What it actually is, where it earns money in a 20–250 person European business, what a first project really costs, and what the EU rulebook asks of you as it stands in August 2026.
Conventional software does what you tell it. Someone writes down the logic — if the order is over €5,000 and the customer is new, flag it — and the machine follows it forever. That is still most of the software in your business, and much of it is excellent.
Machine learning inverts that. You show the system thousands of past cases along with what happened, and it works out the pattern itself. A spam filter is the everyday example: nobody hand-wrote the rules for what spam looks like. The filter saw enormous numbers of messages already marked spam or not spam, and learned which combinations of words, senders and formatting tend to go with each. When a new message arrives in a shape nobody anticipated, it still makes a decent call.
This matters commercially for one reason: it lets you automate judgement calls that were never possible to write down. Which of these 400 open invoices will go past 60 days. Which quote is priced below where it will actually win. Which pump is starting to behave unlike itself. Rules can't capture those well. Patterns in your own history can.
Two clarifications that save a lot of confused meetings. Not all AI is machine learning — a rules engine or a classic chess program is AI and learns nothing. And generative AI is a different shape of tool: an LLM produces content — text, code, images — whereas the machine learning discussed here produces a number or a decision. Both are useful. They are not interchangeable, and they fail in different ways.
Learns from examples where the answer is known and recorded. Predicts a number or a category.
Roughly seven in ten real applications. Almost certainly where your first project sits.
Finds structure with no answer key — natural customer groupings, transactions unlike all the others.
Good for exploration and anomaly detection. Harder to point at a number.
Learns by trial, reward and penalty — the approach behind game-playing systems and robotic control.
Fascinating, still niche. Not a sensible first SME project.
You don't need the maths. You do need six terms, because each one hides a commercial risk that lands on your desk rather than the data scientist's.
The past examples the model learns from. Everything downstream is limited by it.
Your risk: if your history only reflects how you used to work, the model will faithfully reproduce how you used to work.
The inputs — order value, customer age, day of week, region. Choosing and shaping them is called feature engineering.
Your risk: this is where your team's domain knowledge beats any external expert. Don't outsource it entirely.
The recorded right answer for each past case: did it churn, was it fraud, what did it actually sell.
Your risk: labelling is the classic bottleneck and often the largest line in the budget. Ask about it in week one.
The model memorises the quirks and noise of the training set instead of the real pattern — the student who learned the answers, not the subject.
Your risk: the demo dazzles at 96% and production settles at 71%. This is the single most common way pilots disappoint.
Data held back and never shown during training, used once at the end. Any leakage of it into training inflates the score.
Your risk: a number quoted without a clean test set is marketing, not measurement.
The world moves — new products, new suppliers, new prices — and yesterday's pattern quietly stops holding.
Your risk: models decay silently. Monitoring and retraining are a running cost, not an optional extra.
"What was your test set, and was it separated before you started modelling?" — one question, and you will learn more about a vendor than a whole afternoon of slides.
Adoption across Europe is climbing quickly — one in five EU enterprises with ten or more staff used some form of AI in 2025, up from around one in seven the year before — but it is far from even. Large firms are at roughly 55%, and among smaller companies using generative AI, fewer than a third have it anywhere near their core activities. Most of that gap is not budget. It is that nobody picked a decision worth improving.
A decision is a good candidate when three things are true: it repeats hundreds of times a month, the outcome is already recorded somewhere, and being right most of the time — not always — creates value. Miss any of the three and no amount of modelling will help.
| Function | The question | Where the history lives |
|---|---|---|
| Finance | Which open invoices will run past 60 days? | Ledger and payment history |
| Sales | Which quotes are worth a partner's time this week? | CRM, won/lost outcomes |
| Stock & supply | How much of each SKU will move next month, per site? | ERP sales lines |
| Service | Which inbound message goes to which team, and which is urgent? | Helpdesk archive |
| Production | Which machine is behaving unlike itself before it stops? | Sensor logs, maintenance records |
| Pricing | Which jobs are systematically quoted below the margin we assume? | Quote and job-costing history |
Note the pattern: every one of these is a decision your people already make, with an outcome you already store. That is not a coincidence — it is the selection criterion.
Ignore accuracy percentages. The business case is four numbers and one subtraction:
The word that does the most work there is baseline. You are not competing against zero. You are competing against a competent planner with a spreadsheet, and that person is often quite good. If a project can't state what today's error rate is, it can't claim to have improved it.
