The day the industry stopped asking whether. It started asking how.
The VDMA ran their AI practice day as a World Café in Frankfurt — twelve themed tables in Raum 10, plus a thirteenth at the centre that the organisers called the "Wilde 13," the break table where the prepared agenda ran out and people said what they actually meant. The room held around 200 people from the Mittelstand: engineers, production managers, heads of IT, a few CAIOs, and enough consultants to make the engineers nervous. They were not there to be pitched. They were there to compare notes.
That is a different kind of meeting. When a room of people who build physical things sits down to work out what AI is actually for, the conversation goes somewhere a keynote never reaches. The questions were sharper. The frustrations were more specific. And the thing most people agreed on, without quite saying it aloud, was this: the industry is no longer debating whether AI belongs in mechanical engineering. It is figuring out where to start and how to do it without breaking what already works.
This page is my attempt to bring the structure of that day forward — the diagrams, the frameworks, and the honest numbers — so they are useful beyond the room.
The room
Twelve tables, one room, one question per table. The diagram below shows the layout as it was on the day. Each table ran for forty minutes before participants rotated.
AI in technical sales: what the pitch decks promise versus what actually lands in a complex B2B sale.
AI agents in admin: connecting ERP and Office systems without breaking the workflows people actually use.
Talking AI — regulation and legal certainty: where the EU AI Act draws the line, and what the room still did not know.
AI across the value chain: from procurement through manufacturing to assembly — where the data gaps actually live.
AI agents in production: what an agent can actually do on the shop floor today, and what still requires a human.
Governance: who is responsible when an agent acts — the question nobody had fully answered.
Raising the digitisation level and using data systematically: you cannot govern what you cannot see.
Edge, embedded and physical AI: inference running close to the machine, not in a data centre three countries away.
The unprogrammed table. The agenda ran out here. The questions got better.
Strategy is not adoption speed
The strategy panel — "KI: Einfach machen oder strategisch handeln!?" — brought together SSI Schäfer, Zollern Engineering, and others who have moved past the question of whether to implement AI. The conversation that emerged was about something harder: how to ensure that AI strengthens the company's strategic position rather than simply automating what was already being done. The title was a question because, for most of the room, the answer was not yet settled.
One marker of how far the conversation has matured: CAIO — Chief AI Officer — is now accepted vocabulary in this community. Not a curiosity, not a job title borrowed from Silicon Valley. A role that several companies in the room have created, filled, and are actively learning to use. That is a meaningful shift from twelve months ago.
Read more on the CAIO question
“Strategy is not 'how fast can we adopt?' It is where this leads — and whether AI strengthens your strategy or quietly replaces it.”
The EU AI Act: most companies don't know where they stand
Dr. Gorenflos held the EU AI Act table and fielded questions for the better part of an hour. The honest summary of what I heard: most companies in the room do not know which risk tier their applications fall into. They know the Act exists. Several have read summaries. A few have received legal advice. But the question of whether a specific system — a quality control model, a procurement recommendation agent, an anomaly detection pipeline — is high-risk, limited-risk, or minimal-risk under the regulation was, for the majority, still open. That is not a criticism of the people in the room. It is the honest state of the compliance landscape eighteen months into implementation. The regulation is real. The taxonomy is harder to apply than the taxonomy suggests.
Six things an AI agent actually needs
The room converged on six success factors for deploying AI agents in Maschinen- und Anlagenbau. These are not aspirations — they are the conditions under which agents fail or succeed in production. Apuna builds all six into every engagement.
Context & grounding
The agent needs the right context before it can reason. Without grounding in the specific machine, process, and business situation, a model is pattern-matching against its training data — not reasoning about your problem.
Reliability
Reproducible results, not lucky ones. An agent that performs well on Tuesday and differently on Friday is not production-ready. Reliability is an engineering property, not a model property — it lives in the deployment, the configuration, and the test suite.
Permissions
Fine-grained permissions; read-only first. An agent that can read your ERP should not be able to write to it until there is a deliberate decision to grant that right. The blast radius of a mistake should be small by design.
Human in the loop
A person decides. The agent recommends; the human authorises. This is not a concession to caution — it is the correct architecture for any action with real-world consequences. Automate the research; keep the human on the decision.
