No PriorsSAP: Bringing the ‘Operating System’ of a Company into the AI Era with CTO Philipp Herzig
CHAPTERS
- 0:00 – 0:42
Cold open: Enterprise AI as a 30% effort reduction—and the end of “human-driven UI”
Philipp opens with the practical promise of AI inside large enterprises: materially reducing effort and cost, especially in complex legacy environments. He also frames a major shift: software should no longer require the user to be “the intelligence in front of the screen.”
- •Large enterprises can cut substantial effort (e.g., ~30%) getting to outcomes
- •Legacy landscapes make automation and acceleration especially valuable
- •Traditional UX taught humans to click through rigid workflows
- •AI shifts software toward systems that do more of the thinking and execution
- 0:42 – 1:18
Philipp Herzig’s role and SAP’s AI-focused agenda (show setup)
Sarah introduces Philipp Herzig (SAP CTO) and previews the conversation: SAP’s AI strategy, enterprise-scale adoption challenges, predictive analytics, and broader tech/business model implications. This sets the lens on both technology transition and business outcomes.
- •SAP’s AI strategy and product embedding
- •Why SAP survives multiple tech cycles
- •Enterprise scale as the underappreciated challenge
- •Predictive/tabular ML as a complement to LLMs
- 1:18 – 2:51
What SAP does: The “operating system” for finance, HR, supply chain, and customer workflows
Philipp explains SAP’s breadth across core enterprise functions and why it’s best understood as an operating system for running a company end-to-end. He emphasizes cross-functional processes like order-to-cash and source-to-pay that SAP orchestrates globally.
- •~400k enterprise customers across industries
- •Coverage: finance, HR, supply chain, manufacturing, logistics, sales/service/procurement
- •SAP as an end-to-end business operating system, not just an app suite
- •Core value: standardized execution plus real-time insight
- 2:51 – 6:52
Why SAP endures: Standard software, scalable economics, and unwavering focus on outcomes
Sarah asks why SAP has remained durable through major platform shifts. Philipp traces SAP’s origins to the economics of standardization and argues that while technology changes, customer demand for measurable outcomes and ROI stays constant.
- •SAP’s founding insight: repeating bespoke implementations doesn’t scale
- •Standard software emerged from the need for scalable economics
- •Platform transitions: mainframe → client-server → internet → mobile → AI
- •Durability comes from delivering business outcomes, not chasing tech for its own sake
- 6:52 – 8:47
CTO priorities: Going “all in” on AI while re-engineering products for measurable customer impact
Philipp describes SAP’s internal and external AI push: widespread agentic coding and rapid productization like Joule for Consulting. He frames SAP’s work as re-engineering the system to make AI deliver outcomes across complex customer landscapes.
- •Company-wide adoption of agentic coding for developer productivity
- •Joule for Consulting reduces implementation/upgrade effort and cost (~30%)
- •AI agents already live in areas like travel/expense workflows (e.g., Concur)
- •CTO focus: impact across UI, processes, and data layers
- 8:47 – 12:14
Three-layer transformation: Generative UI, agent-driven processes, and a unified data foundation
Philipp compares the AI shift to the cloud transition: you can’t just “bolt it on”—you must re-architect. He outlines three major change layers: dynamic, multimodal UI; flexible agentic process execution; and a harmonized semantic data layer to power AI reliably.
- •Generative UI: interfaces generated dynamically for context and questions
- •Proactive, multimodal systems that surface issues and recommendations
- •Agentic processes blend structured + unstructured work beyond rigid SOPs
- •Data layer: harmonized semantic view combining SAP + external data sources
- 12:14 – 14:22
The hardest part: Making enterprise AI work at scale (documents, identity, APIs, integration)
Philipp argues the core challenge isn’t building a demo—it’s teaching AI to behave correctly at enterprise scale. He highlights escalation points: from 10 documents to thousands, from a handful of APIs to tens of thousands, and from generic answers to identity- and policy-specific responses.
- •POCs are easy; scaling to real enterprise complexity is hard
- •Personalization requires master data (location, payroll, tax rules, policies)
- •Context and orchestration complexity grows rapidly with scale
- •Tooling challenge: SAP has ~20,000 APIs—far beyond typical agent demos
- 14:22 – 17:48
Verifiability and evals: Why coding agents work—and what enterprise agents still need
Philipp explains why agentic coding has advanced quickly: outputs are verifiable (compiles, tests pass). For enterprise workflows, reliability requires explicit expected outcomes, robust boundary conditions (security/privacy), and a renewed emphasis on tests/evals—echoing test-driven development in a new form.
- •Code generation succeeds because verification is straightforward (tests/compilation)
- •Enterprise agents need clear input→output expectations for validation
- •Boundary conditions matter: security, privacy, maintainability beyond “vibe coding”
- •A shift back toward rigorous evals/testing as the core developer discipline
- 17:48 – 20:42
From system-of-record verifiability to “agent mining” and a data flywheel
Philipp describes two lanes: leverage the system of record for ground-truth verification, but recognize it’s insufficient for higher autonomy. He introduces “agent mining” (an evolution of process mining) to capture decision traces and tribal knowledge, turning human-in-the-loop interactions into new training/eval data and process improvements.
