AI in the German Mittelstand looks healthy on paper. Manufacturing AI adoption is rising, and most manufacturers we meet have already run AI pilots. Very few of those pilots are running on a line today.
The causes repeat. Across plants in India, the United States and now the German Mittelstand, four of them account for most stalled programmes, and only one is technical.
Four failure patterns behind stalled AI pilots
01 The data foundation was never built
The model was trained on an export. Live data arrives late, incomplete, and in a shape the pipeline did not expect. The demonstration held. The line did not.
02 Ownership ended with the pilot team
The system kept running and stopped being consulted. Until someone audits it, that is indistinguishable from working.
03 No integration with MES, ERP or the quality system
Output that does not appear where the work already happens does not get used. An operator will not open a second screen while the line is moving.
04 The operators were not consulted while it was built
So they worked around it. That is rational conduct from people accountable for output, and training does not correct it.
Only the first is close to a data science problem, and it is the least interesting of the four.
What the Mittelstand AI adoption data shows
Published figures make this look like a national failing. AI adoption across German companies is quoted at 41%. In manufacturing it sits closer to 28%, and the gradient by company size is steeper than the sector gap, which is where the Mittelstand feels it most.
Adoption by sector shows the same split. The work of describing things has automated faster than the work of making them.
Automotive is the exception at 70.4%, and it is also the sector that has worked with external engineering partners for thirty years. That pattern deserves attention before the next build-or-buy discussion.
Strategy coverage is thinner than adoption. 21% of mid-sized German companies have a formal AI strategy, and 64% of those already using AI operate without one.
We read that as sequencing rather than commitment. Most programmes select the use case first and establish the state of the data second. That order reliably produces a working demonstration and a failed deployment. It is the most common reason AI projects fail before they reach production.
What we changed in our method?
We start from the assumption that the hard part is everything around the model.
So the opening question is not which use case is most interesting. It is which use case sits on data whose lineage can be proven, inside a system an operator already has open, with an owner accountable for the outcome rather than for the model. That is how we sequence manufacturing AI adoption from pilot to production.
What German manufacturers ask before AI adoption?
These questions arrive before anyone asks about accuracy. Not having an answer ends the conversation, so here are ours.
Where does our data sit?
In EU regions. We deploy inside your Databricks or AWS environment, so your data does not leave your tenancy.
Do you train models on our data?
No. Your data is used for your systems only. That is contractual, not a policy statement.
Does the EU AI Act apply?
Most manufacturing quality and operations use cases fall outside the high-risk category. Automated visual inspection on a line is generally not high risk. We assess the classification with you, per use case, and document it rather than assuming it.
What about the works council?
Consultation goes into the project plan in week one rather than being discovered at go-live. We bring a standard briefing pack for the Betriebsrat.
The works council answer is the one foreign suppliers most often get wrong. A system that watches a line touches the terms under which people work. Programmes that meet that fact at go-live lose months, and they usually blame the technology.
On not having an in-house AI team
Most Mittelstand firms tell us they lack the in-house capability. Germany has 109,000 open IT roles and will lose 3.9 million working-age people by 2030, so the constraint is real and it is not closing.
It is also the normal case. 89% of German SMEs already use external AI solutions rather than building in house. The question is which AI implementation partner, not whether to use one.
We do not sell headcount. We work in pods: a fixed core team that stays with you, plus capacity that scales with the phase of the work. The people who scope it are the people who deliver it.
A first engagement: the paid diagnostic
We do not offer free proofs of concept. A free pilot signals that our time carries no value, gets assigned to whoever happens to be free on your side, and opens no procurement path, which is the real bottleneck in German enterprise.
We sell a paid diagnostic instead. Six to eight weeks at a fixed price, delivered as a written assessment covering current state, data readiness, three prioritised use cases with business cases, and an implementation roadmap with costs. The fee is credited against the first build phase if you proceed within ninety days.
Paying for an assessment is normal industrial practice. It is how German engineering consultancies already sell.
If you have a folder of pilots that stopped
Ascentt builds and runs enterprise AI for automotive and manufacturing companies. Munich, Detroit, Indore. Elite Databricks Partner and AWS Partner.
We are in Germany from 22 to 29 September with the Nasscom delegation, and in Munich either side of it. We would rather spend an hour on why a pilot stopped than present a capability deck. Book the hour.
