Automating business processes with AI: what actually works

Automating business processes with AI sounds like something you need a strategy for first. In practice it is a decision per process: this one yes, that one no, and that other one next year. Get that decision right and the first automation pays for itself within a few months. Skip it and start with a tool, and a year later you will recognise yourself in the companies where AI is 'quite handy' and nothing more. In this article: how to judge a process, six processes that almost always work in a small or mid-sized business, four that only look like they would, and what it costs.

First things first: a rule or a judgement?

Most work in a company is a mix of two things. There are steps that follow a fixed rule: every invoice from that one supplier goes into that one folder, every order above a thousand euros comes past you. And there are steps where someone first reads, listens or looks, and then decides what needs to happen.

That distinction drives your whole budget. A fixed rule does not need AI. You can simply connect it: cheaper, faster and more reliable, because a rule does the same thing every day. You only need AI at the moment something has to be judged — an email where the tone matters, a delivery note that looks different at every supplier, a request in which the customer forgets to say what they actually need.

What Dutch companies use AI for right now — and what they do not

Statistics Netherlands (CBS) measures every year which AI technologies Dutch companies use. Among companies with 10 to 49 employees that use AI, the 2025 order is this: automatically reading and interpreting text sits at 69.6 percent, generating text at 46.6 percent, creating images, video or sound at 44.7 percent and speech recognition at 31.6 percent. Machine learning is at 24.9 percent. The category closest to 'AI that actually does the work' — automating workflows and supporting decisions — sits at 21.5 percent. Image recognition reaches 13.9 percent, service robots 5.6 percent.

So the top four are all language: reading, writing, generating and understanding. For every company that lets AI do work, more than three only let AI read and write. (Source: CBS, ICT use by businesses, reporting year 2025, via opendata.cbs.nl; the same measurement appears with two decimals at Eurostat, table isoc_eb_ai.)

You can read that two ways, and both are true. One: there is barely any proven demand for the layer that actually takes work over. The other: four out of five of these companies have already crossed the 'we do nothing with AI' threshold and have stalled on the layer below. The second reading is the interesting one, because the reading-and-writing layer is exactly what you get by taking out a subscription. The work that costs hours sits one layer deeper.

The five questions that tell you if a process qualifies

Run your processes past these five questions. Not in a day-long workshop — just on one sheet of paper, in half an hour, with the people who do the work.

• 1. How often does it happen? Daily, or several times a week, is the floor. Something that happens four times a year is almost never worth it, however annoying it is.
• 2. Does it involve reading or judgement? If someone has to open something, read it and then decide where it goes, that is the kind of work where AI adds something. If it is pure data entry, look at an ordinary connection first.
• 3. Is the input available digitally? Email, PDF, form, speech: fine. A note in someone's head, a paper folder in the van, or a price agreement only the owner knows: then your first project is not an AI project, it is getting that input into digital form.
• 4. What happens if it goes wrong, and would you notice? A wrong label on an email is a shrug. A wrong price to a customer is money. This answer does not decide whether you automate, but whether a human stays in the loop.
• 5. Can you measure the result in something other than hours? Think orders, lead time, error rate, margin per order, or how often a customer has to call twice. If you cannot name it, you will not be able to tell afterwards whether it worked.

Four or five yeses make it a candidate. Two do not, even if that process annoys you most. And if question 3 is a no, you already know what your first job is — and that it has nothing to do with AI.

Six processes that pay off fastest in a smaller company

• The shared inbox. Info@, sales@, service@: everything lands in one place and someone spends every morning sorting, forwarding and answering the same questions twice. In most companies this is the biggest pile of repetition there is.
• Documents that come in and get retyped. Invoices, delivery notes, orders, request forms, inspection reports. Every supplier does it differently, which is exactly why a fixed rule fails here.
• Quotes built from a price list with variants. As soon as volume tiers, options and surcharges come into play, quoting becomes look-up work. And that work sits at the front of your sales process, so it slows your revenue down.
• Anything called 'following up'. Open quotes, unanswered questions, documents the customer still has to send. Nobody misses this work when it slips, which is precisely why it slips.
• The standard customer questions outside office hours. Delivery times, opening hours, where-is-my-order, what-does-a-callout-cost. Not exciting, but the highest volume.
• Data that has to stay current in two places at once. Your own system and your largest customer's portal, or your stock and your web shop. Quiet work that only gets noticed when it goes wrong.

What these six have in common: high frequency, digital input, someone who currently reads and forwards, and a result you can express in something other than hours. So they score high on the five questions above by themselves. That is not a coincidence — it is the same test, written the other way round.

Four processes that look automatable and are not

• Decisions where a single mistake costs you a customer and nobody notices. Prices, discounts, promises about delivery dates, anything legally binding. You can have it prepared, but not sent. Doing it anyway trades hours for a risk that is far more expensive.
• Work that happens a few times a year. Year-end close, the annual price list round, the subsidy application. The build takes just as long as for a daily process, and the return is a fraction.
• Processes that run differently for every customer or project and are written down nowhere. Automation makes nothing faster there; it only makes the mess faster. Write down how it should go first, then build.
• Anything that starts with 'we need to get our data in order first'. That may be entirely fair, but then call it what it is. It is a clean-up project with its own price and its own lead time, and on its own it saves nobody any time.

If an agency says yes to all four, that is not ambition, that is a sales pitch. We would rather tell you something is better left manual than build it and watch you switch it off a year later.

Calculate it in orders and margin, not only in hours

The standard sum is hours times hourly wage. There are two holes in it. The first: that hour usually does not disappear. The employee stays on the payroll, so you only save something once those hours go somewhere else — into more orders, into work you currently outsource, or into the peak you would otherwise have hired for.

