Why is AI use in Czech and Slovak manufacturing so much higher than in Poland or Romania?

The Eurostat data does not explain causes. Both countries roughly doubled between 2024 and 2025, and part of that jump comes from the broader 2025 questionnaire. Treat the gap as a signal of different starting points, not as a ranking of plants.

Can we skip stages and start with predictive maintenance?

You can, on one machine with a clear failure mode and dedicated sensors. Expect the project to rebuild parts of stages 1 and 2 for that machine. Plant-wide AI without stages 1 to 3 usually stalls on missing or unlabelled data.

How long does it take to get from stage 1 to stage 4?

It depends on how quickly reason codes and daily use settle. Analytics needs several months of data collected the same way. We do not publish a typical duration because we have no verified figure.

Do we need data scientists for stages 1 to 3?

No. These stages need production, maintenance and quality people who own the definitions and use the data every day. Specialist skills matter mainly at stage 5.

Does the data layer help with IATF audits?

Yes. Automatically captured stops and counts give a stronger basis for the OEE, MTBF and MTTR objectives that IATF 16949 clause 8.5.1.5 names as examples, and for their monthly trend and review.

Home

Guide · Data and AI · status 22 September 2026

Why only 3–8% of CEE manufacturers use AI, and the data layer that has to come first

Eurostat's 2025 survey puts AI use among manufacturers in Romania, Poland and Hungary at less than half the EU rate. The fix is rarely a better model. It is reliable machine data, collected every shift.

In shortIn 2025, 3.2% of manufacturers with 10+ employees in Romania, 7.7% in Poland and 7.8% in Hungary used at least one AI technology, against 17.3% in the EU (Eurostat). Industrial AI needs complete, timestamped machine states, counts and stop reasons. Build that first: capture automatically, add reasons, use it daily, then analyse, then apply AI to what remains.

The adoption numbers

Enterprises using at least one AI technology, 10+ employees, % of enterprises
Country	All sectors 2024	All sectors 2025	Manufacturing 2024	Manufacturing 2025
Romania	3.1	5.2	1.3	3.2
Poland	5.9	8.4	5.1	7.7
Hungary	7.4	10.4	4.8	7.8
Czechia	11.3	17.6	9.6	16.7
Slovakia	10.8	18.0	8.1	15.8
EU-27	13.5	20.0	10.6	17.3
Germany (reference)	19.8	26.0	16.1	24.4

Source: Eurostat isoc_eb_ai (all sectors) and isoc_eb_ain2 (manufacturing, NACE C), indicator E_AI_TANY, 2024–2025. Eurostat broadened the AI questions in 2025, so part of the 2024→2025 increase is a survey change, not only real adoption.

Two groups stand out. In Romania, Poland and Hungary, manufacturing AI use is under half the EU rate. In Czechia and Slovakia it roughly doubled in a year and is now close to the EU level, although the survey change explains part of that jump.

Across the EU, the most common reason given by enterprises that considered AI but did not use it was lack of relevant expertise, cited by 70.9% of them, followed by unclear legal consequences (52.5%) and data-protection concerns (48.8%) (Eurostat, Use of artificial intelligence in enterprises, 2025). Eurostat does not publish a figure for data quality as a barrier in that release, so we do not quote one.

Why industrial AI projects stall on data

We have no reliable statistic on how many industrial AI projects fail, and we distrust the ones that circulate. The mechanism is easier to show than to count. A model that predicts a failure, recommends a setting or explains a loss needs history. That history must meet four conditions.

Complete. Every stop, including the short ones, with start and end times. A model trained on hand-logged breakdowns never sees the two-minute jams that make up much of the loss.
Labelled. Stops need reasons from a short, stable list. 'Other' as the largest category teaches a model nothing.
Aligned. Machine states, part counts, part numbers, shifts and, where available, process values must share one clock. Aligning a PLC log with a paper shift report after the fact is slow and often impossible.
Long and stable enough. Months of data, collected the same way, on a process that has not been redefined every quarter.

