At Rise of AI 2025 in Berlin, Christian Winterfeldt, CTO & Director AI Solutions at Dell Technologies, argues that organizations fail with AI not because they lack talent or tools but because they lack a common strategy. He divides the AI market into three phases: a 30-year-old pre-generative-AI market of roughly 2,500 highly specified use cases, the recent large language model market that made organizations more efficient, and an emerging enterprise market built on agentic inferencing that changes business outcomes rather than just optimizing tasks. He illustrates this with a procurement manager, Anna, whose agents monitor markets, prices and contracts overnight, and a product manager, Lisa, whose work depends on six core AI capabilities including bringing corporate context to LLMs and managing trusted data. Winterfeldt presents a "strategy house" model, developed and applied inside Dell itself, that moves from vision and use-case selection to a scalable foundation of culture, data, technology and ecosystem, and finally to execution. He closes with Dell's own on-premise AI platform results and a checklist for building a production-grade AI environment, insisting organizations start with a business goal rather than with technology.
Key takeaways
- Winterfeldt frames the emerging enterprise AI market around agentic inferencing, where multiple agents act on an organization's behalf before a manager starts the day.
- A standard LLM knows only about 1% of an organization's internal knowledge, so context must be added through RAG or fine-tuning.
- Dell's internal sales chatbot reached a 96% usage rate among staff after being tuned with reinforcement learning on real answers.
- Running AI inferencing on-premise rather than in the cloud can cut costs by 75% at scale, with an ROI within 12 months for Dell's own platform.
- About 80% of enterprise data sits on-premise, so Winterfeldt argues AI workloads should be brought to where the data already is.
Detailed notes
Three AI markets and the shift to agentic inferencing
Winterfeldt separates AI into three markets: a pre-generative-AI market from 30 years ago covering about 2,500 highly specified use cases, each needing its own architecture; the large language model market built over the last two years with tens to hundreds of millions in investment, producing versatile LLMs used for chatbots, writing and image generation; and a new enterprise market focused on business outcomes rather than optimization. He introduces agentic inferencing through the example of Anna, a procurement manager whose agents continuously track the market, monitor price fluctuations, review vendor contracts and generate negotiation recommendations before she starts her day, arguing this changes whole organizational processes, not just individual efficiency.
Six capabilities organizations need to master
Using product manager Lisa as an example, Winterfeldt lists capabilities organizations must build: bringing corporate context to LLMs (since a standard LLM knows only about 1% of an organization's knowledge) via RAG or fine-tuning; deciding what data to trust and which agents or models may access it; building simple applications such as chatbots without needing to be a programmer; using analytics agents to monitor product performance against competitors; and optimizing physical infrastructure such as production lines and assembly assets. He notes that many organizations run successful pilots that never reach production, and describes a manufacturing company where IT had no knowledge of AI initiatives running in other departments, calling this misalignment the common failure pattern.
The strategy house model
Winterfeldt presents a "strategy house" as a mirror rather than a fixed model: start with vision and strategy (what makes the organization special, which processes can improve with AI), then select use cases, build a scalable foundation covering culture, data, technology and ecosystem, and finally execute with board-level backing. At Dell, this combined a top-down question (what differentiates the organization) with a bottom-up employee survey that produced over a thousand candidate use cases, narrowed to about ten. Dell identified four strategic areas for AI investment: go-to-market, services, R&D and supply chain, and built a sales chatbot now used more than four times a day by over 96% of staff after refining it with reinforcement learning on the underlying data.
Culture, data and ecosystem as the hard part
Winterfeldt describes culture as harder than technology: training managers through encouragement and education, and starting with the roughly 10-20% of employees eager to try new tools before bringing along the rest. He stresses that data is the core problem, since the same product spec can be interpreted differently by R&D, marketing and customer support, requiring reliable "data products" and pipelines. He states that no single company can cover every layer alone, so an ecosystem of partners, including Applied AI in Germany and Hugging Face, is necessary.
Building a production-grade AI platform
Winterfeldt describes the technical stack: running inferencing on-premise on physical servers can save 75% of costs compared with cloud at scale, with an ROI within 12 months on Dell's own AI platform investment. He notes 80% of enterprise data sits on-premise and that AI should be brought to where the data is, including the edge. He outlines a system built with Nvidia, the "AI factory," combining data sources, a data platform with data contracts and vector databases for RAG, a model hub with lifecycle management and governance, and IT-operated infrastructure, which Dell's engineers can stand up as a demo in 30 minutes.
Q&A highlights
Asked how to overcome cultural resistance in traditional Mittelstand companies, Winterfeldt says to start with the roughly 10-20% of employees eager to experiment, show results, and bring the rest along through education rather than mandates. Asked how to demonstrate ROI to convince a manager, he describes setting KPIs and sub-KPIs, citing a customer-service chatbot measured on resolution time and customer feedback quality, and recurring employee surveys tracking time saved. Responding to a comment that starting with a business pain point may be more effective than starting with a business goal, he agrees in principle but notes that organizations investing during times of crisis tend to emerge from it in stronger shape than those that hold back.
Quotes
This is not a model. This is kind of a mirror. So take a picture and really compare this what you're doing at home.
Christian Winterfeldt
Don't start with technology. Start with a business goal worth scaling.
Christian Winterfeldt
Organizations who are investing in times of crisis are going out of the crisis exponentially better than other organizations who don't, who just keep their money together.
Christian Winterfeldt
Questions answered in this talk
Why do AI pilots often fail to reach production?
According to Christian Winterfeldt at Rise of AI 2025, pilots often stall because organizations lack a common AI strategy and because departments, such as IT and the teams running pilots, are misaligned and unaware of each other's initiatives.
How much can on-premise AI infrastructure save compared to the cloud?
Christian Winterfeldt states at Rise of AI 2025 that customers running inferencing on-premise on physical servers save 75% of costs compared to cloud services once they scale to thousands of users, with an ROI on Dell's own AI platform within 12 months.
How should a company choose which AI use cases to pursue?
Christian Winterfeldt describes at Rise of AI 2025 combining a top-down question about what differentiates the organization with a bottom-up employee survey, then narrowing candidate use cases, in Dell's case from over a thousand to around ten, based on data availability and the six core AI capabilities.