AI for SMEs: where to start
Most guides on AI for SMEs start with a list of tools. 25 apps, 3 subscriptions, a prompting course. Your team tries them, everyone writes emails faster, and 6 months later nothing in your numbers has moved.
The tools are not the problem. The starting point is. If you want AI for your small business to change something real, start with the work you sell: the process with a price per file or per hour, that comes back every week, and that your experts still do by hand. That is where most of the value sits, and where most of the mistakes happen.
This article is for the owner, managing partner or operations lead of a Dutch B2B service company with 10 to 200 people. Subsidy advisors, tax, vehicle tax and customs specialists, training and certification providers, compliance consultancies, accountancy and administration offices, inspection firms, brokers. You have no technical team of your own, and your growth is capped by how many people you can hire.
It is not for you if you are still looking for an idea, building a consumer app, mainly comparing hourly rates, need a single Zapier flow, or are preparing a public tender. Better articles exist for those.
In this article you will learn why so many SMEs get stuck at the start, the 3 levels of AI in a service company, how to pick your first process in 5 steps, what it costs, and how other Dutch service firms go about it.
Key takeaways
• Start with a process, not a tool. Work with a price per unit or per hour, with volume, done by hand by your experts, is your best starting point.
• AI in a service company has 3 levels: personal tools for everyone, 1 automated expert process, and that process as a product your customers pay for.
• Level 1 is cheap and you can start this week. It makes people faster, but rarely changes your margin.
• Level 2 holds the most value and the most mistakes. A person approves wherever a mistake is expensive.
• According to CBS, 27% of Dutch firms with 10 to 50 employees used AI in 2025. Of the firms that considered AI and decided against it, 73% named lack of experience as the reason.
• McKinsey finds the difference in organisations that redesign how the work runs, not in organisations that add a tool.
• Your first step costs nothing: list your processes and put a price per unit on each one.
• Then pick 1 of 3 routes: do it in house, hire someone, or have it built outside.
Why so many SMEs get stuck at the start
The good news first: AI use is growing fast. According to Statistics Netherlands (CBS), 17% of all Dutch firms used AI in 2025, 2 times as many as in 2023. Among firms with 10 to 50 employees it was 27%, and among firms with 50 to 250 employees 45%.
Now the less good news. Almost 3 in 4 firms with 10 to 50 employees did not use AI yet. And of the firms that seriously considered AI and still decided against it, 73% named lack of experience as the main reason. 49% named privacy, 42% the legal consequences.
Behind those percentages sits a familiar story. The research firm Dialogic interviewed 25 SMEs for the Dutch Ministry of Economic Affairs. Too little knowledge and skill was by far the most named barrier. A manufacturer with about 30 staff put it plainly: what to do with it, and where to begin, is not clear.
What the owners in that study ask for stands out. Not another tool. They ask for someone who can help map everything properly. That is a precise wish, and it is exactly the starting point most guides leave out.
In short, service firms get stuck for 3 reasons:
• Lack of knowledge: nobody in house can judge what AI can and cannot do in your work.
• No starting point: there are 100 possible uses and no way to choose.
• Privacy: you handle client data and you do not know where it ends up when someone pastes it into a tool.
All 3 can be solved. Just not by starting with the tool.
The 3 levels of AI in a service company
It helps to stop seeing AI in your company as 1 thing. A service company has 3 levels. They differ in cost, in risk, and in who owns them.
Level 1: personal AI tools for everyone
This is ChatGPT, Microsoft Copilot, Claude or Gemini on a business account. Your people use it to write emails, summarise documents, draft meeting notes and set up a first version of a text. The cost is a subscription of a few tens of euro per user per month.
Start here this week. It is cheap, the risk is low on a business account, and your team learns what AI does well and badly. You will need that knowledge the moment you start on level 2.
The limit of level 1 is just as clear. Every employee gets a little faster, but the process stays the same. The same steps, the same handovers, the same queue at the experienced colleague who checks everything. Your margin rarely moves.
Level 2: 1 expert process automated, with a person who approves
Here you take 1 process your experts now do by hand, and you redesign it. A vehicle tax calculation, a subsidy application, an inspection report, a file check. The AI does the searching, the retyping and the first draft. Your expert reviews the result and approves wherever a mistake is expensive.