A 55-person distributor, €12m turnover, €1.8m of stock, 3,000 active SKUs. Replenishment is planned monthly in a spreadsheet by one very experienced person. Baseline forecast error, measured over the last 12 months: 31%.
The €190k is working capital freed, not profit — the recurring gain is the financing cost on it plus fewer emergency orders and fewer lost sales. Presented properly to a board, this is a working-capital case with a service-level bonus, and it survives scrutiny. Presented as "AI will optimise our supply chain", it does not.
And the uncomfortable truth worth naming in front of your board: a substantial share of first machine learning projects deliver nothing. In our experience they rarely fail on the modelling. They fail because the output landed in a dashboard nobody owned, or because the model's recommendation was politically inconvenient, or because the data turned out to be three systems disagreeing with each other. Budget for those risks, not for GPU time.
Your data becomes an asset with a maintenance schedule. Most SMEs discover on day three that the same customer exists under four spellings across three systems. That clean-up is real work, and it is worth doing regardless of whether the model ever ships — which makes it the safest money in the whole programme.
Someone has to own the decision. Not the model — the decision it feeds. If the planner is not obliged to either accept the recommendation or record why they overrode it, you will never know whether it worked, and it will quietly be ignored within a quarter.
Keep the model where you can reach it. Insist on the trained artefact, the training pipeline and the feature definitions in your own repository and cloud account. A model you can't retrain is a subscription with extra steps.
The AI Act has been amended. The Digital Omnibus on AI — Regulation (EU) 2026/1744 — was published in the Official Journal on 24 July 2026 and entered into force on 27 July, moving several deadlines and easing a few obligations, particularly for smaller firms. Here is the short, practical read.
| When | What applies | In practice |
|---|---|---|
| In force | Banned practices and an AI literacy duty | A short list of prohibited uses has applied since February 2025. The literacy duty has been softened by the Omnibus: you must take reasonable measures to support AI understanding among the staff who operate these systems, rather than guarantee a level of it. |
| 2 Aug 2026 | Transparency — the one that touches almost everyone | If people interact with your chatbot, they must be able to tell it is a machine. Synthetic or manipulated media must be disclosed and AI-generated content marked in a machine-readable way — with a short transitional period into December 2026 for systems already on the market. Emotion recognition and biometric categorisation must be declared to the people exposed to them. |
| 2 Dec 2027 | High-risk obligations — deferred, not withdrawn | Stand-alone high-risk systems (Annex III) apply from 2 December 2027; AI embedded in regulated products (Annex I) from 2 August 2028. Recruitment, worker management, credit scoring and access to essential services sit in this tier — the extra runway is for building governance, not for ignoring it. |
Two things smaller firms should know. The Act's proportionality provisions — simplified technical documentation and a more measured approach to fines — have been extended to small mid-caps, so growing out of the SME definition no longer drops you straight into the large-company regime. And the AI Act sits alongside GDPR, it does not replace it: if you are making automated decisions about people using personal data, your existing data-protection obligations, including on solely automated decisions with significant effects, still apply in full.
For the majority of SME use cases in this article — demand forecasting, message triage, maintenance prediction, quote scoring — you are outside the high-risk tier, and the practical burden is modest: know what systems you have, know what they decide, and be able to explain both. Start that inventory now; it takes an afternoon today and a fortnight later.
Resist the strategy phase. A single real decision, improved and measured, teaches your organisation more than any roadmap.
Pick one repeated decision. Write down what it costs to get wrong and how well you do it today, in a number.
Assemble the history and the outcomes. Expect this to be most of the effort. Expect to find problems worth fixing anyway.
Simplest model that works, held-out test set, compared against the human baseline — not against nothing.
Into the tool people already use, with a human deciding, overrides logged, and a date in the calendar to review the numbers.
Figures in the worked example are illustrative arithmetic for a composite business, not measured client results.
We build production systems that keep running after the consultants leave — and we will tell you plainly when a spreadsheet is still the right answer. Tell us what decision you would improve first.
Where deep-tech AI startups should own the stack, where they should rent it, and where ownership has quietly stopped being an option — with decision protocols.
How AI and software engineering cut avoidable cost and shorten lead times for Irish and EU metal fabricators — with a reference architecture and pilot plan.
How AI governance and AI security differ, what each one guards against, and how to combine them into a single layered defence for enterprise AI systems.