AI consulting in the Netherlands — your questions, answered

Pick one process with a measurable outcome and start there. Most failed AI projects start too broad. A first audit maps your data, ranks use cases by payback, and produces a roadmap you can act on. You decide afterwards whether to build anything at all.

• The AI Audit & Strategy is fixed at €2,500 and takes 1–2 weeks, ending in a ranked use-case list and a go/no-go recommendation.
• Strong first candidates are high-volume, rule-heavy tasks: invoice processing, quote follow-up, customer-service triage, document handling.
• If no use case is viable yet, we say so in the audit rather than selling you a prototype.

AI plugs into your existing systems through their APIs. It rarely replaces them. We connect to your CRM, ERP, databases and document stores, then build the pipelines that move and clean the data in between. The goal is AI that fits your current stack, not a rip-and-replace.

• Common Dutch integrations: Exact Online, AFAS, e-Boekhouden, Microsoft 365, HubSpot, Salesforce.
• Where no API exists, we fall back to scheduled exports, direct database reads or file-drop connectors.
• Integration work is scoped during the audit, which is what lets the build price stay fixed. See data engineering.

A working prototype on your own data takes 4–6 weeks. The Proof of Concept is fixed at €20,000 and deliberately narrow, which is what keeps it to weeks rather than months. You agree the success metric up front and get a go/no-go recommendation at the end.

• One use case, one data source, one measurable outcome — that is the scope discipline protecting the timeline.
• Scope creep is the usual reason prototypes slip; a fixed price is the mechanism that prevents it.
• More detail in scoping an AI proof of concept.

Yes. Every engagement is fixed-scope and fixed-price, not open-ended hourly billing. An AI Audit & Strategy is €2,500, a Proof of Concept is €20,000, and a Production Launch starts at €50,000. Ad-hoc work is €150 per hour. The prices are published, so there is no quote round.

• All prices exclude VAT. Most Dutch AI agencies publish no prices at all, so comparison normally requires a sales conversation first.
• For context: big-four day rates run €2,000–€3,000, and a typical SME trajectory there lands at €50,000–€150,000.
• Full breakdown on pricing, reasoning in what AI implementation costs.

Usually yes — your own operational data is often the biggest advantage. Transactions, documents, logs and support tickets already describe how your business actually works, which is exactly what a model needs. The first job is confirming that data is reachable and genuinely relevant to the use case.

• The audit checks three things: can we reach the data, is there enough of it, and does it contain the answer.
• Volume matters less than people expect — a few thousand well-labelled examples often beat a million messy ones.
• Your data stays yours, and you own the code and models built on it. See using existing data to train AI.

Probably, if it exists, is reachable and is reasonably consistent. Perfect data is rare and almost never required. “AI-ready” is judged per use case, not as a general grade for your company. Some cases work fine on messy data; others need cleanup first. The audit tells you which.

• Red flags: no historical record, data locked in a system with no export, or labels that different people define differently.
• Green flags: a consistent process, a single system of record, and at least a year of history.
• Walkthrough of the checks in is your data AI-ready.

No. Cleaning is part of the project, not a prerequisite. Waiting until your data is tidy is a common way to never start. We build repeatable pipelines that clean and transform automatically, so the fix holds instead of decaying the moment someone stops doing it by hand.

• Typical work: deduplication, consistent formats, missing-value handling, and joining records across systems.
• Manual one-off cleanups usually decay within months — pipelines are the durable version.
• See data engineering and data engineering for AI.

Usually yes, with a lawful basis and minimised personal data. Using your own business data for your own purposes is generally permitted. The real obligations are documentation, purpose limitation, and keeping personal data out of the model wherever it is not needed. That is design work, not paperwork afterwards.

• Pseudonymise or strip personal fields before training whenever the use case does not require them.
• Employee and customer data both need a documented basis; a processing agreement covers the supplier side.
• The EU AI Act sits on top of GDPR rather than replacing it — both apply. See training AI on company data under GDPR.

AI is the goal; machine learning is one way to reach it. AI means software doing tasks that normally need human judgement. Machine learning is the technique where a model learns patterns from data instead of being programmed with rules. Plenty of useful AI uses no machine learning at all.

• The nesting runs AI → machine learning → deep learning → large language models, each a subset of the one before.
• A rules engine that routes invoices is AI. A model that predicts which invoices will be paid late is machine learning.
• Most SME projects in 2026 use a pre-trained language model, so no training happens at all. See machine learning vs AI.

Often something simpler works: rules, search, or a prompted language model. Custom machine learning earns its keep only when the pattern is genuinely hard to write by hand and you have the data to learn it. Simpler solutions are cheaper to build, faster to ship and easier to explain.

• Use rules when the logic is known and stable. Use machine learning when the logic is in the data but nobody can state it.
• A language model with your documents attached (RAG) solves many cases with no training at all.
• Custom models add permanent cost: monitoring, retraining and drift management. See machine learning for business.

Yes. A model is a snapshot, and the world moves on. It learned the patterns present in its training data, so as customers, prices or processes change, its predictions drift away from reality. This is model drift, not a malfunction. The fix is monitoring plus periodic retraining.

• Drift shows up as accuracy falling slowly over months, which is exactly why it goes unnoticed without monitoring.
• Budget for maintenance from day one — a model is a running system, not a delivered file.
• Retraining cadence follows how fast your data changes; quarterly is a common starting point. See why ML models stop after training.

Yes — deployment, monitoring, retraining and the pipelines around them. Building a model is the smaller half of the work; keeping it accurate under real traffic is the harder part. Production Launch covers that, starts at €50,000, and includes the data engineering the model depends on.