This level holds the most value, because it touches exactly the work you sell. If a file takes 6 hours today and 1.5 hours after, the same team can handle up to 4 times as many files on that step. That is where your margin moves.
It is also the level with the most mistakes. Not because the AI is bad, but because the process was never mapped properly, the exceptions are unknown, or nobody owns it when it goes wrong. An automation that gets 90% of cases right and quietly gets the other 10% wrong costs more than the manual work.
Which processes pay back fastest is covered in our AI automation payback playbook. This article is about the step before that: how to choose.
Level 3: your process as a product your customers pay for
Once your process runs reliably, the next question is obvious. Why should only your own team use it? An importer who pays you 100 euro per vehicle tax calculation may prefer a portal where they run 200 calculations a month themselves, on a subscription.
That is level 3: your expertise as software your customers buy. It is a revenue line that does not grow with your headcount. It is also the most expensive and longest route, and it only works once level 2 is in place. Building a platform on a process you do not yet control yourself is an expensive way to find that out.
There is a reason to think about this now. The calculation your client pays 100 euro for is exactly the kind of task their own AI assistant will attempt within a few years. The service firm that offers its expertise as a product first keeps the client. The rest becomes a line on a comparison sheet.
Want to put this on paper for your own company? With the free AI product requirements template you describe 1 process: what goes in, what comes out, who decides and what it may cost. Useful preparation for any conversation with whoever will build it.
Start with the process, not the tool
In McKinsey's State of AI survey from August 2026, 80% of respondents say AI has improved their personal productivity. Yet only 37% see a meaningful effect on earnings, the same as a year earlier.
The small group that clearly profits from AI, about 6% of respondents, does something different. Nearly 3 quarters of those high performers have fundamentally redesigned their workflows. They do not put AI on top of the existing process. They rebuild the process around it.
For a service company that is good news. You do not need a 40 page AI strategy. You have a process you run every day, with a price on it. That is your starting point.
A list of 40 AI tools is a fine way to fill a Friday afternoon. It is not a strategy.
A good first process has 4 traits:
• It has a price per unit or per hour: per file, per application, per calculation, per certificate.
• It happens often: tens to hundreds of times a month, not 3 times a year.
• Your experts do it by hand today, the people you find hardest to hire.
• It follows rules: legislation, criteria, tables or a fixed method, even if that method lives only in someone's head.
A process that meets all 4 has visible arithmetic. You know what it costs now, how often it happens, and who gets stuck when volume grows.
AI for SMEs: where to start in 5 steps
This is the order that works, even if you do it without help. Expect a few days of work, spread over 2 weeks.
Step 1: list your processes
Take a sheet of paper and write down every process your clients buy or that you need to deliver. Not the departments, the processes: intake, building the file, calculation, review, report, invoicing. A service company of 20 people quickly gets to 10 to 20 processes.
Next to each process, write who does it. If the same name appears next to 3 processes, you have found your first bottleneck.
Step 2: put a price per unit on each process
For each process, write 3 numbers: how often per month, how many hours each time, and what an hour of the people doing it costs you. Multiply. Now you have a monthly amount per process.
This is dull work, and it is the most important work in this whole article. Without those numbers you choose on feel. And on feel almost everyone picks the process that annoys them most, instead of the one that costs the most.
Step 3: pick 1 process
Pick the process with the highest monthly amount that also follows rules. Not 3 processes at once. 1 process that runs well teaches you more than 3 pilots that are half done.
Torn between 2? Pick the one where a mistake is least expensive. You learn with less risk, and your team gets used to the new way of working before much is at stake.
Step 4: map the process as it really runs
Not as the quality manual describes it, but as it looks on a Tuesday afternoon. Which systems get opened, which data gets retyped, where people wait, who decides when in doubt, and which exceptions come back every week.
If 5 advisors work 5 ways in 5 tools, that is your first finding. You cannot automate what is not settled. So the first step is often 1 source of truth: 1 place with your rules, criteria and examples, which both your team and your AI run on.
