📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=5u7AjPardvo
Beyond the Dashboard: SAP’s CTO on the Death of the UI and the Rise of AI Agents
SAP has outlasted every major technology cycle since the 1970s, successfully navigating transitions from mainframes to the cloud and mobile. Now, CTO Philip Herzig explains how the enterprise giant is re-engineering its core to move beyond simple software and toward a future of autonomous business outcomes.
Core Question: How does a legacy enterprise titan reinvent itself for the AI era while managing the impossible scale of global business data?
Highlights
- The era of teaching humans to click through UIs is over; “Generative UI” will replace rigid dashboards with dynamic, task-specific views.
- Scaling AI from a simple ten-document proof-of-concept to 20,000 enterprise APIs is the defining engineering hurdle of the next decade.
- LLMs are insufficient for high-stakes finance; SAP is developing Relational Pre-trained Transformers (RPT1) specifically for structured tabular data.
- Business models are shifting away from traditional seat-based licenses toward hybrid consumption and outcome-based pricing models.
⏱️ Reading time: approx. 6 minutes · Saves you about 34 minutes vs. watching.
Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇
The End of the “Dumb” Interface
From Clicks to Conversational Outcomes
For decades, enterprise software design focused on teaching humans how to navigate complex menus and click specific buttons to complete a task. Philip Herzig argues that this “dumb software” paradigm is finally dead because AI can now dynamically generate the interface a user needs in real-time.
Instead of hunting through pre-built dashboards, users simply state a business problem, and the system builds the analytical view or executes the workflow on the fly. This shift represents a move toward multimodal and proactive systems that monitor global supply chain risks—like disruptions in the Strait of Hormuz—while the user sleeps.
By morning, the agent doesn’t just present a report; it offers specific recommendations for re-routing shipments based on current inventory data and regional tariff regulations, effectively blending the structured world of databases with the unstructured reality of global news.
💡 Digging Deeper
Q: What is Generative UI?
A: It is a user interface that doesn’t exist until a user asks a question, at which point the system builds a custom view to solve that specific problem.
Q: How does this change the role of a business user?
A: Users stop being “operators” of software and start being “supervisors” who review AI-generated recommendations and strategic options.
The Engineering Reality of Enterprise Scale
Solving the “Context Bloat” Problem
While building a chatbot on ten documents is a trivial weekend project, applying AI across 400,000 global enterprise customers is a massive technical challenge. Herzig points out that when you move from ten APIs to 20,000, standard “off-the-shelf” AI solutions suffer from massive context bloat and hallucination.
At SAP, the engineering focus isn’t just on the models themselves but on the “tribal knowledge” that lives outside the system of record. Success requires connecting AI to master data—like knowing exactly which payroll taxes apply to a German employee versus a US employee—to ensure every agent-driven action is legally and operationally compliant.
Reliability in the enterprise depends on “Agent Mining,” where the system records every decision trace to create a data flywheel for future improvement.
💡 Digging Deeper
Q: Why is enterprise AI harder than consumer AI?
A: Because of the scale of data and the need for 100% accuracy in regulated fields like finance and HR where “hallucinations” are not an option.
Q: What is “Agent Mining”?
A: It is the process of recording how AI agents and humans interact to find “anomalies” or “improvements” in standard operating procedures.
Why LLMs Fail at Predictive Finance
The Rise of Relational Pre-trained Transformers
Large Language Models are masters of the unstructured world, but they are fundamentally ill-equipped to handle the tabular data found in general ledgers. To solve this, SAP spent two years researching “RPT1,” a transformer architecture designed specifically for relational databases rather than sequences of words.
Predicting cash flow or demand isn’t just about text; it involves complex regression and classification across hundreds of thousands of database tables. By using these specialized models, companies can achieve high-accuracy predictions with smaller datasets, democratizing data science for teams that can’t afford a massive department of PhDs.
💡 Digging Deeper
Q: Can’t you just use an LLM with a calculator tool?
A: For simple math, yes, but for deep pattern recognition in millions of rows of financial data, specialized tabular models are far more efficient.
Q: What is the main business benefit of RPT1?
A: It allows businesses to make high-accuracy predictions, like demand forecasting, without needing to hire a massive team of data scientists.
The Future of the Enterprise Business Model
Transitioning from Seats to Outcomes
As AI agents begin to perform the “mundane work” traditionally handled by humans, the traditional seat-based licensing model is becoming obsolete. SAP is moving toward a hybrid model that prioritizes consumption and, eventually, verifiable business outcomes.
This shift mirrors the transition from on-premise to cloud software, requiring companies to rethink how they measure return on investment. If an AI reduces consulting effort by 30%, the value is no longer in the time spent but in the speed and accuracy of the result achieved.
Ultimately, Herzig believes the winners in the AI race will be those who make the technology “disappear” by delivering seamless business outcomes.
💡 Digging Deeper
Q: How does SAP price AI today?
A: It is currently a hybrid of seat-based and consumption-based models to give customers predictability while they transition.
Q: Will AI replace finance and HR teams?
A: No, it up-levels them; junior employees become supervisors, spending less time on PowerPoints and more time on strategic decision-making.
Key Takeaways
The transition to AI in the enterprise is less about the models themselves and more about the integration of those models into the “heritage” and complexity of global business. SAP’s strategy hinges on making AI invisible—embedding it so deeply into the business process that the technology fades into the background while the outcomes come to the fore.
We are entering an era of “Service as Software,” where the distinction between a software tool and a service provider blurs. For organizations to thrive, they must move past the hype of “innovation races” and focus on the “outcome race,” ensuring their data is harmonized and their AI is verifiable at a global scale.
Q&A
Q1: How has SAP stayed relevant for over 50 years?
A: By focusing on “Standard Software” economics—building solutions that scale across many customers rather than custom-coding for every individual client.
Q2: What is the biggest mistake entrepreneurs make with enterprise AI?
A: Underestimating the challenge of scale; what works for a small demo often breaks when confronted with 20,000 APIs and complex regional master data.
Q3: Is “test-driven development” coming back?
A: Yes, because while AI can write code, humans must define the “evals” (evaluation metrics) and boundary conditions to ensure that code is safe and reliable.
Q4: What role does “tribal knowledge” play in SAP’s AI?
A: It is the context stored in Slack, emails, or people’s heads. SAP aims to capture this through agent interactions to improve future automated decisions.
Q5: Why did SAP build its own “Knowledge Graph”?
A: To act as the “glue” between natural language queries from users and the highly structured data sitting in financial and supply chain tables.
Q6: How does Philip Herzig spend his day as CTO?
A: He balances high-level strategy and team reviews with “hands-on” prototyping, often running command-line interfaces to test new AI models personally.
Q7: Will AI make software cheaper?
A: It reduces the cost of reaching an outcome, but the complexity of maintaining reliable, secure, and integrated AI at scale remains a significant investment.
