Autonomous, multi-step AI agents that use your tools and APIs to complete real work — planning, tool use, RAG grounding, guardrails and human-in-the-loop. Built for the Dutch MKB by a senior-led team, EU AI Act and GDPR first.
Last updated: 11 June 2026
An AI agent is autonomous software that plans a multi-step task, uses your tools and APIs to carry it out, and finishes real work rather than only answering questions. Crux Digits builds these agents for Dutch SMEs with RAG grounding, permissions, guardrails and human-in-the-loop review, so they act safely and stay EU AI Act and GDPR compliant.
Most AI projects stop at the conversation. A chatbot answers a question, a model drafts an email, and a human still has to log in somewhere, copy the answer across, and click the buttons that actually finish the job. Agentic AI closes that last gap. An AI agent plans a multi-step task, calls your tools and APIs to carry it out, checks its own progress, and only hands back to a person when it should. The difference is the one that matters to a business: an assistant tells you what to do, an agent does it.
Crux Digits builds these autonomous agents for Dutch SMEs (MKB) from our base in Nieuwegein, in the province of Utrecht. We are a boutique, senior-led AI consulting and implementation company, founded in 2022, working bilingually in English and Dutch. If you are weighing an AI agent development company rather than another vendor selling a generic chatbot, the question to ask is whether they can safely let software take real actions in your systems. That is the work described here.
An agent is not a bigger chatbot. It is a system built around a language model that can reason about a goal and then act on it across several steps. Where a standard assistant produces text, an agent decides what to do next, uses the tools available to it, observes the result, and adapts. Four capabilities separate genuine agentic AI development from a thin wrapper around a public API.
Real tasks have more than one step. A refund involves checking an order, confirming a policy, issuing the credit, and notifying the customer. An agent breaks a goal into those steps, sequences them, and recovers when one fails — that orchestration layer is where most of the engineering lives, and where weekend projects fall apart.
An agent is only useful when it can reach the systems you already run. We give agents controlled access to your CRM, ERP, ticketing, email, databases and internal APIs, so they can read a record, update a status or trigger a downstream process. The model never touches your systems directly; it calls defined tools with defined permissions, and nothing else.
To act correctly an agent needs facts, not guesses. We ground agents in your own documents, policies and data using retrieval (RAG), so decisions trace back to a real source with a citation rather than to the model's imagination. This is the same accuracy-first approach behind our generative AI builds, applied to systems that take action instead of only answering.
Autonomy without limits is a liability. Every agent we build runs inside explicit guardrails: what it may do, what it must escalate, and where a person signs off before anything irreversible happens. A support agent might resolve a routine query end to end but route a refund above a set amount to a human. You decide where the line sits.
Across our 13 case studies — spanning healthcare, NLP, computer vision and forecasting — the pattern that pays off fastest is the same: a narrow, well-bounded task an agent can own completely, with a clear definition of done. Start broad and an agent becomes hard to trust and harder to measure; start narrow and you have something live and earning its keep within weeks. Three shapes come up repeatedly for Dutch SMEs, and each can be extended once it has proven itself in production.
In each case the agent is measured on completed work, not words produced. That is the bar we hold our own builds to, and the one we recommend you hold any vendor to.
It helps to see where agents sit in the wider picture. Our generative AI service is the broad build layer — assistants, document Q&A, content and knowledge tools grounded in your data. Our AI automation service is business-process automation, connecting and streamlining the steps in a workflow. Agentic AI is the autonomous action layer that sits on top: it uses generative models to understand and decide, and it uses automation plumbing to execute, but it adds the judgement to plan a task and adapt mid-stream. Many engagements combine all three. When you are not sure which layer your problem needs, our AI consulting in the Netherlands exists to answer exactly that before any code is written.
An agent that takes real actions has to be trusted, and trust is earned with measurement. Before an agent goes live we evaluate it against a test set of real cases, scoring whether it completed the task, stayed inside its guardrails, and got the facts right. We treat that evaluation suite as part of the deliverable: it is how you confirm the agent is ready, and how you catch a regression the day a model or a policy changes. Once live, every agent is monitored — each action is logged, success and escalation rates are tracked, and drift is caught before it becomes a problem. You can always see what the agent did and why, which is both good engineering and, for an agent that acts on customer data, a regulatory requirement.
For Dutch and European businesses, the EU AI Act and the GDPR/AVG are not paperwork added at the end — we treat them as design inputs from the first decision. We classify where an agent's use case sits under the AI Act's risk tiers, keep personal data minimised and documented, and record how the agent reaches each decision so it can be explained to an auditor, a regulator or a customer. Building an agent that acts on your systems makes this discipline non-negotiable, and it is a clear reason to choose a local partner over an offshore team unfamiliar with the rules you answer to.