- •System-of-record data enables baseline verification of process outcomes
- •Tribal knowledge often lives outside systems (calls, Slack/Teams)
- •Agents ask clarifying questions; capturing answers creates reusable context
- •Agent mining: record decision traces, detect anomalies, elevate improvements
- •Creates a data flywheel: more traces → better evals → better agent outcomes
- 20:42 – 22:11
Tool calling vs. computer use: Why APIs will dominate (with legacy UI automation as a bridge)
Sarah asks whether enterprise agents will operate via UI “computer use” or via APIs/tool calling. Philipp expects tool calling to be the primary mode for reliability and structure, with computer-use approaches filling gaps where APIs are missing or legacy systems persist.
- •UI automation is improving but can be slow and brittle
- •Tool calling/background agents are preferred for structured integrations
- •Headless browsing can help but isn’t the default enterprise pattern
- •Computer use remains valuable for legacy systems and missing APIs
- 22:11 – 24:58
Where agents deliver value first: Unstructured work, knowledge workflows, and natural language analytics
Philipp notes LLMs excel in unstructured domains, making early ROI clearest in support, services, sales, and document-heavy knowledge work. He also highlights progress in bridging structured enterprise data with natural language—enabling conversational analytics and ad hoc dashboarding without heavy IT cycles.
- •Fastest returns: unstructured text/document-heavy workflows
- •Early domains: services, support, sales, consulting enablement
- •Tool orchestration requires disambiguation (e.g., multiple meanings of “order”)
- •Conversational BI: NL-to-SQL, iterate to insight, then “pin” views temporarily
- •SAP Knowledge Graph as glue between language and structured enterprise data
- 24:58 – 29:05
Why LLMs aren’t enough for predictive/tabular analytics—and SAP’s RPT-1 approach
Philipp argues planning and decision-making require predictions (demand, cash flow, payment risk), which are poorly served by pure next-token LLM architectures. He describes the scaling limitations of traditional ML (many models per country/business unit) and introduces SAP’s research on Relational Pre-trained Transformers (RPT-1) to bring foundation-model-like leverage to structured/tabular prediction tasks.
- •Core enterprise needs: forecasting and prediction (classification/regression/time series)
- •LLMs excel at language but are not designed for tabular prediction accuracy
- •Traditional ML doesn’t democratize well; requires scarce data science talent
- •Operational scale problem: many geographies → many models (e.g., 180+)
- •RPT-1: transformer-based approach tailored to relational/structured data
- 29:05 – 31:54
Barriers to enterprise AI adoption: Data fragmentation, scale, and security readiness
Philipp frames adoption gaps as practical constraints: disaggregated data across systems (including M&A sprawl), the complexity of integrating experiences at scale, and enterprise-grade security requirements. He emphasizes that innovations from the open ecosystem often need significant hardening before deployment in regulated environments.
- •Data is fragmented due to buying decisions, silos, and M&A landscapes
- •AI value is limited without integrated, high-quality, harmonized data
- •Unified experiences are harder as landscapes grow more complex
- •Security is a gating factor; open-source innovations often need hardening
- •Risk examples: credential/key exposure vulnerabilities are unacceptable in enterprise
- 31:54 – 34:03
How AI ‘uplevels’ enterprise roles: Less mundane work, faster decisions, more scenario planning
Philipp predicts AI will remove repetitive information-gathering and presentation work, accelerating decision cycles and improving quality. He compares enterprise operations roles to junior developers: people move up a level into supervision, feedback, and higher-leverage strategic work while agents handle routine execution.
- •Automation shifts effort away from manual collection and slide-building
- •Teams run more scenarios with deeper insight and faster iteration
- •Work becomes more strategic as agents absorb repetitive execution
- •Analogy: finance shared services roles evolve like devs using coding agents
- •Humans focus on oversight, guidance, and defining “what to build/decide next”
- 34:03 – 36:35
AI forces pricing evolution: From seats to consumption—and eventually outcome-based models
Sarah asks how increased automation and “service as software” affects SAP’s business model. Philipp explains the shift from primarily seat-based licensing toward consumption-based pricing, with outcome-based models as a longer-term possibility—while acknowledging customers still demand predictability and cost control, leading to hybrid approaches today.
- •SAP historically seat-based, with some consumption exceptions
- •AI pushes pricing toward consumption; outcome-based is a future step
- •Customers want predictability and may not yet fully trust outcomes
- •Risk of cost “explosion” is a customer concern in pure consumption
- •SAP’s approach: hybrid model that meets customers where they are
- 36:35 – 38:53
What makes an AI-era winner: Make technology disappear and deliver outcomes fast
Philipp argues the differentiator won’t be the novelty of technology but the speed and reliability of customer value realization. He emphasizes architectural flexibility, partnering where tech commoditizes, and focusing investment on differentiators—while ensuring enterprise-grade qualities and rapid “time to value.”
- •Winning equals adoption + measurable customer outcomes
- •Technology should be invisible; outcomes must be front-and-center
- •Flexible architecture and partnerships reduce over-indexing on any one model/vendor
- •Invest in differentiators; expect commoditization in parts of the stack
- •Fast activation and short time-to-value preserve ROI and strengthen business cases
- 38:53 – 39:48
Day in the life of SAP’s CTO: Progress reviews, prototyping, and external signal scanning
Philipp describes his daily work as a mix of guiding teams across the stack (database to models to UI), staying current on external developments, and hands-on prototyping to learn what works. He then feeds those learnings back into product direction and execution.
- •Review progress and provide guidance across multiple layers of the stack
- •Continuously study external ecosystem changes
- •Prototype frequently to validate ideas and constraints
- •Translate hands-on findings into direction and inspiration for teams
