Insurance is a business of documents, decisions and risk. We build AI that settles claims faster, sharpens underwriting and catches fraud — while keeping every decision explainable and compliant.
Last updated: 11 June 2026
Every claim, quote and renewal is a decision made from data — and much of it still passes through manual hands. AI is built for this: reading claim documents, scoring risk consistently and flagging the patterns that signal fraud, all with a clear, auditable rationale.
We build for a regulated world — explainable models, strict data handling and EU AI Act & GDPR alignment from the start.
Read, triage and pre-assess claims so straightforward ones settle in hours, not weeks.
Consistent, explainable risk models that support faster, fairer underwriting.
Spot suspicious claims and patterns early, with fewer false positives for investigators.
Extract and check data from policies, forms and correspondence automatically.
Grounded assistants that answer policy and claim questions from your own data.
Portfolio, churn and loss-ratio analytics built on your own historical data.
We map your highest-value use cases, data and compliance constraints.
By the second call you get a working prototype on your use case — not a spec.
We harden, integrate and ship with audit trails and access control in place.
We watch accuracy and drift so decisions stay reliable and defensible.
Walk through a mid-sized Dutch verzekeraar or volmacht and the cost is rarely in the headline product. It sits in the handling: a schadeafhandeling team re-keying the same loss-adjuster report into three systems, an acceptance desk that gives two near-identical risks two different premiums depending on who picked up the file, and a special investigations unit drowning in alerts that turn out to be a customer who simply moved house. None of that shows up cleanly on the P&L, but it is exactly where claims cycle time stretches, where the loss ratio drifts, and where the combined ratio quietly creeps past 100%.
AI for insurance in the Netherlands is worth doing only when it moves one of four numbers: claims cycle time, fraud caught net of investigation cost, underwriting consistency, and the loss ratio. Everything below is framed against those, because a model that does not touch one of them is a science project, not a business case. We are an AI engineering partner, not a marketing agency — the deliverable is a working system wired into your policy and claims platforms, owned by your team.
The slowest part of a claim is almost never the decision — it is the waiting. A first-notification-of-loss email lands, sits in a queue, gets read, gets summarised by hand, then routed to whoever has capacity rather than whoever has the right expertise. A claims-automation pipeline collapses that. The moment a FNOL arrives, a document-understanding model reads the email, the attached PDF estimate, the photos and the police report, extracts the structured fields your claims system needs, checks them against the policy, and proposes a route: straight-through settlement, fast-track with a light human check, or full adjuster review.
The operational change is concrete. Simple, low-value motor and travel claims that used to take days of back-and-forth can be triaged in minutes and a meaningful slice settled the same day. Across the industry, well-scoped claims automation typically takes 30–50% out of average cycle time on the high-volume, low-complexity segment, while the genuinely complex claims get to a senior adjuster faster because the queue is no longer clogged with easy ones. Your handlers stop being data-entry clerks and start spending their time on the files where judgement actually matters.
This sits squarely on our AI implementation and data engineering work — the unglamorous plumbing of getting clean, governed data out of your policy administration and into a model is usually where the real effort goes.
Two underwriters, the same risk, two different answers. That inconsistency is invisible day to day but expensive in aggregate — it shows up as adverse selection, as risks you priced too cheaply staying on the book and risks you priced too dearly walking to a competitor. The point of AI in underwriting is not to remove the underwriter; it is to give every risk the same disciplined first read, so human judgement is spent on the exceptions instead of the routine.
A risk-scoring model trained on your own portfolio and claims history produces a consistent, explainable score for each application. For standard SME and personal lines, that means a large share of clean risks can be auto-accepted within an agreed envelope, referrals are triggered by genuine risk signals rather than by whoever happens to read the file, and the acceptance team handles materially more volume without growing headcount. The measurable outcomes are a tighter, more consistent acceptance band, fewer mispriced policies entering the book, and — over a renewal cycle or two — a loss ratio that reflects deliberate pricing rather than drift.
Because acceptance and pricing decisions affect people, explainability is not optional here. We favour models that can justify a score — which factors pushed it up, which pulled it down — so the decision stands up to an auditor, a regulator and the policyholder who asks why. The same machine-learning and forecasting techniques behind this are described under machine learning, and the portfolio-level analytics that sit on top — churn, retention and loss-ratio modelling — build on the same governed data foundation.
Most fraud models fail not because they miss fraud but because they cry wolf. An SIU that gets a hundred alerts to find three real cases stops trusting the system. The goal is the opposite: fewer, sharper alerts where the hit rate is high enough that investigators chase signal, not noise. That means combining classic rules with anomaly detection and network analysis — spotting the same bank account across unrelated claims, the repair shop that appears far too often, the claim filed suspiciously soon after a policy upgrade, the photo metadata that does not match the stated date of loss.