We deliver fixed-scope projects, not bodies for hire. Crux is not a staffing or body-shop firm — there is no dedicated-team rental, no nearshore bench, no staff augmentation. Instead, the senior engineer who scopes your agent stays on it through launch, and the finished solution — code, prompts, evaluation suite and documentation — transfers to you in full. You own what we build and can run or extend it without us.
Pricing is fixed-step and published, all excl. VAT. An AI Audit & Strategy at EUR 2,500 maps where an agent genuinely pays off and where a simpler tool would serve you better. A Proof of Concept at EUR 20,000 puts a working agent in front of real tasks in weeks, not quarters. A production launch starts from EUR 50,000, with ad-hoc work guided around EUR 150 per hour. Full detail sits on the pricing page.
The honest place to begin is the audit. In a short working session we map one process end to end, decide whether an autonomous agent is the right answer or whether plain automation would do, and define the guardrails and the human-in-the-loop points before anything is built. If the case is strong, a Proof of Concept follows — one agent, one real workflow, measured against results you agree up front. That is how you build AI agents that survive contact with your actual data and customers, rather than a demo that impresses in a meeting and breaks in production. To scope an agent for your business anywhere in the Netherlands or Europe, reach Tom Joseph and the team at info@cruxdigits.nl or +31 6 44384676.
A chatbot answers questions in text; you still do the work it describes. An AI agent goes further — it plans a multi-step task, calls your tools and APIs to execute it, observes the result, adapts when a step fails, and completes the job, escalating to a person only where you decide it should. The shorthand we use with clients: an assistant tells you what to do, an agent does it. That action layer is the whole reason to build one.
The fastest wins are narrow, well-bounded tasks an agent can own end to end. Common builds are customer-support agents that read a ticket, retrieve the answer from your knowledge base and resolve routine requests in Dutch and English; document and back-office agents that read invoices or contracts, extract the fields that matter and enter them into your finance or ERP system; and operations agents that monitor stock or shipments and act on anomalies. Each is measured on completed work, not words produced, and can be extended once it has proven itself in production. We usually launch one agent on one workflow first.
Every agent runs inside explicit guardrails: what it may do, what it must escalate, and where a person signs off before anything irreversible happens. The model never touches your systems directly — it calls defined tools with defined permissions, and nothing else. An agent that touches your CRM, ERP, ticketing and email needs each of those connections defined and permissioned individually, which is exactly the work that separates a reliable agent from a demo. We add human-in-the-loop checkpoints at the points you choose, log every action for audit, and ground decisions in your own data with RAG so each one traces back to a real source rather than to the model guessing.
Generative AI is the broad build layer — assistants and document tools grounded in your data. AI automation streamlines and connects the steps in a business process. Agentic AI is the autonomous action layer on top: it uses generative models to understand and decide, and automation plumbing to execute, adding the judgement to plan a task and adapt mid-stream. Many projects combine all three layers. If you are unsure which one your problem needs, our AI consulting engagement answers that before any code is written.
They are when built that way, and that is how we build them. We treat the EU AI Act and GDPR/AVG as design inputs from the first decision, not paperwork added at the end — classifying the use case under the AI Act risk tiers, minimising and documenting any personal data, and recording how the agent reaches each decision so it can be explained to an auditor, regulator or customer. For an agent that acts on real customer data, that traceability is both good engineering and a legal requirement, and it is a strong reason to choose a Netherlands-based partner.
Pricing is fixed-step and published, all excl. VAT: an AI Audit and Strategy at EUR 2,500, a Proof of Concept at EUR 20,000, and a production launch from EUR 50,000, with ad-hoc work guided around EUR 150 per hour. The honest place to start is the audit — in a short working session we map one process end to end, decide whether an autonomous agent is the right answer or whether plain automation would do, and define the guardrails and human-in-the-loop points before anything is built. We base the recommendation on your real data and will say plainly when an off-the-shelf tool would serve you better.
No. Crux delivers fixed-scope projects, not staffing — there is no dedicated-team rental, nearshore bench or staff augmentation. The senior engineer who scopes your agent stays on it through launch, and the finished solution, including the code, prompts, evaluation suite and documentation, transfers to you in full. You own what we build and can run or extend it yourself, or hand it to another partner, without depending on us.
Start with an AI Audit and Strategy. We map one workflow end to end, decide whether an autonomous agent is the right answer, and define the guardrails before any build — fixed price, EU AI Act and GDPR first. Talk to Tom Joseph and the Crux Digits team.
Book your AI audit