90-day engagements for European scale-ups and SMEs. I embed in your business, understand your workflows, and build concrete systems your team actually uses.
Under the operational load of every company sits an original mission.
I build AI that lifts that load — so the mission gets room to breathe again. And can succeed, with the discipline every mission deserves.
Most AI initiatives end in a report, a pilot, or a slide deck.
A stalled pilot costs more than its budget. It teaches the organisation that AI “doesn’t work for us” — and pushes the next serious attempt back months, sometimes years, exactly when that capability is worth the most. The real bill is the internal mandate to ever try again.
So I don’t just sell a working system, but the assurance that your first serious attempt won’t be your last. My approach as a Forward Deployed AI Engineer is different.
I embed in your business, understand your workflows, and build concrete systems your team actually uses. No consulting hours, no vague roadmaps, no pilots that never ship. Instead: working software, knowledge transfer, ROI within 90 days.
That starts from two principles. Safety and human control are built in from the design: no black-box automation, people stay in charge of the decisions that matter — with audit trails, confidence-scoring and clear escalation paths. And technology translated to daily practice: many software builders ship elegant systems that never land in production because the translation to the floor is missing. My experience spans both worlds — technical enough to build, close enough to the work to understand why adoption succeeds or fails.
A Forward Deployed AI Engineer (FDE) works embedded with the client, not at a distance. The term comes from Palantir’s playbook and is now being adopted by firms like BCG X and EY — for consulting work that goes beyond reports. The role: into your organisation, understand the workflows, and build concrete AI systems that go into production. No pilot phase handed over to an “implementation team”. The same person who writes the roadmap also delivers the working software.
Delivers a report and a roadmap, then leaves. The building — the part where it usually goes wrong — stays on your plate.
Solves one bounded problem but doesn’t know your workflow. You bend to the tool, not the tool to you.
Finding the right person takes months of recruiting and onboarding. Only then does the building start — if the market hasn’t moved faster than your vacancy.
A Forward Deployed Engineer builds working software inside your organisation and leaves the knowledge behind. No handover to another team, no tool wrapped around your workflow, no waiting before something exists.
Start small, expand. Each tier ends at a natural decision point — no annual contracts, no hidden subscriptions.
Workflow analysis of your organisation, identifying real AI opportunities — and the places where AI adds no value. Includes mapping your existing processes and translating them into a concrete, prioritised roadmap with an implementation proposal.
One concrete AI system built and taken into production. Agentic workflow, document intelligence, customer communication, or operational automation. Price scales with the number of data sources, integrations and the complexity of your existing systems. Ends with working software, governance and knowledge transfer.
Embedded in your organisation as a Forward Deployed Engineer. Building multiple AI systems in parallel, coaching your internal team, setting up governance. The highest return for companies serious about scaling.
The heaviest obligations — for high-risk systems — are being postponed via the Digital Omnibus: stand-alone systems (Annex III) from 2 August 2026 to 2 December 2027, and systems embedded in products (such as medical devices) to 2 August 2028. The European Parliament adopted the text on 16 June 2026; formal adoption by the Council and publication in the Official Journal are still to come. Until that publication, 2 August 2026 remains the formally applicable date — count on the new dates, but be prepared. What does not change: the AI-literacy obligation and the ban on unacceptable uses have applied since February 2025, the GPAI obligations since August 2025, and enforcement starts 2 August 2026. Fines run up to € 15 million or 3% of worldwide annual turnover for high-risk breaches, and up to € 35 million or 7% for prohibited uses. The delay is not a cancellation — it is exactly the time to build high-risk systems properly, with audit trails, confidence-scoring and human control, rather than sprinting to a date.
Status: 17 June 2026
Six recurring building blocks from my work on pcdOS and with other clients. Not ticks on a feature list, but the foundations that get AI systems working and reliable in production.
Not one AI agent for one task, but a team of agents that runs a complete process together — from deal sourcing through due diligence to an investment memo, say, or from patient intake through triage to record-keeping.