Traceability
Full audit trail. Every recommendation, every action, every data source that contributed to a result must be logged and retrievable. In a regulated industry, this is not optional. In any industry, it is what earns trust.
Guardrails
Schema-level output validation. Before a model's output reaches a downstream system, something must verify that it matches the expected structure. Not 'is this a good answer' — that is a separate problem. 'Is this output structurally valid before it touches production data.' A JSON that fails a schema check stops at the gate.
The four stages: from data inventory to agentic procurement
The companies making the most progress were not the ones that had launched the most AI features. They were the ones that had built the foundation first and added capability in a sequence that each step made possible. Four stages came up repeatedly as the right order of operations.
Transparency dashboard
What data do we have, and where does it come from? Before a model touches anything, there must be an honest inventory: which systems produce data, what quality it is in, how often it is updated, and who is responsible for it. Most companies find that this stage alone surfaces problems they did not know they had.
API foundation
Reliable, queryable data surface. Once the inventory is honest, the data needs to be accessible through a stable interface — one that downstream systems, agents, and dashboards can depend on. This is not glamorous work. It is the work that makes everything else possible.
Recommending agent (read-only)
The agent recommends; a human decides. At this stage, the agent has access to the data foundation and can form and explain recommendations — maintenance windows, procurement suggestions, anomaly alerts. It cannot act. That is a deliberate constraint, not a temporary limitation. Read-only first earns the trust that makes the next stage possible.
Predictive maintenance + agentic procurement (with HITL loop)
At this stage, the agent can initiate actions — but not without a human in the loop. The HITL loop is not a bottleneck; it is the architecture. The agent does the research and the legwork; the human makes the call.
The HITL loop in agentic procurement
- Low stock detected
- Agent retrieves quotes
- Agent synchronises delivery dates
- Human authorisation
- Order placed
Predictive maintenance: reference architecture
The architecture for predictive maintenance is not complex in principle. It is hard in practice because the data inputs are rarely clean, rarely in the same format, and rarely owned by the same team. The diagram below shows the logical flow — from data inputs through the agent's reasoning loop to a traceable recommendation.
Data inputs
Live data / digital twin
Real-time sensor data and the current state of the digital twin — the authoritative source of what the machine is doing right now.
History
Historical operating data: past faults, maintenance records, repair histories. The context that makes the current anomaly meaningful or unremarkable.
Documentation (RAG)
Manuals, technical specifications, service bulletins — retrieved at inference time via retrieval-augmented generation so the agent reasons against the actual documentation, not its training data.
Business data (ERP)
Spare parts inventory, supplier lead times, planned production windows. A maintenance recommendation that ignores the production schedule is not useful.
Flow
- Anomaly detected
- Maintenance agent correlates signals across data inputs
- Agent forms hypothesis
- Hypothesis evaluated against context (history, documentation, ERP)
- Recommendation issued — with reasoning and full audit trail
From product to data-driven service
The business case for AI in Maschinen- und Anlagenbau is not 'replace human labour.' It is 'build a service business on top of the product business you already have.' The machine is in the field. The data it produces is yours. The question is whether you use it.
New service models
Predictive maintenance contracts, condition-based service agreements, performance guarantees tied to uptime data. The machine becomes a monitored asset, and the monitoring becomes a revenue line. Companies that have made this shift report that service revenue now grows faster than equipment revenue — and with better margins.
Lower service costs
Reactive maintenance is expensive: emergency parts, urgent travel, production downtime while the technician is in transit. Predictive maintenance replaces the emergency with the scheduled visit — parts ordered in advance, technician briefed before arrival, downtime planned rather than suffered. The cost reduction is not marginal. It is structural.
Customer retention
A customer who receives a maintenance alert before they know the problem exists experiences the manufacturer as a partner, not a vendor. That relationship is harder to displace than a price comparison. The data connection is also a switching cost — not an artificial one, but a real one: the history, the calibration, the baseline are all in your platform.
Better products
Fleet data from machines in the field is the most honest product feedback a manufacturer can receive. Failure modes that the test bench missed. Usage patterns the design team did not anticipate. Wear profiles that inform the next generation. The companies building this feedback loop into their product development cycle are getting smarter faster than those that are not.
