Almost every business I speak to this year has the same item somewhere on the list: "do something with AI". It is rarely more specific than that. And the vendors are not helping — every SaaS tool you already pay for has grown an AI button, and every consultancy has an AI offering, and none of it tells you whether it will do anything for the process that is actually costing you time.
So let's be concrete. This is about integrating AI into software you already run: your portal, your internal tool, your order flow. Not about building an AI product. Where it pays off, where it does not, what it costs, and what has to be true before you connect anything to a model.
The three integrations that actually pay off
In business software — not consumer products — nearly all the value sits in three patterns.
1. Reading unstructured input. Invoices, order e-mails, PDFs from suppliers, scanned forms, free-text tickets. Someone reads them and types the contents into a system. A model can extract structured fields from that mess at accuracy levels that were not possible three years ago. This is the single highest-return AI integration in most Dutch businesses, and it is the least glamorous one.
2. Routing and classification. Which department does this ticket belong to, is this an urgent complaint or a question, does this application meet the criteria for a fast track. Decisions that are cheap individually but consume a person's day in aggregate.
3. Search and retrieval over your own material. Answering "what did we agree with this client about delivery times" from a decade of contracts and e-mail threads. This is what people usually mean by a "chatbot on our own data", and it works — provided the material is findable in the first place. If your documents live in five systems and three shared drives, that is the project, not the AI part.
Notice what is not on that list: generating text nobody asked for, dashboards that summarise things a human already understood, and assistants that answer questions nobody was asking. Those demo well and get switched off within a quarter.
What an AI integration actually costs
Two separate costs, and people usually only budget the first.
Build cost. For a well-scoped integration into existing software — document extraction into your ERP, classification in your ticket flow — realistically €8,000 to €25,000. That covers the integration itself, the prompt and evaluation work, an interface for the humans who check the output, and the plumbing to get data in and out safely. Retrieval over your own documents is usually €15,000 to €40,000, because the majority of that project is getting the documents into a usable state, not the model call.
Running cost. Model usage is metered per token and it is cheaper than most people expect: a few hundred euros a month for a process handling thousands of documents is common. The cost that surprises people is human review. In the first months you check a meaningful share of the output, and you should. Budget that time explicitly instead of discovering it.
The comparison that matters is not "what does AI cost" but "what does the current manual step cost". Two hours a day of copying data is roughly €25,000 a year in loaded salary. Against that, an €18,000 build that removes 80 percent of it pays back inside a year. If the manual step costs two hours a month, no AI integration will ever be worth it, and anyone telling you otherwise is selling something.
The four questions to answer before you connect anything
Where does the data go? If you send customer data to a model, you are sending personal data to a processor. Under the AVG that needs a legal basis, a processing agreement, and clarity on where the processing physically happens. The major providers offer EU-region processing and contractual guarantees that data is not used for training — but only if you configure it that way and sign the right agreement. The default consumer tier is not that.
What happens when it is wrong? Not if. A model that extracts invoice totals correctly 97 percent of the time is excellent, and it is also wrong on three invoices in a hundred. The integration is only as good as what happens to those three. Confidence thresholds, a human review queue, and an audit trail of what the model produced versus what was corrected — that is the actual engineering work, and it is what separates an integration that survives a year from a demo.
Can you tell whether it is working? You need a set of real examples with known-correct answers, scored before you ship and re-scored whenever the prompt or the model changes. Without that you are guessing, and you will keep guessing after a provider silently updates a model underneath you.
Are you locked in? Model providers change pricing, deprecate versions and shift terms. An integration written against one provider's specific API shape is a migration project waiting to happen. Written behind a thin internal interface, swapping providers is an afternoon. This is the same vendor lock-in question as always, in a faster-moving market.
Start with the boring one
The pattern I would push almost any business towards for a first AI project: pick the single most repetitive place where a human reads something and types it somewhere else. Automate that one, with a review queue, and measure it for a quarter.
It is unglamorous, it will not impress anyone at a conference, and it is the only version of this that reliably returns money. The ambitious projects go much better once the organisation has one working example and a realistic sense of what these systems do and do not do.
Where this fits
An AI integration is, structurally, an integration project: getting data out of one system, through something, and into another, reliably and observably. The model is one component. The unglamorous parts — error handling, retries, audit logging, the review interface — are what determine whether it still runs in eighteen months.
If you have a process in mind and want an honest read on whether AI is the right tool for it, or whether plain automation would do the same job for less, describe the process and you will get a straight answer within 24 hours — including "don't do this" when that is the right answer. For Dutch businesses, how Zoyare works locally covers the AVG and hosting side in more detail.