The outcome to watch is twofold: more genuine fraud caught and fewer false positives per real case. When the alert quality goes up, the same investigation team works through a higher-value caseload, recoveries rise, and — just as important — honest customers stop being delayed by a manual fraud check they never deserved. Insurance fraud quietly inflates the loss ratio for everyone; tightening it is one of the few levers that helps both margin and the customer at the same time. In the Netherlands this has to be done carefully: profiling and automated decisions touch the AVG and the EU AI Act head-on, so models must be explainable, fair across groups, and never act as a hidden blocklist. We build the audit trail and human-in-the-loop checkpoints in from day one, not as an afterthought when the toezichthouder asks.
Insurance is, underneath everything, a document business — polisvoorwaarden, schade-aangiftes, medical reports, broker correspondence, KvK extracts, repair invoices. Most of it still gets read and typed by a person. Intelligent document processing reads these end to end: it classifies the document, extracts the fields, validates them against your records, and flags only what is genuinely unclear for a human to resolve.
The effect compounds. Cut manual document handling and you do not just save the keystrokes — you remove the transcription errors that cause downstream rework, you speed up every claim and every quote that depends on those documents, and you free experienced staff from work that wastes them. For a busy claims or acceptance team, taking the bulk of routine extraction off their plate is frequently the single change that frees the most capacity, and it is usually the fastest to show a return. It also feeds everything upstream: cleaner extracted data makes the claims models, the risk scores and the fraud flags all more accurate.
Where the documents are dense or visual — damage photos, scanned forms, handwritten notes — this pulls in computer vision alongside the language models that handle document understanding and grounded assistants. For policyholder- and broker-facing help, a retrieval-grounded assistant answers policy and claim questions from your own documents only, so it cannot invent cover that does not exist.
Two realities shape every AI build in Dutch insurance, and both are easy to get wrong. The first is regulation. Underwriting, pricing and fraud decisions are high-stakes uses under the EU AI Act, and they process personal — sometimes special-category — data under the AVG. That is not a reason to avoid AI; it is a reason to build it properly: documented data lineage, explainable decisions, human oversight on anything that affects a customer, bias testing across groups, and the option of private or on-premise models where sensitive data cannot leave your environment. Done this way, compliance becomes a selling point with brokers and customers rather than a brake.
The second is language. Your claims, your correspondence and half your customers are in Dutch; the rest is in English. Models that only really work in English quietly degrade on Nederlandstalige documents and conversations. We build and test bilingually because that is the market you actually operate in — and being a Dutch firm in the province of Utrecht, that is the market we know.
We are deliberately a boutique, senior-led shop — the opposite of both the big enterprise consultancies and the web agencies that rebranded as "AI" over a weekend. The people who scope your problem are the people who build it, and the solution is structured so your team ends up owning it. The path is fixed and transparent, with pricing excl. VAT:
Behind this sits a track record of 13 delivered case studies across healthcare, computer vision, NLP, forecasting and predictive maintenance — client names confidential, the engineering discipline the same. The forecasting and anomaly-detection work that powers claims and fraud, and the document and NLP work behind intelligent processing, are the same capabilities, pointed at insurance.
If claims cycle time, your fraud hit rate, underwriting consistency or the loss ratio is where the pain sits, that is exactly where we start. See our case studies, the full picture of AI consulting in the Netherlands, or transparent pricing — then book a free consultation and we will map a compliant path to value on your actual numbers.
Yes — explainability matters in insurance, so we favour models that can justify a risk score or claim decision for regulators, auditors and policyholders.
Yes — we connect to your policy administration, claims and CRM platforms through secure APIs.
We tune models on your real patterns and add context so investigators focus on genuine fraud, not noise.
Yes — GDPR-compliant handling, access control and private models where sensitive data can't leave your environment.
We do — Crux Digits, a boutique applied-AI firm in the Utrecht region, founded in 2022. We work with Dutch and EU insurers, MGAs and brokers on claims automation, underwriting and fraud detection, in English and Dutch. A small senior team builds the system with you, no offshore hand-off, and you own the source code and IP.
We begin with a short paid discovery to scope the use case and check your data, then deliver at a fixed price agreed after that. Because underwriting, pricing and fraud decisions are high-stakes, we build for AFM conduct rules and the duty of care, keep decisions explainable, and align with the EU AI Act and GDPR from day one — with human oversight and audit trails built in, not bolted on later.
Find at-risk customers and the levers to keep them — so marketing targets the right cohorts instead of spraying spend.
Turn demand signals into a production plan, smooth mould-change losses and defend market position — for a €235M footwear manufacturer.
Tell us where the cost and delay sit — we'll map a compliant path to value in a free consultation.
Book a free consultation →