify the intake under the AI Act and build the associated obligations directly into the design. Chat flow, retention policy, encryption, sub-processor choice and LLM residency end up in a single architecture document that your DPO can also read. We align with an existing ISMS.

3

Prompt design and RAG build

The system prompt and the structured output definitions take shape based on the schema from step 1. We build the RAG layer on the organisational knowledge the model needs in order to answer in the same style your organisation would. From day one, prompts and evaluation cases live in a git repository with code review.

4

Building in sprints with continuous evaluation

A working build on staging after every short sprint. We build incrementally: first the basic flow with structured output, then the RAG layer, then the downstream integration, then the human-in-the-loop step, then the audit trail. With every prompt or schema change, an evaluation set of realistic cases runs and scores per cohort — a change that improves overall quality but worsens results for one target group is something we want to see before release. An intake that silently fills in the wrong fields is worse than one that occasionally looks a little slower.

5

Rollout, training and further development

Phased rollout: first a pilot group with a dual intake (AI and the existing form side by side), then a broader rollout, then a full switchover. Brief training for the staff who need to pick up intake records, a guide for the end user, and ongoing maintenance for model updates, prompt iterations and security patches. We monitor intake quality per cohort and keep fine-tuning.

Compliance is not an afterthought.

An AI intake layer often touches sensitive data and, on the way to a decision, has an impact on a person — a patient, a client, an applicant. For healthcare, legal and public-sector applications, a well-built AI intake is by definition also a well-documented, GDPR-compliant and AI Act-compliant integration. We handle that layer from the first sprint, not as a closing paperwork exercise.

GDPR & DPIA

Every intake produces personal data, often sensitive: a healthcare complaint, a legal case, a BSN (Dutch citizen service number), a conversation report. We carry out a targeted DPIA, put data processing agreements in place with sub-processors (LLM provider, hosting, vector DB), and build retention and deletion policies into the intake flow itself.

AI Act & high-risk classification

The AI Act has been in force since 2026. A general B2B discovery intake is low-risk, but as soon as the intake becomes input for medical triage, legal decisions or HR assessments, the system falls into the high-risk category. We carry out the classification up front and take care of the risk management documentation, data quality requirements, transparency layer and human oversight mechanism.

NEN 7510 for healthcare

For healthcare organisations, we follow the NEN 7510 control set from the start: access management, logging, encryption, supplier management. We align with the existing ISMS. See also our approach to NEN 7510-compliant software.

GDPR for young people

Intakes involving minors — youth psychiatry, school care, youth services — are subject to additional rules. For under-16s, consent from a parent or guardian is generally required; that consent should be part of the intake flow itself, without making it needlessly burdensome for the young person.

eIDAS and digital signatures

For intakes that end with a formal declaration or consent — an insurance claim, an onboarding contract, consent for sharing medical data — we build in an eIDAS-compliant signature step. This uses DigiD or eHerkenning for public bodies, or an advanced signature for B2B.

Audit trail & human oversight

Every intake has a traceable provenance: which questions, which answers, which model and version, which RAG sources, which human actions, and what was sent to the underlying system. For high-risk applications, we add explicit human-in-the-loop steps before the outcome affects any decision-making.

Frequently asked questions.

What clients typically want to know before we start on an AI intake integration.

