Large European firms took up language AI faster than anyone planned. Our analysis of Eurostat data shows 55.0% of EU firms with 250 or more staff used AI in 2025, against 28.4% in 2021. Text mining and text generation now lead. The next gain sits one layer deeper: AI inside a workflow or a decision, which 75.6% of large firms do not yet run.
This piece is for the CIO, IT director or informatiemanager at a Dutch or Belgian organisation of 250 to 5,000 FTE who has the licences in place and is asked what comes next. Crux Digits builds process AI alongside your ICT partner or IT team; we are not that partner.
Is process AI the next step after language AI?
Yes: in 2025 only 24.4% of large EU firms used AI to run workflows or assist decisions, against 35.0% for text mining.
The order flipped in four years. In 2021 workflow automation was the most used AI technology among large EU firms, at 14.6%. By 2025 text mining (35.0%) and generating text, speech or code (31.7%) had passed it, and workflow automation sat sixth of eight. The figures come from Eurostat table isoc_eb_ai, which we queried again for this article.
That is a good result, not a warning. Language AI spreads quickly because it needs a licence, a browser and a document. It has put AI in front of thousands of employees and taught organisations how to buy it, secure it and write a usage policy. That groundwork is what process AI needs.
Process AI is different work. It puts a step inside a system that already has an owner: a purchase-to-pay flow in SAP or AFAS, a ticket queue in TOPdesk or ServiceNow, a planning run in a TMS. It reads the case, prepares or takes the decision, and hands it on. The gain is no longer one person writing faster. It is a hand-off that no longer waits.
What do Finland and Belgium do that others do not?
The leaders run both kinds at high levels: 51.7% of large Finnish firms and 45.0% of Belgian ones use AI in workflows or decisions.
Finland leads the league at 79.4% and Belgium is second at 76.4%. Finland also leads the EU on text mining at 68.7%. Their advantage is balance. We divide the share using text mining by the share using AI in workflows: the EU figure is 1.43, Belgium 1.31 and Finland 1.33. The Netherlands is at 1.56: 47.7% of large Dutch firms use text mining and 30.6% use AI in workflows.
They also go deeper. Of large firms that use AI, 84.8% in Finland and 75.6% in Belgium run three or more AI technologies, against 62.6% in the Netherlands. And they still build. In both countries about two in five large AI users (39.1% and 39.0%) have AI developed by their own employees; in the Netherlands it is 30.8%. Belgium also has the second highest share using an external provider to develop or modify AI, 54.6%.
The Netherlands is seventh at 67.6%, so the position is strong. The gap to the leaders is on the process line, not the headline. The full country detail is in the league table of large EU firms.
Should a large organisation build process AI or buy it?
Buy the language layer and build the decision layer that sets your process apart; that is the pattern the leaders show.
In all 20 countries with data, more large AI users run ready-to-use commercial AI than AI their own staff developed. Ready-made AI rose in 17 of 19 countries between 2021 and 2025. Buying is what made the fast growth possible, and for summarising, drafting and search it remains the right call.
Process AI is where off-the-shelf fits less well. The decision rules, the exceptions and the system it writes back to are yours. A tool built for every company knows none of them. That is why building has held up among the leaders. Our study also finds that building and process AI move together only weakly across countries (correlation 0.33). Building alone guarantees nothing; building the right step does.
A decision rule we use: if the AI step only touches text that a person then handles, buy. If it changes a record, routes a case or proposes an approval in a system of record, build it or have it built, with code and logs you own.
Which processes should go first in 2027?
Start where large EU firms already use AI most and where hand-offs are frequent: ICT security (26.1%), business administration (23.9%) and, for logistics firms, planning.
Among large EU firms, ICT security is the most common purpose for AI, ahead of business administration and management (23.9%) and marketing or sales (20.9%). Production follows at 18.4% and logistics at 8.2%. Belgium leads the EU on business administration at 46.8%. In logistics Belgium (18.9%) and Finland (16.8%) run at about twice the EU figure, which makes transport, warehousing and wholesale the clearest opening in the data.
Three examples of a first process, offered as illustrations rather than cases: triage and routing of security alerts before they reach the service desk; matching invoices to orders and receipts in the ERP, with the exceptions sent to a named person; and proposing a transport plan that a planner approves. Each has a clear owner, a system of record and a measurable wait between steps.
What changes when AI prepares or takes decisions?
Decision AI about people can be high-risk under Annex III of the AI Act, with those obligations applying from 2 December 2027.
Annex III point 4 lists AI used to recruit, to allocate tasks based on individual behaviour, or to monitor and evaluate the performance of workers. Point 5(b) lists AI that evaluates the creditworthiness of natural persons. After the Digital Omnibus, the rules for these stand-alone high-risk systems apply from 2 December 2027. Our summary of the Omnibus has the other dates.
For a deployer, Article 26 asks for logs kept at least six months and, at the workplace, for workers' representatives and affected workers to be informed before the system goes live. National labour law adds its own step. In the Netherlands, Article 27(1)(l) of the Wet op de ondernemingsraden gives the works council a consent right on facilities suited to monitoring attendance, behaviour or performance. In Belgium, CAO nr. 39 requires firms with at least 50 workers to inform in writing and consult on new technology with important collective consequences, at the latest three months before it is introduced.
Most process AI is not about people: an invoice match or a route plan is outside Annex III. Make that call in writing at the design stage, and put the works council on the plan for any step that touches staff.
What is the twelve-month playbook?
Five moves, each tied to a finding, take a 250 to 5,000 FTE organisation from language AI to its first process AI in production.
1. Benchmark on the process line. The 55.0% headline says only that AI runs somewhere. Compare yourself on AI in workflows: 24.4% EU, 30.6% Netherlands, 45.0% Belgium, 51.7% Finland.
2. Count your own ratio. List the processes where AI reads or writes text and those where it prepares or takes a decision. A ratio above 2 means AI lives in your documents; the leaders run near 1.3.
3. Pick one process with a frequent hand-off in an area where leaders are ahead: business administration, ICT security or logistics. Name the owner and the system of record before anyone writes code.
4. Buy for language, build for the decision. Keep the licences. Build the step that writes back to your ERP or ITSM, and own its code, logs and tests.
5. Settle the rules at design time. Record whether the step falls under Annex III, keep the logs Article 26 expects, and plan the works council timing: consent in the Netherlands, three months' notice under CAO nr. 39 in Belgium.
If the 2021 to 2025 direction holds, the share of large firms using some AI will stop telling organisations apart. Our reading is that 2027 rewards the organisations that moved one decision into a system and can show how it behaves.
Where does Crux Digits fit?
We design and build the process AI step. That covers the decision logic, the integration with your existing systems, the logging and the evidence an auditor asks for. We work alongside your ICT partner or IT team, who keep running your infrastructure. We are not an ICT partner, and the code we write is yours.
Read the numbers behind this piece in our study of large EU firms, and see what a build like this costs on our page on custom AI development cost.
Last updated: 28 October 2026.
A consultant tells you where AI pays off; Crux Digits also builds it. A fixed price per step, one named expert, from Utrecht.
AI consultancy in the Netherlands →
