Which processes should an SME automate with AI — and which ones not

The question that lands on the table is almost always: what can we do with AI? That is the wrong question. It produces a list of possibilities nobody can act on, because there is no criterion underneath it.

The useful question is narrower: which process in my company costs time structurally, goes wrong regularly, and is predictable enough to hand over? In the average mid-sized company there are three or four of those. The rest is noise.

That distinction is exactly where it goes wrong. Among Dutch companies that considered AI but did not adopt it, 73 percent name a lack of experience as the main reason (Statistics Netherlands, December 2025). The barrier is not the technology. Not the budget. It is not knowing where to start.

This article gives you the criteria.

Five criteria: is this process a candidate?

Run a process past these five. If it scores four or five, it is a candidate. If it scores two or fewer, leave it alone — no matter how irritating the work is.

1. It happens often enough Volume × frequency. Fifty invoices a week is a process. Five a month is a task — you do not automate that, you just do it. The cost of building and maintaining has to run off against something.
2. The outcome is verifiable Can you see afterwards whether it went right? A booking proposal sitting next to the purchase order takes three seconds to check. A price recommendation that appears out of nowhere never gets checked. Without a control point you are building a black box.
3. The process is stable Not: "we usually do it this way, but Jan does it differently." If the process changes every month, the automation changes with it — and then it costs you more than it saves.
4. The data lives in a system Not in someone's head, not in a chat group, not in a spreadsheet on a desktop. What is not recorded cannot be handed over — not to software, and not to a new colleague either.
5. An error is affordable What does it cost when it goes wrong? A misrouted email costs ten minutes. A misrouted payment costs more. The more expensive the error, the tighter the human control — or the clearer the conclusion that this process should not be automated at all.

Where is the AI, actually?

This is where the market consistently blurs the line. "AI automation" is sold as one thing, while it is two. The difference determines what you are paying for.

In most of the processes Den Otter Solutions builds, eighty percent is rule work and twenty percent is AI. That is not modesty — it is the reason it works. Spread the AI layer across the whole process and you end up with something you can neither verify nor maintain.

For your selection this means: a process without a judgement step does not need AI at all. A connection between your webshop and your accounting package is a connection. Fine to build — just do not call it AI, and do not pay an AI price for it.

Four processes that lend themselves to it

• Invoice processing. High volume, layouts that differ per supplier, and a hard check afterwards (the purchase order). AI reads, the system matches, you sign off. The textbook case.
• Inbound email and customer questions. Classifying and routing, or preparing a draft reply. The judgement step sits in understanding the question; the rest is rule work.
• Order to accounting. Mostly rule work — which is precisely why it makes a good first step. Retyping is the most expensive form of administration there is.
• Exception alerts. Stock below level, margin slipping, a customer going quiet. The system detects, AI summarises what it is about, a person decides.

Four processes to leave alone

• Processes that are already broken. If your purchasing process rattles, you automate the rattling. Fix first, hand over second. This is the most common mistake.
• The customer relationship itself. Negotiating, handling a complaint that hurts, agreeing a price. Drafting a reply: fine. Handing over the conversation: no.
• Decisions with financial or legal consequences and no control point. Releasing payments, approving contracts, granting credit. AI may prepare. A person signs.
• The work that is your craft. The exception you handle better than anyone is not an inefficiency — it is your margin. Automate the thirty emails around it, not the judgement itself.

The sequence: process first, tool second

Most automation projects start with a tool. Someone sees a demo, gets enthusiastic, and goes looking for a process to fit it. That is backwards.

The sequence that works: map the process as it actually runs (not as the manual describes it), apply the five criteria, identify where the judgement step sits, and only then choose what to build it with. That first step often reveals that part of the process can simply be deleted. That is the cheapest automation available.

And decide up front what it has to deliver. Not in "time saved", but in hours, errors and lead time — numbers you can see in your dashboard once it runs. Automation you cannot measure is automation you cannot defend when the next investment comes up.

Where to start

Pick the process with the highest volume and the hardest check. Not the most impressive one — the most boring one. Invoice processing, order entry, mail routing. Build one thing, let it run for a month, measure what it saves. After that you know enough to choose the second, and you have something to show internally.

Den Otter Solutions delivers this work remotely for clients in the Netherlands and abroad, at a fixed price per process, built on your own accounts and handed over documented.

The bottom line

Automation is a selection question, not a technology question. Five criteria, four good candidates, four no-go areas. What remains is usually less than you hoped for and worth more than you expected.

And the rule underneath it: process first, tool second, and AI only where it adds something. Not everything has to be AI.