Process Automation: What to Automate First

Share

In short

Automate the boring, high-volume, rule-based process that everybody already agrees about, not the painful one everybody argues about. The first automation has to succeed visibly, because it buys permission for the next five. A process is a good candidate when it runs often, follows written rules, has a clear right answer, and someone already checks it. If people disagree about what the process even is, that is a process problem and automating it just makes the disagreement faster.

Choose first, then build

What to automate and what to leave alone

The gain is in picking the right process. The damage is in automating the wrong one perfectly.

Candidate
What makes a good process
Daily, describable in rules, and running on data that already exists somewhere. Three yeses make a candidate.
Leave alone
Where judgement is the work
If every case looks different and experience decides, you are automating away the judgement that was the quality.
The numbers
What actually happened
Dutch companies automated admin and intake most, and quietly dropped their bespoke processes.
Business case
The staffing argument turned
Automating against scarcity weakened as an argument in 2026; quality and lead time gained ground.
How it fails
Broken stays broken, just faster
Automating a bad process does not improve it, it makes it consistently wrong. Repair first, then accelerate.

Choosing

What makes a process a good candidate

Score any candidate against these four. Three out of four is a good start; two is a research project.

Volume. It happens often enough that saved minutes add up. Something monthly rarely repays the build.
Rules. The decision can be written down, even if it currently is not. If the rule is "it depends on experience", start by extracting the rule.
A checkable answer. You can tell afterwards whether it was right. Without that there is no way to measure or to improve.
Tolerable failure. When it gets one wrong, a person catches it and the cost is small. Start where errors are recoverable, not where they are regulatory.

What to avoid

The processes that look ideal and are not

Three categories consistently disappoint, and all three are usually the first suggestion in the room.

The process nobody agrees on. Two departments describing it differently means you are being asked to arbitrate, not to automate.
The one that changes every quarter. Automation freezes a process. Freezing one that is still moving buys a rebuild.
The rare, high-stakes exception. Low volume, severe consequences and a judgement call. That is exactly where a person should stay, and where automation has the worst risk-to-return ratio available.

The evidence

What Dutch firms actually automated, and what they quietly dropped

There is a public dataset that settles a lot of arguments about this, and almost nobody in the automation market quotes it. Statistics Netherlands publishes what Dutch companies actually use, split by size, and its March 2026 report on micro-enterprises reports the 2025 figures for seven separate technologies.

The headline is ordinary enough: 13.8% of firms with 2 to 9 staff used at least one of them in 2025, up from 10.6% in 2024 and 6.8% in 2023. For companies with 10 to 249 staff it is 29.8%, and for 250-plus it is 66.2%. Adoption tracks headcount almost perfectly, and it does so inside the micro band too: 11.6% at two staff, 14.1% at three or four, 19.0% at five to nine.

The interesting number is the one underneath. Broken out by technology, robotic process automation sat at 2.3% in 2025, against 2.1% in 2024 and 1.7% in 2023. Three years, essentially flat. Over the same period text mining went 3.1%, 6.9%, 9.8%, and speech recognition went from 2.9% to 4.4%.

Our reading, and it is a reading rather than a finding: the automation that grew is the automation that reads and writes language. The automation that drives screens did not move at all. That matches what we see in quoting work. Screen-driving robots demo well and then break on a layout change, so they get built once and quietly retired, while anything that turns a document into structured data survives contact with a real back office. It is also a useful corrective if you have been told that RPA is what process automation means. In the Dutch data it is the one thing that never took.

The sector spread is the other thing worth knowing before you benchmark yourself against a headline. Among micro-enterprises, information and communication runs at 50.4% and specialist business services at 27.1%, while transport and storage sits at 3.6%, construction at 4.6% and hospitality at 4.2%. If you are in one of the low bands you are not behind your peers. You are looking at an argument that has not been made to your industry yet.

The business case

The staffing argument got weaker in 2026

For most of the past five years the automation pitch in the Netherlands leaned on hiring: you cannot find people, so automate. That argument has softened, and it is more honest to say so than to keep using it.

In the second quarter of 2026 there were 95 vacancies for every 100 unemployed people, with vacancies down by 3,000 and the number of unemployed down by 17,000 over the quarter. The market is no longer past parity. It is still tight by historical standards, but "we physically cannot hire" is not the automatic closing line it was in 2022.

That is not an argument against automating. It is an argument for a business case that survives the question. Throughput, error rate, and how long a customer waits for an answer are all measurable before and after, and they hold up whether or not the labour market is tight. Headcount avoidance is the weakest of the four and the first one a finance director will interrogate.

How it goes wrong

Why automating a broken process makes it worse

Automation is an amplifier. A process with a 4% error rate and a person catching most of them becomes a process with a 4% error rate and nobody catching anything, running ten times as often.

The fix is dull and reliable: map the process as it actually runs, not as the procedure says, and fix the steps that are simply wrong before you automate any of it. Roughly half the value in a typical automation project is realised in that mapping, before a line of code is written, and it is the half nobody budgets for.

Keep a person in the loop on anything with consequences. Not as a permanent crutch, but because the review data is what tells you when the system is ready to be trusted, and you cannot get that data any other way.

Buyer guides

Compare further

API integration AI automation Custom software or a package AI with Exact, AFAS and e-Boekhouden

FAQ

Frequently asked questions

Where should we start?

With one process, end to end, that meets the four criteria above and that a team already complains about. Narrow and complete beats broad and partial, because a finished automation changes behaviour and a half-finished one becomes a workaround.

Is this the same as RPA?

No, and the Dutch data suggests RPA is the smaller story. RPA drives the user interface as a person would: quick to deploy, and fragile because it breaks whenever a screen changes. Working against an API instead is more work up front and far more stable. Statistics Netherlands puts RPA use among micro-enterprises at 2.3% in 2025, barely moved from 1.7% in 2023, while text mining nearly tripled to 9.8% over the same three years. We reach for RPA only when there is genuinely no other way in.

What does process automation cost?

A single, well-scoped process is usually a proof of concept at €20,000, with production from €50,000 depending on how many systems it has to touch. The audit at €2,500 exists mainly to work out which process is worth doing first, and it frequently rules candidates out.

Will it replace jobs?

In the work we do it usually removes a task rather than a role, and the honest reason is scale: the processes that are good candidates are the ones people already dislike and defer. If your goal is headcount reduction rather than throughput, say so at the start, because it changes which process to pick.

Process automation: what to automate first, and what to leave alone

What to automate and what to leave alone

What makes a good process

Where judgement is the work

What actually happened

The staffing argument turned

Broken stays broken, just faster