Three things I am taking away from the room
IoT delivers data. Platform makes it usable. Agent makes it value-creating. In that order. Skipping the platform step and going straight from data to agent is where most projects stall.
Agents need context, permissions, and traceability — not just a model. The six success factors the room agreed on are not nice-to-haves. They are the conditions under which agents work reliably rather than impressively.
For machine and plant builders, the platform is the lever from product to service. The companies that are moving fastest are the ones that invested in the data foundation first and are now pulling the service business lever. The ones still debating the pilot are still building on sand.
The optimisation reflex
There is something specific about the German engineering culture that makes the AI question harder here than in the US or China. It is not technophobia and it is not conservatism. It is a professional formation built around making what already exists work better. The optimisation reflex — find the inefficiency, eliminate it, raise the yield — is what produced some of the most precise manufacturing in the world. It is also what makes it difficult to ask whether the system being optimised is the right system to be running.
In Frankfurt, I heard this most clearly at the governance table. The questions were about how to make AI fit the existing process — how to ensure agents behaved within current workflows, current compliance structures, current reporting lines. These are legitimate questions. They are also the wrong first question. The first question is whether the process is worth fitting AI into, or whether the process is the artefact of constraints that AI removes. The companies that are moving fastest in this space are the ones that asked the second question first. They are building new service models, not better versions of old processes.
China and the United States are not optimising their way into AI leadership. They are building. The distinction matters: optimisation improves what you have; building creates what you don't yet have. The Mittelstand's competitive advantage has always been depth of craft — the thirty-year relationship with the customer, the machine that runs for twenty years because it was built to, the service technician who knows the specific installation by memory. AI does not replace that advantage. But building the data platform that makes the advantage legible — monitorable, defensible, scalable — is not optimisation work. It is foundation work. And it has to come first.
Cost signals from the room
A speaker at the VDMA Praxistag KI
“AI cost will soon surpass the cost of a qualified human who is fully responsible for the work. This is not a warning about AI being expensive — it is a statement about the accountability premium. When AI cost crosses that line, the question of who is responsible for the output becomes commercially unavoidable.”
Romano Roth, Zühlke
“GitHub Copilot has moved to token-based billing. Developers are being observed spending up to €50 per hour on tokens. FinOps-for-AI is now a real discipline — not a future concern. Usage without visibility is a budget problem waiting to happen, and it is happening now.”
A VDMA TOP-5 speaker from a leading industrial pump manufacturer put two framings into circulation that I think will stick. The first: the "gondola or on foot?" metaphor for AI adoption speed — you can reach the summit either way, but the gondola does not build the legs. The choice of pace is a strategic decision, not just a technical one, and organisations that ride the gondola exclusively may find themselves at the summit without the capability to do anything useful once they arrive.
The second: a "new Moore's law for AI agents" framing — the capability curve for agents is compounding in a way that changes the planning horizon. The agent that is adequate today will be significantly more capable in eighteen months. Building for the current capability ceiling is building for obsolescence.
He also cited the numbers that came up at several tables: 70% of AI projects fail due to poor data quality and integration — not bad models, not wrong use cases, but the pipes. And approximately zero percent productivity gain from bolting AI onto broken processes. These are not pessimistic numbers. They are the numbers that tell you where to start: fix the foundation, then add the intelligence. Foundation-first industrial AI work of this kind — the sort seen at leading industrial manufacturers — is one of the better examples of the approach in practice.
The questions are right. The work is clear.
The industry is asking exactly the right questions. Who governs the agent? Which risk tier does this application fall into? How do you build a service business on top of a product business when the data is there but the platform is not? What does it actually cost, and who is accountable when it goes wrong?
The answers to most of those questions exist. They are not mysteries. They are engineering decisions and organisational decisions that companies are making and figuring out, in rooms like Raum 10 in Frankfurt. The missing piece is not insight. It is the discipline to build the foundation rather than optimise around its absence — to do the unglamorous data work, the permissions architecture, the audit trail, the HITL loop — before adding the impressive capability on top.
That is the work. It is the right work. And it is available now.