Apr 23, 2026 · 39m · no-priors
SAP: Bringing the ‘Operating System’ of a Company into the AI Era with CTO Philipp Herzig
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
SAP CTO Philipp Herzig joins Sarah Guo on the No Priors podcast to discuss how the enterprise software giant is re-engineering its platforms for the generative AI era. He outlines technical strategies spanning tabular foundation models, agentic workflows, scalable evaluation frameworks, and the commercial transition toward outcome-based software.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 17.7% of the talking time here. How this is scored →
The hosts as informed peer 5.2 Guest teaching 4.3 Guest disagreement 0.3 The hosts pushing back 0.7
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:40 Rejecting LLMs as a solution for tabular forecasting
Philipp firmly rejects the conventional industry assumption that LLMs can handle enterprise predictive analytics, arguing that sequence-to-sequence token prediction is fundamentally unsuited for tabular regression and classification.
Hardest push from the hosts ▶ 2:51 Challenging SAP's durability against startup disruptionSarah presses Philipp on how SAP has managed to survive successive platform shifts when standard venture capital logic predicts incumbents will be unseated by emerging startups.
Biggest teaching moment ▶ 26:15 Explaining the massive scaling limits of classical tabular MLPhilipp educates Sarah on the operational impossibility of traditional AutoML at multinational scale, explaining that a single payment delay problem across 90 countries demands managing 180 bespoke models without a foundational relational transformer.
The host holds their own ▶ 17:06 Drilling into enterprise agent verifiability bottlenecksSarah leverages her deep technical understanding of AI evaluation to challenge Philipp on whether enterprise business agents can realistically achieve the compounding improvements seen in code generation without formal verifiability.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Defining SAP's Breadth as the Enterprise Operating System | 6 | 4 | 1 | 2 | Sarah articulates the venture capital thesis regarding technological disruption cycles and notes SAP's enterprise market cap dominance over newer players like Salesforce. Philipp contextualizes SAP's durability by recounting its founding in 1972 around standard software economics and its focus on enterprise business outcomes. | |
| Three Pillars of SAP's Technical Re-engineering Strategy | 4 | 5 | 0 | 0 | Sarah prompts Philipp on his architectural priorities as CTO. Philipp delivers an extensive breakdown comparing the cloud transition to AI across three architectural pillars: generative UI, agentic business processes, and harmonized data layers. | |
| Addressing the Engineering Bottlenecks of Enterprise Scale and Evals | 6 | 5 | 0 | 1 | Philipp explains enterprise scaling hurdles, contrasting simple 10-API MCP prototypes with SAP's 20,000 APIs and complex localized master data. Sarah readily connects the discussion to earlier dialogues on writing evals and the revival of test-driven development. | |
| Verifiability, Tribal Knowledge, and Agent Mining Flywheels | 7 | 4 | 1 | 2 | Sarah probes deeply into whether enterprise business logic can compound like code generation given the lack of clear verifiability. Philipp agrees and details the distinction between deterministic system-of-record checks and capturing unstructured tribal knowledge via agent mining. | |
| Comparing API Tool Calling and Computer Use in Enterprise Workflows | 5 | 4 | 0 | 1 | Sarah asks Philipp to compare computer use against API tool calling for enterprise automation. Philipp explains why API tool calling remains the dominant path while computer use serves legacy edge cases, detailing orchestration challenges with ambiguous enterprise terms. | |
| Limitations of LLMs and Relational Pre-trained Transformers (RPT-ONE) | 6 | 6 | 1 | 1 | Sarah highlights Philipp's non-consensus bet on tabular foundation models over pure LLMs. Philipp details the mathematical shortcomings of sequence-to-sequence language models for regression and classification tasks, presenting SAP's NeurIPS-published RPT-ONE architecture. | |
| Enterprise Adoption Bottlenecks: Data Fragmentation and Security | 5 | 4 | 0 | 0 | Sarah inquires about enterprise deployment roadblocks in the outcome race. Philipp identifies data fragmentation from historical M&A and points out recent open-source security vulnerabilities like LightLLM credential leaks that scare CISOs. | |
| The Future Evolution of Enterprise Roles in Finance and HR | 4 | 4 | 0 | 0 | Sarah asks how operational roles in finance and HR will change over the next few years. Philipp draws an analogy to junior software engineers using coding assistants, predicting routine data preparation will be handled by agents while humans supervise outcomes. | |