Agents check their own work before it reaches your team. Strict rules define what 'good enough' is; output that doesn’t clear the bar stays inside the system and is revised automatically.
Light, fast AI for routine work; more powerful AI for complex judgement. The right engine per task — saving significantly on running costs without losing quality where it matters.
AI that answers from your own documents, SOPs and regulations — not from general internet knowledge. Every answer comes with a direct reference to its source, so verification is always possible.
The system starts with human approval on every decision. Autonomy grows step by step per workflow, based on a proven track record. No automation without a safety net, and no needless bottleneck either.
Models that run scenarios for strategic decisions. Example from practice: a throughput-time simulation with multiple adoption scenarios and a clear tipping point. Not predicting, but exploring options before the investment.
Anonymised example from practice. Three collaborating AI agents and one human decision together produce a complete investment memo. Each step has its own confidence score; the human can step in at any moment and has full visibility into what the agents did.
An illustrated view with fictional data — this is how the building blocks look in production: confidence-scoring per step, a human who decides, answers with source references, and scenario simulation.
Preliminary verdict: strong commercial traction, but an elevated risk profile on governance and funding timing.
Being honest about what fits saves us both time. Read the two columns — pick which side you're on.
Niels Roest is a Forward Deployed AI Engineer (Netherlands-based) who builds production AI for European scale-ups and SMEs in 90-day engagements.
“A builder with 12+ years in military and police special operations. Direct, accountable, no fluff.”
PCD CareHub — where I am Chairman & CAIO. Not an independent client; but the place where I built all of this, solo, from scratch.
Investment and care processes that leaned on manual work — research, memos, intake, record-keeping — spread across disconnected tools and people.
pcdOS — since November 2025 an AI-native platform with 23+ modules in production: an investment-memo generator, an investor-support agent with confidence-scoring, self-evaluating agents (Zod output schemas) and an internal Academy. The human stays in charge — each step has a confidence score and an audit trail; autonomy only grows on a proven track record. Stack: Next.js, Supabase, Prisma, pgvector, MCP.
“Since November 2025, Niels has strengthened our team at PCD, and his impact is immediately visible. With his strong leadership, his experience building high-performance teams and a sharp vision on digitalisation and AI, he has played a key role in the development PCD has gone through since.”
“What sets Niels apart is his proactive attitude. He continuously spots opportunities to improve processes and genuinely takes initiative. He gets to grips with new subject matter quickly; if he doesn't yet know something, he makes sure he understands it. On top of that, he brings not only substantive knowledge but energy and innovation.”
Disclosure: Coby and Patrick are co-founders of PCD CareHub, where I am Chairman & Chief AI & Innovation Officer. The recommendations are translated from Dutch; read the originals on LinkedIn.
Start with Tier 01, the Discovery. €5.000 excl. VAT, 2–4 weeks, depending on your team's availability. You get a concrete roadmap with prioritised AI opportunities, explicit "no AI here" recommendations and an implementation proposal. Then you decide whether a follow-up makes sense — Sprint, Retainer, or stop. No obligation. Most organisations that continue after Discovery do one Sprint (€15.000, 4–8 weeks, one working system delivered) for the first use case. If parallel systems or internal coaching are needed, we scale up to an FDE Retainer. But that decision comes after Discovery, not before. What I mainly check in Discovery: whether there's a clear problem owner inside your organisation, whether the data is genuinely available, and whether your team is mandated to change processes. Without those three, any AI initiative becomes expensive theatre.
Hybrid. The first two weeks are usually intensive on-site, to really get to know your business — observing workflows, shadowing operators, asking questions of people outside the project team. Not what managers think happens, but what actually happens. After that, remote with a fixed weekly on-site moment, usually half a day, for reviews and blockers. On an FDE Retainer I'm embedded 2–3 days per week; that only works on-site, because most of the value sits in informal corridor conversations, not in scheduled Zoom calls. For clients outside the main hubs: I'm happy to come on-site, provided the travel is workable. For clients outside the Netherlands (EU/UK): hybrid with 3–5 day on-site blocks per month usually works best. Tell me what suits you and we'll find the rhythm.
