Home / AI solutions

AI — solutions by problem

AI solutions: which one your problem actually needs

Almost every list of AI solutions is organised by technology — models, agents, forecasts. That is the wrong entrance, because you do not have a technology problem, you have a work problem. This page inverts it: you start with what is going wrong in your business and end at the category of solution that fits. Including the ones that get sold often and deliver rarely.

By Tom Joseph · Last updated: 18 August 2026

Free consultation Pricing

In short

Nearly every workable AI solution in a Dutch SME falls into six categories: extracting information from documents, answering questions from your own documents, forecasting, sorting and routing incoming work, visual inspection, and executing multi-step work. Which one you need follows from the symptom, not the technology — and most businesses recognise their own problem faster in a complaint from the floor than in a product name. We start with a €2,500 audit that translates the symptom into a category and attaches a figure to it, before anything is built.

From symptom to category

Which problem do you recognise?

You do not have a technology problem, you have a work problem. Start with what is going wrong.

Retyping
Documents
Orders and invoices arrive as PDF and someone keys them in. The most predictable application in the SME segment.
Finding
Your own documents
The answer costs half an hour of searching plus the two people who know. Showing the source is the condition.
Forecasting
Too late or too much
Purchasing runs on experience and a spreadsheet, and breaks at a peak. Two years of history is the floor.
Sorting
One inbox
Everything lands in one place and someone divides it. Less visible than a chatbot, and usually worth more.
Looking
Visual inspection
Checking by eye depends on who is standing there. The scarcity is examples of what goes wrong.

The entrance

Start with the symptom, not the technology

Ask a supplier what AI solutions exist and you get a product list. That list is not wrong, but it is useless for choosing, because two very different problems can carry the same product name and one problem sometimes fits three categories at once.

The useful question is: which action does someone here perform by hand, every day, that follows the same pattern each time? That is where AI means something. Actions that are different every time and require judgement are not automation candidates, however advanced the model.

Retyping

"We retype data out of documents"

Orders, invoices, delivery notes or forms arrive as PDF, email or scan and someone keys them into your system. This is the most common and most predictable AI application in the SME segment.

The solution extracts the fields, checks them against what you already know — does this supplier exist, is this order number valid — and produces structured data. It works well because the outcome is objectively checkable: the amount is right or it is not.

Where it goes wrong: businesses expect 100% and get 95%. The question that matters is not whether it is flawless, but whether the 5% it is unsure about visibly reaches a person rather than passing through silently.

Finding

"Nobody can find anything in our own documents"

The knowledge sits in manuals, contracts, quotes and old project folders, and answering a customer question costs half an hour of searching plus the two people who know.

The solution answers questions from your documents and shows the source alongside. That last part is the whole condition: an answer without a location cannot be checked and therefore is not trusted, rightly.

This is also the category where quality depends entirely on how your documents are organised. If three versions of the same contract exist and nobody knows which applies, the system gives three answers — the problem was never the searching.

Forecasting

"We order too late or too much"

Purchasing, staffing or production runs on experience and a spreadsheet, and it breaks at a peak, a promotion, or a supplier running late.

The solution learns the pattern from your own history — season, day of week, promotions, lead times — and produces a number to plan against. It pays off where the error costs money in both directions at once: lost sales on one side, stock standing still on the other.

The hard requirement is history. Two years of usable data is a realistic floor; below that the model mostly predicts noise. If you do not have it, collecting data is the first job and not building a model.

Sorting

"Everything lands in one inbox and gets divided by hand"

Email, requests, reports or complaints arrive in one place and someone reads through them to decide who gets what. The work is not difficult, but it is constant, and it delays everything behind it.

The solution classifies what comes in and routes it onward — with an urgency, a category and an owner. It is less visible than a chatbot and usually worth more, because it removes the waiting time before the real work starts.

Looking

"We check by eye and occasionally miss something"

Products, packaging or parts are checked visually, and the quality of that check depends on who is standing there and what time it is.

The solution assesses images for defects. This is the category with the heaviest requirements outside the model itself: consistent lighting, a fixed camera position, and enough examples of what can go wrong. That last one is usually the constraint — you have thousands of good units and often ten of each failure type.

Executing

"It is more than one step and there is a decision in it"

The task runs across systems: read something, check it against your ERP, take a decision, record it somewhere and reply. This is where the word agent belongs.

It is also the most expensive category, and the reason is not the model: an agent needs access to your systems and permission to act. Agent projects are integration projects wearing an AI hat, and they should be budgeted that way.

If the individual step is not yet reliable, the chain certainly will not be. Start with the weakest link on its own.
Decide in advance what the agent may do alone and what needs sign-off. That is a business rule, not a setting added later.

Rarely works

Solutions that get sold often and used rarely

Three things recur in quotes that rarely deliver what they promise in an SME setting:

A website chatbot as the first project. Visible, easy to demo, and it usually solves no problem you had — unless you demonstrably receive many of the same questions.
A dashboard that gives "AI-powered insight". If nobody looks at the current dashboard, a smarter one changes nothing. The problem is the decision, not the chart.
A custom-trained model where an existing one would do. Training your own is sometimes right, but it is an expensive default and should be a reasoned exception.

Always required

What every category needs, whichever it is

Whatever the category, the same three conditions apply, and projects almost always fail on one of them rather than on the technology.

A decision that changes. If nobody acts differently because of the output, it is a report, not a solution.
Data that already exists. Not perfect, but present. Building a model on data you still have to start collecting is two projects budgeted as one.
Someone who owns the exceptions. Every solution has them. With nobody named, the floor resolves them informally and use dies out within a quarter.

What it returns, and who builds it

This page identifies the category. Two questions follow from it, and both are covered elsewhere.

For whether it pays — by department, by company size, including what the Dutch schemes cover — see AI for business: where it actually pays off. For how such a solution gets built and taken into production, see AI automation.

FAQ

Frequently asked questions

What AI solutions exist for a business?

In practice six usable categories, and you choose on symptom rather than technology.

Extracting information from documents, answering questions from your own documents, and forecasting.
Sorting and routing incoming work, visual inspection, and executing multi-step work.
The question that makes the choice: which action does someone perform daily by hand that follows the same pattern each time?

Which AI solution delivers fastest?

Usually document processing, because the outcome is objectively checkable and the work visibly recurs daily.

The amount is right or it is not — so the effect is measurable without argument.
Sorting and routing incoming work often comes second and is worth more than it looks.
Do not expect 100%. The question is whether uncertain cases reach a person rather than passing through silently.

Do we need a lot of data for an AI solution?

It varies sharply by category, and it is the question asked too late most often.

Forecasting needs history — two years of usable data is a realistic floor.
Document processing and question answering work on what you already have, provided it is organised.
Visual inspection mainly needs examples of what goes wrong, and those are usually scarce.

Is a chatbot a good first AI solution?

Rarely, unless you demonstrably receive many of the same questions.

A chatbot is visible and easy to demo, which is not the same as useful.
If your questions vary widely, it moves the work rather than removing it.
Answering from your own documents — with sources shown — is usually the more useful version of the same idea.

What does an AI solution cost?

It varies by category, and the technology is rarely the largest line.

Price is driven by how many systems are touched and how many exceptions must be handled.
With us: a €2,500 audit translating symptom into category with a figure attached, a €20,000 proof of concept on your own data, production from €50,000.
Multi-step agents are the most expensive category, because they are integration projects wearing an AI hat.

Know the symptom but not the category?

The €2,500 audit translates what is going wrong into a concrete solution category with a figure attached — and says plainly when none of them fits.

Book a conversation