At the Rise of AI Virtual Chat in 2020, Fabian Westerheide interviews Dr. Ulli Waltinger, head of Siemens' Machine Intelligence Research Group and technology head of the Siemens AI Lab, about industrial applied AI. Waltinger explains that after the big-data hype and a second AI hype, Siemens' management set up a top-down program called Company Core Technologies with an analytics-and-AI track of around 200 to 230 researchers, trained managers company-wide on core concepts, and placed application centers close to business units such as mobility, energy and healthcare so AI teams work directly with customers. He describes roughly 80 percent of proof-of-concept projects stalling before production, and names C-level sponsorship, early involvement of operations staff, and including customers from the start as the factors that raise the odds of a project reaching deployment. He then explains why trust and robustness matter more in B2B industrial AI than in consumer AI, since industrial systems increasingly make autonomous decisions with physical and financial consequences, and describes Siemens' techniques for measuring model robustness and explainability, including domain-adaptation scoring, Bayesian graphical models and rule-based approximations of black-box models.
Key takeaways
- Ulli Waltinger says roughly 80 percent of Siemens' AI proofs of concept never reach production deployment.
- He identifies C-level sponsorship, early involvement of operations staff, and early customer involvement as the main factors that get a project past the proof-of-concept stage.
- Siemens ran a company-wide training program teaching managers hands-on AI concepts like CNNs and backpropagation, not just data scientists.
- He argues industrial B2B AI needs robustness and explainability more than consumer AI because systems increasingly take autonomous, physically consequential actions.
- Siemens measures model robustness through domain-adaptation scoring (e.g. temperature scaling) that flags when input drifts outside the training domain.
Detailed notes
Why B2B AI integration lags behind the hype
Waltinger and Westerheide open by noting AI moved from a bullish 2014-2017 investment hype into a 2019 "sobering up," where the gap is not that research or AI itself is insufficient, but that too few companies have actually integrated it into products. Waltinger says he has lived through two hype waves at Siemens, first big data, then AI, and that Siemens deployed neural networks in steel production and a natural-language question-answering system for energy assets around the same time as IBM Watson, well before AI became a boardroom term.
Building AI readiness top-down at Siemens
Waltinger describes Siemens' Company Core Technologies program, initiated by the management board, which includes an analytics-and-AI track with roughly 200 to 230 dedicated researchers plus many business colleagues. Alongside this, Siemens ran a company-wide training program, "Applying Digitalization to Our Business," teaching managers hands-on concepts such as neural networks, CNNs, activation functions and backpropagation so the whole organization speaks the same language about the technology, not just data scientists and engineers. AI teams and application centers were then placed regionally, close to specific verticals such as mobility, energy management and airport services, so products are developed together with customers rather than designed centrally and handed over.
From proof of concept to production
Waltinger estimates around 80 percent of Siemens' AI proofs of concept end at the POC or MVP stage rather than reaching production, a rate he compares to general industry figures cited by firms like Gartner. He identifies three factors that raise a project's chances: consistent C-level attention on the specific use case, involving the operational staff who will eventually own the deployed system from the very beginning, and including the customer in the process from the start rather than only at handover. As an example of scale, he lists use cases spanning shared services (speech recognition, document extraction, a virtual assistant called "Carl"), robotic process automation, financial risk assessment, predictive maintenance, visual quality inspection, energy control via reinforcement learning, and quality monitoring in additive manufacturing.
Trust and robustness in industrial AI
Waltinger contrasts consumer AI, where imperfect results (a bad search result, an imperfect recommendation) are broadly accepted, with industrial B2B AI, where systems increasingly take autonomous actions with real physical and financial consequences, such as collaborative robots making their own stop decisions. He argues this demands two properties beyond raw accuracy: robustness, since models trained on past data can silently lose accuracy when deployed in changed conditions (different lighting, season, camera angle) domain adaptation problem, and explainability. Siemens addresses robustness partly through domain-adaptation scoring, using techniques like temperature scaling to flag when input has drifted out of the model's training domain, and adapts models without full retraining. For explainability, Waltinger describes using Bayesian graphical models and knowledge graphs, including for drug-repurposing work in healthcare, and training simpler rule-based models to approximate a black-box model's output as a way of surfacing bias without opening the underlying network.
Security and opening the ecosystem
Asked about hacking risk, Waltinger says the biggest vulnerability is usually not the machine-learning system itself but human factors like infected USB sticks or phishing links, and argues no organization can achieve full security alone; Siemens participates in a cross-corporate security charter built on transparency about incidents. On strategy, Waltinger distinguishes "depth of value" (optimizing internal efficiency and throughput) from "breadth of value" (deciding what Siemens will not build itself and instead sourcing from ecosystem partners to capture network effects), arguing this shift, giving up some value to enable transactions between partners and their customers, is a needed mindset change for Europe's AI ecosystem more broadly.
Closing reflections
A live audience poll during the talk found the top reasons for lagging B2B AI integration were insufficient company-wide AI awareness, lack of frameworks and methodology for implementation, and a mix of fear of making mistakes and insufficient top-management prioritization. Asked to summarize ten years in AI and look ahead a decade, Waltinger cites a New York Times report that two-thirds of men in a study preferred a mild electric shock to being alone with their thoughts for 50 minutes as evidence that consumer attention-economy incentives have become unhealthy, and argues the AI community should instead orient its work around the UN Sustainable Development Goals and values that serve more than a single business partner.
Quotes
It's not that AI's not good enough or the research is not good enough. It's the opposite. It's integration is lacking behind.
Fabian Westerheide
If you don't have C-level attention on your use case, it will never see a light.
Ulli Waltinger
There will be no hundred percent security, I'm sure.
Ulli Waltinger
Questions answered in this talk
Why do most industrial AI proof-of-concept projects fail to reach production?
According to Ulli Waltinger at a Rise of AI Virtual Chat in 2020, around 80 percent of Siemens' AI POCs stall, most often due to missing C-level attention, leaving out operational staff who will own the system, or not involving the customer from the start.
How does Siemens train managers to understand AI?
Waltinger explains at Rise of AI in 2020 that Siemens ran a company-wide, hands-on training program teaching managers core concepts like neural networks, CNNs and backpropagation, so the whole organization shares a common language about the technology.
Why does trust matter more for industrial AI than for consumer AI?
According to Ulli Waltinger at Rise of AI in 2020, industrial AI increasingly makes autonomous decisions with real physical and financial consequences, such as collaborative robots making their own stop decisions, which demands measurable robustness and explainability that consumer AI applications do not require to the same degree.