If these conditions are not met, the first phase of an AI project becomes a data project, whether it was planned that way or not. It is cheaper to budget the data layer as its own project, with its own payback in daily loss reduction, and treat AI as a later option on top.

The expertise gapThe 70.9% expertise barrier is also a data-layer argument. Stages 1 to 3 below need production and maintenance people, not data scientists. Stage 5 is the only one where specialist AI skills are essential, and by then the problem is narrow enough to buy in.

A five-stage roadmap

Capture machine states and counts automaticallyEvery key machine reports running, stopped and producing, with part counts, from a signal rather than a pen. Done looks like: for any shift in the last month you can show every stop with its start and end, and the count matches production reporting within a small, known difference.
Put reasons on stopsOperators pick a reason from a short list, at the machine or on a tablet, while the stop is happening. Maintenance owns the list. Done looks like: more than nine in ten stop minutes have a reason, and 'other' is not in the top three.
Use it daily in shift meetingsThe start-of-shift meeting uses yesterday's losses by machine and reason, not memory. Actions are logged with owners. Done looks like: supervisors ask for the screen when it is missing, and the top loss on each key line has an open action.
Analytics on stable dataWith several months of consistent data, look at patterns: losses by part number, shift, changeover type, time since maintenance. This is where MTBF, MTTR and OEE trends become reliable enough for audits and customers. Done looks like: you can answer 'why did line 3 lose four points last month' in an hour, with numbers.
ML and AI on the problems that remainPick a narrow, costly problem that analytics did not solve: a failure mode with warning signs, a quality drift linked to process values, a scheduling problem. Add the sensors or data that problem needs. Done looks like: a model with a measured effect on one named loss, owned by someone in production.

Stages 1 to 3 also produce the evidence that IATF 16949 clause 8.5.1.5 asks for on maintenance objectives; see what auditors expect on OEE, MTBF and MTTR. With energy sub-meters added, the same layer answers customer energy-per-part questions.

Free data

CEE Manufacturing Data Pack 2026

Every indicator on this page for 7 countries, EU-27 and Germany, with Eurostat dataset codes and links. Excel.

Get the data pack →

Checklist: is your data ready for AI?

Answer for one key line, not for the plant
Check	Ready	Not ready yet
How are stops recorded?	Automatically from the machine signal, with timestamps	On paper or typed in at the end of the shift
Are short stops visible?	Yes, down to the minute or below	Only stops longer than a few minutes
Share of stop time with a reason	Above about 90%, 'other' small	Large 'other' or 'unknown' category
Part counts	From the machine, per part number	From the shift report or ERP postings
Common clock	States, counts and process data share one time base	Separate systems, aligned by hand
History	Six months or more collected the same way	Weeks, or definitions changed recently
Daily use	Used in shift meetings; errors get noticed and fixed	Looked at monthly or for audits
Ownership	Named owner for reason codes and definitions	Nobody, or IT only

Thresholds are practical rules of thumb, not published benchmarks.

If most answers are in the right-hand column, an AI project will spend its first months fixing them. That work is worth doing, but call it what it is when you budget and when you apply for funding.

How this links to funding

Several CEE digitalisation programmes name production data collection, MES or IoT as eligible costs, which fits stages 1 to 3 well. Call status changes often. Our sister site lists the schemes by country with dates and official links: CEE factory digitalisation funding. Check the current call documentation with the agency; this is not grant advice.

Where the data layer can come from

Three common routes
Route	Fits when	Watch for
MES	You also need orders, traceability and work instructions on the floor	Longer projects; machine-state capture on older machines may still need extra hardware
Local integrator (PLC and SCADA data)	Machines have accessible controllers and you have in-house IT support	Definitions and dashboards are custom; keep the documentation
Monitoring product (such as TeepTrak)	You want states, stop reasons and counts quickly, including on older machines	Subscription cost; integration with ERP or MES if you need it later

This site is published by TEEPTRAK SAS, which sells production-monitoring and OEE software (450+ factories equipped in 30+ countries; CEE team in Bucharest). We are not neutral about the third route. All three can deliver stages 1 to 3. So can a disciplined spreadsheet on one line, as a first test of whether your teams will use the data.