• Included: deployment, logging, drift monitoring, alerting, a retraining routine and documentation.
• Typical production timeline is from 8–12 weeks, depending on how many systems it touches.
• You own the code and models, so you can move the system in-house or to another supplier at any point. See machine learning in production.

Yes, grounded in your own documents so it does not invent answers. We build these with retrieval-augmented generation: the assistant looks up your content at answer time and cites it. Where the stakes are high, a person stays in the loop before anything reaches a customer.

• Typical sources: product documentation, policy documents, past support tickets and price lists.
• A website assistant normally fits inside the €20,000 Proof of Concept scope, delivered in 4–6 weeks.
• Ask any supplier about the refusal policy — a good assistant says “I don't know” instead of guessing. Tell us your case via a free consultation.

RAG lets a language model look up your documents while answering. Instead of relying on what the model memorised during training, it retrieves your content first and answers from that. You need it whenever an assistant has to be accurate about your specific products, policies or prices.

• RAG stays current without retraining: update the document, and the answer updates with it.
• The alternatives are fine-tuning (teaches style and format, not facts) and long-context (simple, but expensive at scale).
• RAG also returns citations, which makes answers auditable. See what RAG is and RAG vs fine-tuning.

Several Dutch consultancies and software teams do, including Crux Digits. RAG has become standard practice rather than a specialism, so the useful question is less who can build one and more who will tell you honestly whether you need it. Ask any supplier how they measure retrieval quality.

• A good test question: “show me retrieval accuracy on our own documents, not on a demo set.”
• Crux Digits builds RAG assistants and search on client data from Nieuwegein, in Dutch and English.
• A RAG pilot on your own documents is normally scoped inside the €20,000 Proof of Concept. Start with a free consultation.

Yes, on fixed scope and milestones rather than supplied headcount. Alongside SMEs we deliver for larger organisations that want a small senior team accountable for one outcome. We stay boutique deliberately: the people who scope your project are the people who build it.

• Good fit: a department-level use case, a stalled pilot that needs to reach production, or an EU AI Act review.
• Poor fit: multi-year transformation programmes or long-term secondment — those genuinely belong at a large consultancy.
• The same fixed prices apply regardless of company size. More on about us and our services.

It depends on your use case's risk category under the Act. Most business automation falls into the lower-risk tiers, which still require documentation, transparency and human oversight. High-risk uses — hiring, credit and similar decisions about people — carry heavier obligations. Regulation (EU) 2024/1689 defines the categories.

• Article 4 requires AI literacy for staff who use AI at work, which reaches almost every employer.
• Being a deployer (you use someone else's AI) carries lighter duties than being a provider (you build or rebrand it).
• Our free EU AI Act risk checker classifies a system in four questions, no email required. Background in EU AI Act compliance in the Netherlands.

Data minimisation, agreed boundaries and no unnecessary copies. We take only the fields a use case needs, put a processing agreement in place, and keep data inside environments you approve. Where data is sensitive, we prefer architectures that keep it under your control rather than a vendor's.

• Options for sensitive data: EU-hosted models, private deployment, or on-premise inference.
• Ask any supplier whether your prompts and documents are used to train their models. They should not be.
• Deletion and exit are written into the agreement rather than left as an assumption. See our services.

€2,500 for an audit, €20,000 for a proof of concept. A Production Launch starts at €50,000 and ad-hoc work is €150 per hour, all excluding VAT. Those are published fixed prices, not estimates. Across the Dutch market, AI consultant rates generally run €125–€250 per hour.

• Big-four day rates run €2,000–€3,000, with a typical SME trajectory there landing at €50,000–€150,000.
• Most Dutch AI agencies publish no prices, so comparing them usually means going through a quote round first.
• The Dutch WBSO scheme can offset part of the R&D wage cost on qualifying AI development. See pricing.

Three fixed-price stages, and you can stop after any of them. The audit (€2,500) finds and ranks the use case. The proof of concept (€20,000) proves it works on your data. Production launch (from €50,000) deploys it and keeps it running.

• Typical timelines: audit 1–2 weeks, proof of concept 4–6 weeks, production from 8–12 weeks.
• Each stage ends with a decision point and a deliverable you keep, whether or not you continue.
• Skipping the audit is possible when the use case is already clear and the data is known. Stages map directly to pricing.

Fixed prices, a senior team, and you own everything we build. Crux Digits is a boutique firm, so the person who scopes your project is the one who delivers it. There is no lock-in on code or models. We do custom AI projects only — not secondment, BI tooling or marketing.

• Where a large consultancy fits better: multi-year programmes, board-level change management, or 24/7 managed operations.
• Crux Digits B.V. was founded in 2022, is based in Nieuwegein, and is rated 5.0 on Google across 5 reviews.
• Delivery is bilingual in Dutch and English — uncommon among Dutch AI boutiques, which are usually Dutch-only. We also benchmarked our AI against human experts.

Yes, fully bilingual. We are based in Nieuwegein, near Utrecht. Crux Digits B.V. works in Dutch and English across the Netherlands, remotely or on site, and with international teams elsewhere in Europe. Managing director Tom Joseph leads delivery. The company was founded in 2022.

• Address: Vlierhoeve 100, 3438 MW Nieuwegein, province of Utrecht.
• Contact: info@cruxdigits.nl or +31 6 44384676. A free 30-minute consultation is the fastest route to a straight answer.
• Every page on this site exists in both English and Dutch at the same slug. More about us.