How to build that map step by step is covered in how to map the AI opportunities in your workflow.
Step 5: decide who builds it
With a mapped process and a price on it, there are 3 routes: do it in house, hire someone, or have it built outside. Which fits depends on who you already have. Further down we put them side by side.
A worked example
This is an example, not a client. Take a subsidy advisory firm of 12 people, 8 of them advisors. The firm writes 600 applications a year. A first draft of an application takes an advisor 6 hours on average: checking the scheme, collecting the client's data, writing the text, setting up the budget.
That is 3,600 hours a year on first drafts. At an internal cost of 90 euro an hour, that is 324,000 euro a year for this one piece of work alone.
After automating it at level 2, it looks like this. The intake arrives through a fixed form. The AI looks up the scheme's criteria in the firm's own knowledge base, fills in the data and writes a first draft. The advisor reviews, adjusts and approves. That now takes 1.5 hours per application.
• Hours before: 600 x 6 = 3,600 hours a year
• Hours after: 600 x 1.5 = 900 hours a year
• Capacity freed: 2,700 hours a year, more than 1.5 full time advisors at 1,650 productive hours a year
• AI running cost: say 2 euro per application, so 1,200 euro a year
• Build: 18,000 to 45,000 euro, once
• Operations: 1,250 euro a month, so 15,000 euro a year
What is it worth? That depends on what you do with the 2,700 hours. Fill them with more applications and your revenue grows without hiring another advisor. Leave them empty and you mostly get a calmer team. Both are choices, but only the first changes your results.
Watch what is not in the sum. The time of your best advisor to explain the process. The weeks in which the team adjusts. The exceptions that only show up after 3 months. Count them in, or your arithmetic is too pretty.
How Dutch service firms use AI in 2026: practical examples
Below are 3 examples of how Dutch service firms put AI to work. They are anonymised, typical examples, not client cases.
Example: an administration office
An administration office of 25 people has AI pre sort incoming invoices and bank transactions and propose the bookings. The employee only sees the items where the AI is in doubt, with the reason attached. That is level 2: a person approves the exceptions, the rest flows through.
Example: a training and certification provider
A training provider for technical staff tracks thousands of certificates, each with an expiry date. AI reads the certificates, checks them against the requirements per role and prepares the reminder emails. Later the same overview becomes a portal for the clients themselves. That is level 3.
Example: an inspection firm
An inspector records findings by voice on site. AI turns the recording into a report in the fixed format, with the right standards attached. The inspector reads it and signs. The evening typing after the last inspection disappears.
From our own work
3 anonymised examples of what we built or mapped:
• A vehicle tax (BPM) file processor: from 4 to 8 hours per file to under 5 minutes.
• A subsidy matching engine: from hours of expert searching to seconds.
• A subsidy advisory where 5 advisors worked 5 ways in 5 tools. The first question there was not which AI, but how 5 ways of working become 1 source of truth.
The pattern is the same every time. It starts with 1 process with a price. It gets mapped first. And a person keeps deciding wherever a mistake is expensive.
Which AI tools are useful, level by level
The question which AI tools are really useful has a different answer per level.
• Level 1: a business account for ChatGPT, Microsoft Copilot, Claude or Gemini. If you work in Microsoft 365, Copilot is the obvious choice. If you work in Google Workspace, Gemini.
• Level 1, specific: a transcription tool for meetings and a tool that makes your own documents searchable.
• Level 2, simple: n8n, Make or Zapier for steps without expert judgement, such as passing a form to your CRM.
• Level 2, expert process: custom software that connects to Exact, AFAS or your sector tool, with a knowledge base, logging and an approval step.
• Level 3: your own platform with accounts for your clients, billing and operations.
A warning on level 2. Many vendors sell rented n8n or Make flows, with an entry price of 1,500 to 5,000 euro. For a simple connection that is fine. For an expert process where a mistake is expensive, you often miss the knowledge base, the exceptions, the logging and someone who picks up the phone when it stops.
So before you sign, ask who owns the flow, where your data lives, and what happens when the underlying platform changes its price.
