At Rise of AI 2025 in Berlin, Feiyu Xu, Professor of Industry AI, opens the Applied AI Stage by arguing that scaling AI in industry depends on ecosystems, governance and a clear strategy rather than on model quality alone. She points to Llama 4 Scout, which runs on a single H100 GPU with a context window of almost 10 million tokens, as a step toward industrial adoption, and urges every company to define an "AI north star" that turns hundreds of scattered use cases into a roadmap, using autonomous driving for carmakers and passenger experience for airports as examples. She stresses that companies need a full data inventory, an application stack they control themselves, and a central AI team responsible for strategy, the development stack, and scalable delivery. Her longtime collaborator Patrick Monk then presents agentic AI for enterprise workflows, demonstrating an email-triggered ordering agent and comparing model accuracy on business-process automation, from GPT-4o to DeepSeek R1 and OpenAI's o4-mini.
Key takeaways
- Feiyu Xu argues companies need an explicit AI north star to turn hundreds of scattered use cases into a prioritized roadmap.
- She says companies must keep control of the application layer themselves rather than handing it to hyperscalers.
- A central AI team's three jobs are strategy and investment, building the application stack, and delivering scalable applications with business units.
- Patrick Monk frames agent design as writing a standard operating procedure with defined escalation paths, accepting about 80% task coverage.
- Monk reports DeepSeek R1 reached about 87% task success at ten times lower cost than GPT-4o, pushing OpenAI to release o3-mini and o4-mini.
Detailed notes
Defining an AI north star
Feiyu Xu opens with recent generative AI releases, citing Llama 4 Scout as significant for industry because it runs on a single H100 GPU, including on-device, and offers a context window of nearly 10 million tokens, which she says allows combining long-term and short-term demand forecasting with real world data. She notes that many companies now have hundreds of AI use cases, in some cases more than 800, but boards are reluctant to invest large sums without direction, so she argues every company needs an "AI north star" to prioritize the top use cases, set roadmaps, milestones and investment. Working through examples with the audience, she proposes autonomous driving and embedded AI in manufacturing and business processes as the north star for carmakers, AI co-pilot and safe landing for Boeing, digital twins combined with AI for aircraft manufacturing, and passenger experience for airports such as Berlin.
Data, the application stack and Europe's infrastructure
She lists three data questions every company must answer: whether it has a full inventory of data it owns, produces and can access; the current state of its data ecosystem; and how many AI application stacks it actually needs, arguing for example that a company like Boeing needs at least two or three. She recalls spending months defining the legal framework for customer data access at SAP and tells the audience not to use GDPR as an excuse. On infrastructure, she argues hyperscalers such as Microsoft Azure cover infrastructure, data processing and foundation models, but companies must keep control of the application layer themselves, either directly or with partners. She names GCO as one of the few European infrastructure providers and mentions edge AI as necessary for disconnected assets such as offshore wind turbines, a project she is involved with at Siemens Energy.
Building a central AI team and governance
Responding to executives who call central AI teams useless, Xu argues a central AI team has three essential tasks: setting strategy and make-or-buy/partnership decisions, building the AI application stack and development lifecycle, and running a small product team with business units to make applications scalable. She adds that board and supervisory board members need AI governance covering six components of risk management, framing this as necessary for companies to build AI applications responsibly. She closes her segment with the prediction that AI will eventually move into the human body itself, for example as a sensor in artificial knees, hips or livers.
Agentic AI in enterprise workflows
Patrick Monk argues agents now work because models have gained common sense and reliable instruction following, crediting DeepSeek R1 for being cheap and strong enough to push OpenAI to release o3-mini and later o4-mini. He describes agents as suited to "swivel chair" tasks across multiple enterprise interfaces without good APIs, and recommends treating agent design like outsourcing a business process: write a standard operating procedure, define escalation paths and service-level agreements, and accept roughly 80% automation coverage. He cites an SAP research paper showing 99-100% precision but only about 80% recall on text-to-SQL tasks. He demonstrates an agent that reads an email ordering an iPad mini, checks budget and authority in an internal database, and completes the order in a ServiceNow portal using only the page's HTML DOM tree. On task success rates he cites roughly 67% for GPT-4o, 75% for Claude Sonnet, 87% for DeepSeek R1 at ten times lower cost, and about 98-99.3% for o4-mini with a standard operating procedure and no traditional programming.
Q&A highlights
Asked whether there is a contradiction between models trained on past data and their use for future problems, Monk says agentic systems mainly need past process inputs, outputs and descriptions layered on a baseline model, and that teams should measure performance before considering fine-tuning, comparing it to not tuning a car before ever driving it. Xu adds that open-weight models like Llama 4 make fine-tuning on past data accessible, and that combining a general-purpose multimodal model with real-time, in-context data lets companies draw on both past experience and current information.
Quotes
Without digitalization, without the data strategy, AI is only a good will, good intention.
Feiyu Xu
Don't use GDPR as your excuse.
Feiyu Xu
Chatbots are really kind of interns that are waiting for you to give them instructions, and agents give us the option to actually get work done.
Patrick Monk
Questions answered in this talk
What is an AI north star and why does a company need one?
According to Feiyu Xu at Rise of AI 2025, an AI north star is a defined strategic direction, such as autonomous driving for a carmaker, that lets a company prioritize its use cases, set roadmaps and milestones, and decide where to invest instead of pursuing hundreds of disconnected pilots.
What kind of enterprise tasks are AI agents best suited for today?
According to Patrick Monk at Rise of AI 2025, agents work best on 'swivel chair' tasks that span multiple enterprise interfaces lacking good APIs, treated like outsourced business processes with a standard operating procedure, defined escalation paths and an expectation of around 80% automation coverage.
What should a company's central AI team be responsible for?
According to Feiyu Xu at Rise of AI 2025, a central AI team should own strategy and partnership decisions, build the AI application stack and development lifecycle, and run a small product team with business units to make applications scalable.