RPA vs AI: where rule-based automation stops and AI agents begin
The cleanest way to scope an AI automation project (AI automatisering, as Dutch clients usually search for it) is to separate the two technologies that get lumped together. Robotic process automation (RPA) records a fixed sequence of clicks and keystrokes: open the ERP, paste a value, hit submit. It is fast and low-cost when the screen never changes and the input is already structured. The instant a supplier redesigns its PDF, a field moves, or a customer writes a sentence instead of ticking a box, the bot fails silently and someone spends a morning untangling it.
AI changes the failure mode. A language or vision model reads the messy input, infers what it means, and decides the next step, the way a junior colleague would after a week of training. That is the real RPA vs AI distinction: RPA executes a script, AI exercises judgement. In most Dutch MKB processes the right answer is not one or the other but a layered design: AI for the reading-and-deciding, deterministic workflow automation for the moving-and-posting, so every irreversible action still runs through a rule you can audit.
Where AI agents earn their keep
An AI agent is a model given tools and a goal rather than a single prompt. Instead of "summarise this email", it can look up the order in your CRM, check stock, draft a reply, and flag the three cases a human must approve. Agents pay off when a task spans several systems and needs a decision in the middle:
- Triage and routing: incoming mail, tickets or applications sorted by intent and urgency, then handed to the right queue with a drafted first response.
- Multi-step lookups: pulling facts from accounting, inventory and a knowledge base to answer one customer question that today bounces between three people.
- Exception handling: the 15% of cases your existing rule-based business process automation kicks out, which is usually where the manual cost actually lives.
We design agents with hard boundaries: read-only by default, narrow tool access, and a human in the loop for anything that spends money or touches a customer record. That keeps the failure mode boring instead of expensive. For the full picture of that work, see AI agents & agentic AI development.
What AI automation actually costs to run, and how to measure the return
The build price is visible; the running economics are where projects quietly succeed or fail. Beyond the fixed-step ladder (a €2,500 AI Audit & Strategy, a €20,000 Production-ready MVP and production from €50,000) every automation carries an operating cost: model usage (priced per token or per call), hosting, monitoring, and the periodic re-tuning a model needs as your data drifts. A serious vendor models all of this before you commit, not after.
The numbers that decide whether to automate
We assess each candidate process the same way, so the decision rests on arithmetic rather than enthusiasm:
- Volume × handling time: a task done 4,000 times a month at six minutes each is worth automating; the same task done twice a week is not, however tempting the demo looks.
- Error and rework cost: what a single mistake costs downstream, multiplied by how often it slips through today.
- Straight-through rate: the share of cases the system can finish without a human. A realistic 70–85% with confident human review of the rest beats a fragile "100% automated" that nobody trusts.
- Payback window: most well-scoped MKB automations should pay for the build within roughly a year; the audit puts a defensible figure on each one.
The point of starting with a proof of concept is to replace these estimates with measured numbers on your own data before the larger investment. You see the straight-through rate and the saving on real cases, then decide.
Governance, the EU AI Act and AVG: built in, not bolted on
Automation that reads invoices, customer messages or HR data is processing personal and commercial information, which puts it squarely inside GDPR (AVG in Dutch) and, increasingly, the EU AI Act. For most MKB use cases (document processing, customer-query handling, reporting) the obligations are proportionate, but they are real, and retrofitting compliance after launch is the expensive path.
What compliance-first looks like in practice
- Risk classification up front: we categorise each automation under the EU AI Act so you know on day one whether it is minimal-risk or carries transparency and documentation duties.
- Data residency by design: for sensitive data we default to EU-hosted or self-hosted models (for example self-hosted workflow automation on your own infrastructure), so records never leave an environment you control.
- An audit trail on every decision: what the model saw, what it decided, and who approved it, so a regulator or an auditor gets a straight answer.
- Human oversight where it matters: explicit approval gates on actions with legal or financial weight.
This is also a procurement advantage. A weekend-rebranded "AI" web agency rarely thinks about the AI Act at all; a large enterprise consultancy will, but at a day rate and on a timeline most SMEs cannot justify. Our day-rate guidance sits around €150/hour, and compliance is part of the method rather than a separate workstream you pay to add on.
Making AI automation stick: integration, monitoring and adoption
The hardest part of an automation project is rarely the model: it is everything around it. A pilot that works in a demo and a system your team relies on every Monday are different things, and the gap between them is integration, monitoring and adoption.
The work that turns a pilot into production
Going from proof of concept to a dependable service means wiring the automation into the tools you already run and giving it somewhere to fail safely:
- Real integration: connected to your CRM, accounting package, mailbox and document store, not a parallel spreadsheet the AI quietly maintains on the side.
- Monitoring and alerting: dashboards on straight-through rate, accuracy and cost, plus an alert when the model starts behaving differently so you catch drift before customers do.
- A confidence threshold: low-confidence cases routed to a person automatically, which is what keeps trust intact in the first months.
- Documentation and handover: because the deliverable is a system your team understands and can operate, not a black box only the vendor can touch.
This is where our boutique model is deliberate. The senior people who scope the audit stay on through build and handover (there is no junior-staffing swap halfway through) and the explicit goal is that you own the running solution. That ownership is what separates an automation that survives its first edge case from one that gets quietly switched off after the consultants leave. You can read how that engagement model works on our about page.
How to start without betting the company
Almost every failed automation we have seen shares one cause: it tried to do too much at once. The reliable path is the opposite: one bounded process, measured honestly, then expanded once it has earned the trust of the people who use it.
A sensible first project
Good first candidates share a profile: high volume, high manual effort, tolerant of a human-review step, and contained within one or two systems. Invoice and document processing, first-line customer-query handling, and report generation almost always make the shortlist for a Dutch SME. We start there, prove the saving, and only then connect it to adjacent processes.
If the question in your head still begins with "where" or "whether" (where does automation pay off, whether a given process is even feasible) the right first step is AI consulting rather than a build. If you already know what you need and want it built and integrated properly, that is AI implementation, and for the generative-AI pieces specifically, generative AI. Either way it starts the same: a fixed-price audit that tells you, with numbers, which process to automate first, and what it will be worth when it runs.
Crux Digits