Competitive advantage from AI does not come from level 1
Any competitor can buy the same ChatGPT subscription tomorrow. Level 1 is therefore not a competitive advantage, it is the new baseline. Skip it and you fall behind. Do it and you are level with everyone else.
The advantage sits in what a competitor cannot buy: your process, your rules, your examples from 10 years of files, and the way your experts decide borderline cases. Once that is captured and automated, you can do something others cannot. More volume with the same team, a shorter turnaround, or a lower price per file at the same margin.
At level 3 the advantage becomes a product. Your client logs in with you instead of sending an email. That is harder to replace than an advisor who also works for the firm next door.
Privacy without standing still
49% of the firms that considered AI and decided against it named privacy as a reason. That is fair. You handle client data, and a free tool where an employee pastes a file is a real risk. The answer is just not to do nothing.
• On level 1, use only business accounts where your data is not used to train models, and write down what may and may not go in.
• On level 2, choose processing in the EU where you can, and sign a data processing agreement with every supplier.
• Log what the AI did and who approved what. Then you can explain it later to a client or a regulator.
• Let a person decide wherever a mistake is expensive or affects an individual.
This is not legal advice. For the GDPR and the EU AI Act in your situation, it pays to have it checked properly once.
Do it in house, hire, or have it built
With a mapped process and a price on it, 1 question remains: who builds it?
Do it in house
If you have someone technically strong with time to spare, start yourself. For level 1 and simple flows you do not need an outside team. The risk is that it becomes a side project of 1 person. When that person leaves, nobody knows how it works.
Hire someone
An experienced developer in the Netherlands earns 65,000 to 95,000 euro gross a year, according to recruiter Search Company. With about 30% employer costs, that is roughly 85,000 to 125,000 euro a year, and finding one often takes months. That pays off if you have an ongoing roadmap. For 1 process it is often too much, and your new colleague still has to learn the process.
Have it built outside
Then you choose a team that maps the process, builds it and keeps it running afterwards. Ask every supplier 3 questions. Who exactly works on it? What happens after launch? And what if month 1 disappoints?
This is how we answer them. It starts with a free 20 minute conversation. Then, if it fits:
• The Blueprint: 2,500 to 10,000 euro depending on the number of processes, in 10 to 20 working days. Your process mapped as it runs, bottlenecks priced in hours and euro, AI opportunities ranked, a roadmap with fixed prices per milestone, and a first rough prototype in your own repository. It is yours to keep, even if you build the rest in house.
• The Foundation: 1 source of truth your company and its AI run on. A fixed price from the Blueprint, 4 to 8 weeks.
• The Automations: a fixed price per automation, 18,000 to 45,000 euro per major process, 1 bottleneck per milestone, integrated with Exact, AFAS or your sector tool, with a person approving wherever a mistake is expensive.
• The Platform: your expertise as a product for your clients, staged, from 25,000 euro, in year 2.
• Operations from month 4: monitoring, a named engineer on the alert, the AI cost per unit capped, and a 1 page report every month. Planning value 1,250 euro a month per automation. The response time is set in the Blueprint.
• Exit after month 1: pay for what is delivered, keep everything.
The same 2 engineers work on it, by name, from the Blueprint to year 2. All prices above are planning values. The fixed price per milestone is set in the Blueprint.
What you can do this month
AI in your company does not start with a tool but with a process. Give your team a business AI account this week, that is level 1. Then list your processes, put a price per unit on each, and pick 1. That is the start of level 2, and that is where most of the value sits.
Want to describe that one process properly before you talk to anyone? Download the free AI product requirements template and fill it in for your first process.
Codelevate turns Dutch B2B service companies into software companies: we map the expert process you run by hand, automate it so the same team handles the volume, and build the platform your customers pay for, with 2 engineers by name from the blueprint to year 2.
We have worked from Amsterdam and Sofia since 2019. 24 published cases, 8 of them B2B platforms. Clutch 5.0 over 10 reviews, and 7 of those 10 clients came by referral.
1 of the 2 engineers reads it with you in 20 minutes, and you get 1 of 3 answers in writing the same day: do it in house, hire, or the Blueprint.
.svg){kind=small}
.svg)