Discovery is short and bounded. €5.000, 2–4 weeks, fixed scope. At the end you decide whether a follow-up makes sense — no pressure, no sales pitch. No annual contracts, no hidden subscriptions. A Sprint runs 4–8 weeks; then I deliver and we close it out. If there's nothing left to build: fine, no pressure to extend. A Retainer can be ended at natural break points: quarterly, with 30 days' notice. If during a Sprint I sense we're heading the wrong way, I'll say so — an honest early rethink beats an expensive post-mortem. And the other way round: if your team doesn't move with it or the organisation doesn't adopt the results, we close it out without fuss. My goal is that clients recommend me to other leaders, not that they're locked into a contract. Reputation outweighs one year of revenue.
My defaults: Next.js + TypeScript, Supabase + Postgres + pgvector, Prisma, Claude (Sonnet for reasoning, Haiku for throughput), MCP for tool integrations, Vercel for hosting. For agentic workflows: Claude Code as orchestrator, Zod for output validation, and custom self-eval loops. But no religion. If you already have a serious stack — Python, Azure, AWS, n8n, LangChain, whatever — I build on it, provided it's genuinely production-grade. What I don't do: vendor lock-in as a goal, or introduce new tools just because they're fashionable. What I do: choose stacks your internal team can keep building on independently. For healthcare clients I specifically check the stack is compatible with health-security standards (e.g. NEN 7510), HL7 and FHIR, and that we can stay EU-hosted (Supabase EU, Anthropic EU endpoints). For financial services: SOC 2-aware vendors and log trails by default.
Built in from day one, not a compliance afterthought. EU AI Act: each system gets a per-use-case classification (minimal, limited, high-risk, or unacceptable). For high-risk: conformity assessment, risk-management documentation, logging, and human oversight by design. GDPR: data minimisation as default — we only send the model the fields it genuinely needs. EU hosting where possible (Supabase EU, Anthropic via EU endpoints, no US round-trips). If you don't have a DPIA template yet, we provide one with every implementation. For healthcare clients, health-security standards (e.g. NEN 7510) plus HL7/FHIR integration checks are in scope. For finance: SOC 2-aware. What I don't do: use GDPR/AI Act as an excuse to build nothing. What I do: make sure the legal clarity is there before we deploy, so you get no surprises after launch. When in doubt, I work with an AI lawyer I can bring in on call.
Yes — that's usually the whole point. AI systems without integration are isolated parcels: a chatbot that doesn't know what's in your CRM, or a document AI that can't write its output back into your workflow tool. Not useful. Via the Model Context Protocol (MCP) and direct APIs I connect to what's already running: HubSpot, Salesforce, Pipedrive, Notion, Confluence, Slack, Microsoft 365, Google Workspace, common ERPs (SAP and others), and healthcare systems (Salesforce Health Cloud, Epic and other EHRs). For tools without an API we look at alternatives: scraping (legitimate and with permission), email parsing, webhooks, or as a last resort RPA-style automation. In practice we always find a way. What I ask upfront: a list of your top-5 systems and which of them already have a serious API. We build around those. Removing existing tools is usually more expensive than integrating them.
No problem — it's the rule rather than the exception. Most organisations I help have smart people who've never worked with LLMs or agentic workflows. That's not a blocker, it's a given. Knowledge transfer is an explicit part of every tier, not an optional add-on. On a Sprint: I deliver not just working software, but README documentation, an internal playbook with the design choices, and video walkthroughs of the key flows. Your team can maintain, debug and make simple extensions themselves by day 30. On a Retainer: I coach your technical people one-on-one — pair programming, code reviews, and weekly sessions on what happened this week and why. The goal is that within 3–6 months they can build the next generation of systems independently, with me as a sparring partner on call. Not a dependency you never wind down.
A 30-minute intro, not a sales pitch. We map where AI pays off most for you — and where it has no place at all.
I work solo and take on a limited number of engagements each quarter.
Or call directly: +31 6 28 71 03 83