At Rise of AI 2024 in Berlin, Feiyu Xu, Non-Executive Director of Airbus Group and board member at ZF, argues that generative AI marks a shift from AI that classifies and recognizes to AI that creates content, which she says brings it structurally closer to human intelligence and makes adoption easier because companies no longer need large teams of data engineers to label data. She notes that at least seven of the world's ten largest public companies by market cap are AI-first, and situates the moment against the 2023 OpenAI leadership crisis, the Hollywood writers' protests against AI, and the EU AI Act. She lays out a framework for positioning a company in the generative AI ecosystem across infrastructure, models and applications, data ownership and access, and industry vertical, and recommends companies start with business-process automation before moving to more intelligent products. She then details persistent gaps: large language models lack deep engineering, chemical and geometric domain knowledge, struggle with real-time short-term forecasting, and cannot reliably handle escalated customer complaints or troubleshooting, so she recommends a holistic data governance strategy and treating GenAI as one part of a larger system rather than a complete solution.
Key takeaways
- Generative AI adds content creation to prior AI's classification and recognition strengths, which Feiyu Xu says brings it structurally closer to human intelligence.
- GenAI lowers the adoption barrier because companies no longer need large teams of data engineers to label training data.
- Most European companies sit only at the application layer of the GenAI value chain, which creates dependency on foreign infrastructure and model providers.
- Feiyu Xu recommends companies start GenAI adoption with business-process automation before moving to smarter products and hardware.
- Current large language models lack strong domain data in engineering, chemistry and geometry, and are not suited to short-term, real-time forecasting.
Detailed notes
The context: hype, power struggles and regulation
Xu opens by noting the surge of investment and capability in generative AI and foundation models since late 2022, helped by the open-source movement, which she says lets both large companies bypass slow procurement processes and startups run early experiments cheaply. She points to the prior year's leadership crisis at OpenAI as a conflict between researchers who wanted to slow commercialization and better understand the technology's "black box" and executives pursuing large-language-model revenue, alongside Hollywood writers protesting AI over job displacement and the European Union's newly released, risk-based AI Act.
Why GenAI is a bigger shift than earlier AI
Xu argues generative AI differs from prior AI, which was mainly good at classification and recognition, by adding the ability to create content such as images, text, signals and answers, which she says brings machines structurally closer to how humans recognize, learn and create. She gives three reasons industries see particular potential in GenAI: first, this classification-plus-creation combination mirrors human capability; second, adoption is easier because companies no longer need large teams of data engineers to label training data, since models can learn more directly from raw data; and third, foundation models offer one broad model serving many capabilities, compared with the narrow, single-purpose AI interfaces still typical even at large companies such as Google's separate search, image search and translation products. She also argues regulation is necessary because even model developers cannot fully predict the emerging capabilities of large models, so laws alone are not enough and processes and technical monitoring are also needed, including across national borders.
Positioning a company in the GenAI ecosystem
Xu offers a framework for companies and investors to position themselves: first, the value chain layer, from computing infrastructure and hardware, to large-language-model providers, to applications built on top, noting that most European companies operate only at the application layer, which she flags as a dependency risk; second, the data dimension, covering data ownership versus data access and the legal conditions that govern whether a company may use text, image, video, audio or code data it holds; and third, the industry vertical, such as finance, automotive or aerospace. She recommends companies use these dimensions to define a clear position, for example building multimodal foundation models specifically for the automotive industry.
Where to apply GenAI: processes, products and promotion
Xu recommends companies start applying AI to business processes such as customer relationship management with chatbots, supply chain and logistics network optimization, and human resources, before moving on to making products themselves more intelligent, for example autonomous driving or in-car digital assistants, or infusing intelligence into hardware so that device manufacturers become combined hardware-software companies, or building entirely new AI-dependent products such as drones. She urges companies to actively promote this progress internally and externally rather than treating it modestly, noting that marketing AI-driven improvements is a continuous effort many European companies underinvest in.
The gap between demos and scaled production
Xu warns that running convincing generative AI proofs of concept is very different from delivering scaled, high-quality processes and products, and that hiring AI experts before establishing data access and governance risks losing those hires quickly. She stresses working with legal, IT and business units to determine which data can be used, under which conditions, and for which purposes, since obligations differ for continuous improvement of existing products versus creating new ones, and for structured versus unstructured data. She identifies specific technical gaps: current large language models have strong data coverage in software and medicine but little in engineering, chemistry or geometry, which limits applications such as automotive simulation; trustworthiness and hallucination depend heavily on how complete a model's domain data is, not only on training and instructions; and models are not close enough to real time to support short-term forecasting such as supply chain demand prediction. She also cautions that generative AI should not be used to handle escalated, angry customers or full troubleshooting, saying it can competently cover roughly 80 percent of fact-driven customer requests but the remainder still needs human experts.
Recommendations for scaling GenAI
Xu argues an unscaled proof of concept counts as a failure, not a success, and that scaling requires a holistic corporate strategy: clear data strategy and governance, defined ownership across ideation, deployment and continuous improvement, and a pricing and marketing approach that starts by confirming user satisfaction before monetizing further. She recommends leveraging the broader AI ecosystem rather than building every technology in-house, and advises companies choosing a foundation-model provider to check what data it was trained on, its governance and testing practices, and its energy use and sustainability. She closes by saying vertical, industry-specific AI and AI embedded into hardware are strongly needed, alongside continued vigilance about deepfakes and misinformation.
Questions answered in this talk
Why is generative AI a bigger shift for industry than earlier AI?
According to Feiyu Xu at Rise of AI 2024, generative AI adds the ability to create content on top of the classification and recognition that earlier AI was good at, which she argues brings machines structurally closer to human intelligence and makes adoption easier since companies no longer need large data-engineering teams to label training data.
Where should a company start when adopting generative AI?
According to Feiyu Xu at Rise of AI 2024, companies should start by applying GenAI to business processes such as customer relationship management, supply chain optimization and HR, then move on to making their products and hardware more intelligent, and should scale from proofs of concept only after confirming users are satisfied.
Can generative AI handle customer service escalations and troubleshooting?
According to Feiyu Xu at Rise of AI 2024, generative AI can competently cover an estimated 80 percent of fact-driven customer requests, but escalated, angry customers and full troubleshooting should still be handled by human experts.