| Shifting Enterprise Software Pricing from Seats to Consumption | 5 | 5 | 0 | 0 | Sarah openly asks about SAP's current pricing structure and how agentic workflows alter software monetization. Philipp confirms the ongoing migration from seat-based licensing to consumption and outcome models while highlighting customer demand for enterprise cost predictability. | |
| Differentiating Winners from Losers in the Enterprise AI Era | 6 | 4 | 0 | 1 | Sarah pushes back against the broad 'SaaS is dead' panic, asking what specifically separates enterprise software winners from losers. Philipp asserts that winning vendors must make the underlying technology disappear and focus strictly on customer business outcomes rather than commoditized tech layers. | |
| A Day in the Life of SAP's CTO and Real-Time Prototyping | 3 | 2 | 0 | 0 | Sarah asks a closing rapid-fire question about Philipp's day-to-day routine as CTO. Philipp shares that he actively runs terminal prototypes and tests developer tools even during interviews. |
Statements from this episode (13)
Herzig: SAP powers core operations for 400,000 enterprise customers
“SAP is the market leader, right, in enterprise of software applications and platforms, right? It's a 400,000 enterprise customers, and usually I just running their finance and HR and, you know, supply chain, manufacturing, execution, logistics, warehouse manag…”
Herzig: Enterprise AI adoption gap is widening rather than narrowing
“We also see that the AI adoption in the enterprise is still not where we want to see it, right? Like there's this Gartner curve, right? Where say like there's this AI innovation race, and then there's this AI outcome race, right? Then the gap almost increases,…”
Herzig: The era of software requiring manual user intelligence is over
“The time is clearly over where you design software, where the dump software, where the, that requires the intelligence to sit in front of the computer.”
Herzig: Agentic coding succeeds because outcomes are automatically verifiable
“The problem of why does agent decoding work so well, Sarah, is of course you can verify the outcome, right? You can either say, hey, is the program compiling or are your unit tests, right? Does it work, et cetera.”
Herzig: Developers never actually adopted test-driven development in practice
“The reality is nobody did it. At least I never did it because it was so much more fun. It was not very popular at the end.”
Herzig: System of Record Data Is Insufficient for Autonomous Enterprise Agents
“If you just look into the system of record today, that data is insufficient for this grand vision that everybody has, that it becomes this autonomous enterprise or like the agency of these agents is increasing right over time.”
Herzig: Agent Decision Traces Create a Data Flywheel for Enterprise AI Evals
“And that then leads to this kind of, I call this then this data flywheel, so to speak. So because with every trace, every input a user gives you with all the observability that an agent writes you have new data sources that can then lead to new evals where som…”
Herzig: Most enterprise agents will rely on tool calling, not GUI
“I still believe for the most part it will, The majority will live with tool calling, right? And agents running in the background and so on, right? Because you also don't, you know, maybe want to have the browser open all the time. Okay, we can do this with hea…”
Herzig: SAP built a knowledge graph to bridge natural language and structured data
“We are seeing through, for example, the knowledge graph, the SAP knowledge graph that we've built, which is kind of the glue between natural language and the structured data in the system to really bring this together.”
Herzig: Large language models are not designed for predictive enterprise tasks
“Now, if you want to do these predictions, quite frankly, then the challenge is large language models are not made for this, right? The way how they, you know, generate just one token after another, essentially, in a sequence-to-sequence modeling. I mean, They'…”
Herzig: SAP developed RPT-ONE tabular foundation models after two years of research
“And that led, actually, this was two years of research. We published it also at NeurIPS and a bunch of other conferences. We call this RPT-ONE, so RAPID-ONE stands for relational pre-trained transformers. It's still based on the transformer architecture, but w…”
Herzig: SAP software is mostly licensed per-seat today with few exceptions
“For the most part, SAP software is seat based, licensed today with a few exceptions like a conquer or a field glass, for example, or the business network.”
Herzig: AI will transition software pricing to consumption, then outcome-based models
“Very clearly with AI, it was very clear for us that, you know, step by step, it will go towards this consumptive world, right? First consumptive, and then maybe in the next step, once we have more verifiability in the system, then also towards maybe an outcome…”