The second hole: almost everyone uses the wrong hourly figure. According to CBS, an hour of work cost a Dutch employer an average of 47.60 euros in 2025 — that is labour cost per hour actually worked, including employer contributions and paid holiday and sick leave. It varies by sector: business services 51.70 euros, industry 51.10, construction 45.30, transport and storage 44.40, trade 39.50 and hospitality 27.20. And an average worker puts in 1,436 hours a year, not 2,080. (Source: CBS, The labour market in figures 2025.)

The better sum is about your orders. How many quotes went out the same day last month, and how many after three days? What is your margin on an order, and how many more orders could you handle without hiring? How often did you have to redo something because wrong information went out? Those are the numbers that show, six months in, whether the investment made sense.

What it looks like in practice

Two Dutch examples. Both were published by the agency that built the automation, so these are the builder's own figures, not independent measurements. Read them as a picture of what such a project looks like, not as a promise about what you will get. They are also not Socialo projects.

The first is a Dutch internet and services company that, after years of growth, was spending two to four hours a day managing its info@ and service mailboxes: labelling, forwarding to the right colleague and clearing up the leftovers. There was one hard condition: part of the mail contains sensitive data, so the company flatly refused to work with the well-known AI tools. They chose a model that runs locally instead. It labels and routes, and for mail with an invoice attached it first checks whether this is an existing supplier and whether the price deviates by more than ten percent. Existing supplier: booked in. New supplier: on to a human. When in doubt the system sends a colleague a message asking them to take a look, and simply leaves the mail sitting in the inbox. The reported outcome is around three hours per working day, over 760 hours a year. (Source: case page on synai.eu.)

The second is a growing food producer whose product count exploded through acquisitions and new lines. Suppliers change their ingredient lists constantly, and every change had to be checked, applied and synchronised with customers' portals by hand. Two extra people had been hired just for that. First the whole process was mapped, from supplier contact through to entry in the customer portals, and only then were the checking, updating and synchronising automated. So no new system — just tying together what was already there. Reported result: fifteen percent less time on ingredient management, the two extra full-time roles no longer needed, and no more entry errors in customer portals. (Source: case page on tideconsulting.nl.)

What both examples have in common matters more than the numbers: the process was mapped first, the exception goes to a human, and no new system was bought.

Buy off the shelf or have it built?

National statistics help here too. CBS measures not only whether companies use AI, but where that AI comes from. Among Dutch companies with 10 to 50 employees, the share that had AI built by an external party fell from 40 percent of AI users in 2021 to 21 percent in 2025. Off-the-shelf commercial software went from 5 to 15 percent of all companies in that size class over the same period: a threefold increase. (Source: CBS, ICT use by businesses, 2021 to 2025.)

Read that honestly: most of what companies do with AI today, they simply buy. That is often the right call. Our rule is simple — what exists off the shelf, you buy off the shelf. Custom-building what you can buy for a few tens of euros a month never pays for itself, and we say that even when it costs us the job.

What is left over is the seam between your systems. Your package does its thing, your customer's portal does its thing, your mailbox sits in between, and your way of working is written down nowhere. That part does not exist off the shelf, because it is different at every company. That is where custom work belongs, and only there.

Why it usually is not the money that stops you

One of the most useful figures in the 2025 CBS data is not about AI use but about the brake on it. In no sector surveyed is money the biggest brake — a lack of knowledge and experience is, and across all sectors together knowledge holds companies back roughly four times as hard as cost does. In specialist construction the gap is even wider. (Source: CBS, ICT use by businesses, by sector, reporting year 2025.)

The European version of the same survey shows how pronounced this is in the Netherlands. Of Dutch companies with 10 to 49 employees that have ever considered AI, 79.5 percent name a lack of relevant expertise as a barrier and 22.0 percent name cost. The European average is 70.9 against 38.8 percent. Mind the denominator: these are percentages of companies that have ever considered AI, not of all companies. (Source: Eurostat, table isoc_eb_ai, 2025.)

Across four editions of that same series something visible happened. The objection 'too expensive' halved between 2021 and 2025, from 43.8 to 22.0 percent, and 'not useful for our business' fell from 22.9 to 13.4 percent. At the same time, concerns about data protection rose from 24.3 to 45.1 percent and legal uncertainty from 22.3 to 36.0 percent. In short: the objections you can remove with a price are disappearing, and the objections that need a conversation are growing.

What it costs and the order to start in

Our price list looks like this. A small project of one to two weeks costs between 300 and 5,000 euros. A medium project of three to six weeks sits between 5,000 and 12,000 euros. A large project from six weeks upwards runs from 12,000 to 25,000 euros. Maintenance and optimisation afterwards are monthly, from 100 euros, cancellable month to month. The AI usage itself is passed on at cost plus margin.

The order matters more than the size. Pick one process that scored high on the five questions and that can be finished within six weeks. Write down before you start what you are going to measure — how many emails land in the right place by themselves, how many quotes go out the same day, how often something had to be redone. Measure after eight weeks. Only then start the second process. That is how you build evidence inside your own company, and that is the only evidence worth anything.

• Step 1: a free first conversation. Getting to know each other, and honestly establishing whether there is anything to gain at all.
• Step 2: an AI scan. A review of how your business runs, with a concrete report: which processes, in which order, with what expected return.
• Step 3: a proposal covering what we build, what it costs and what it is expected to deliver.
• Step 4: building in weeks, connected to what you already use.
• Step 5: delivery and maintenance for a fixed monthly amount, no annual contract.