Why not just use an off-the-shelf chatbot from a vendor?
For general information bots, a SaaS product works perfectly well. An AI intake is different: it has to match the schema of your underlying system exactly, the protocols of your sector and the tone of your organisation. You also want to keep the prompts, evaluation cases and RAG sources under your own control, because they directly affect the data that ends up in your EHR, HRIS or CRM. With a closed vendor, you never get full visibility into those. A custom intake also produces a far better audit trail — relevant as soon as the AI Act or the regulator comes knocking.
Which LLM do you use?
That depends on the use case, data category, languages and privacy requirements. For most healthcare, government and financial clients, we land on Azure OpenAI in Western Europe or Claude via AWS Bedrock in Frankfurt. For clients who, as a matter of policy, don't want a US cloud provider, we run Mistral or Llama in your own environment. We work vendor-independently: you own the prompts, the schema and the evaluation set, and we can swap the model if price or performance gives us reason to.
How do you prevent the model from hallucinating during an intake?
Three layers. First, we use structured outputs / function calling, so the model must shape its output according to a predefined JSON schema. Second, we use RAG on the organisation's official knowledge, so that answers and follow-up questions are grounded in approved sources rather than the model's training data. Third, an evaluation pipeline with realistic cases validates every model or prompt change. For high-risk applications, a human review always sits in between before the intake outcome affects any decision-making.
Can an AI intake process patient or client data?
In principle, yes, provided the right compliance layer is in place: a DPIA, a suitable data processing agreement, sub-processors in line with your GDPR policy, and LLM inference within the EU. For some healthcare organisations we opt for on-premise or dedicated cloud for broader reasons — for example, with particularly vulnerable groups. We advise on what fits for each project and work with your data protection officer, using NEN 7510 controls as a framework. Something similar applies to law firms, given professional secrecy.
How does an AI intake handle people who would rather talk to a person?
Every AI intake we build has a visible "I'd rather talk to a person" route. At any moment, users can step out of the flow and reach a human member of staff — by phone, a callback request or a handover to live chat. It isn't an option but a design requirement: some groups don't suit an AI flow, and forcing them into one is a bad idea. The handover route is also an AI Act transparency requirement for high-risk applications.
Can you integrate the intake with our EHR, HRIS or CRM?
Yes. We have experience with HiX, Epic, ChipSoft, CGM and Cura on the EHR side; AFAS, Workday, BambooHR and Personio on the HRIS side; and Salesforce, HubSpot and Dynamics on the CRM side. The integration typically runs via HL7 FHIR for healthcare, a REST API for modern systems, or a SOAP layer for older ones. If your system is built in-house, we integrate against a custom API. See also our page on case management system development.
How do you measure the quality of an AI intake?
We build an evaluation pipeline with realistic cases, often based on real, anonymised historical intakes from your existing form. Every prompt or schema change automatically runs against that set and produces a score per cohort: how often the classification is correct, how often the structured fields are filled in correctly, how often the AI escalates to a human and how often that is justified. Alongside these quantitative measurements, we look qualitatively at failures, cases where the AI silently fills in the wrong field, because those cause the most harm.
How do you handle the AI Act for our intake?
We start with an AI Act classification: low risk, limited transparency obligations, or high risk (intake as input for medical triage, legal decisions, credit refusals, HR assessments). For high-risk applications we take care of the mandatory risk management documentation, the data governance requirements, the transparency layer, the human oversight mechanism, the logging the regulator is entitled to inspect, and post-market monitoring after go-live. We align with an existing ISMS.
Who owns the code, prompts and data?
You do. We hand over the complete source code, prompts, JSON schemas, evaluation cases and build pipelines. The RAG sources and intake records live in your cloud or with a hosting partner of your choice. No lock-in on the technology or on the choice of LLM: we earn our money from good work that lasts, not from clients who are stuck with us.
What are the typical pitfalls of an AI intake project?
Three patterns. One: the schema isn't properly thought through, so the AI has to produce data that the back end doesn't know what to do with. Two: there is no evaluation pipeline, and the first prompt change silently breaks an edge case that nobody notices until a user complains. Three: the human side of the workflow isn't taken into account. The AI now delivers neat intake records, but the team that has to pick them up still works to the old rhythm. Our sprint approach addresses all three structurally.

Talk to us about your AI intake.

A no-obligation introductory conversation. Tell us who will use the intake, which system the record ultimately lands in and what compliance context applies. We'll think along with you and be honest about whether a custom integration is the right answer. Sometimes the right first step is a short advisory phase via AI development; sometimes it's time to build straight away. Prefer email: fabian.vandijk@appfront.nl.

Edit content