No PriorsSAP: Bringing the ‘Operating System’ of a Company into the AI Era with CTO Philipp Herzig
CHAPTERS
- 0:00 – 0:42
Cold open: AI as the new interface and 30% efficiency gains
A quick teaser frames SAP’s AI thesis: companies can materially reduce effort and cost by embedding AI into enterprise workflows. Philipp argues the era of software that requires “intelligence in front of the screen” is ending as UIs and processes become more autonomous.
- •AI can reduce enterprise project effort by ~30% in complex environments
- •Cost and time savings come from faster path to outcomes
- •Shift from human-driven clicking through UIs to AI-driven task completion
- •Enterprise complexity (heritage systems, varied policies) shapes what’s possible
- 0:42 – 1:18
Meet SAP’s CTO Philipp Herzig and the episode roadmap
Sarah introduces Philipp and sets expectations: SAP’s AI strategy, adoption challenges at scale, predictive analytics beyond LLMs, and how business models shift with agentic software. The framing emphasizes that AI is not only a technology transition but a business model transition.
- •CTO Philipp Herzig’s scope includes AI strategy and product direction
- •Focus on enterprise-scale adoption, not just impressive demos
- •AI changes how software is built, used, and monetized
- •Predictive analytics and real-world optimization problems will matter
- 1:18 – 6:53
SAP as the ‘operating system’ for enterprises
Philipp explains SAP’s breadth across core business functions—finance, HR, supply chain, manufacturing, procurement, and customer-facing operations. He positions SAP as an end-to-end system of record enabling “order-to-cash” and “source-to-pay” at global scale.
- •~400,000 enterprise customers across industries and geographies
- •Runs core processes: finance, HR, logistics, warehouse, procurement, sales/service
- •“Operating system of a company” metaphor: system of record + execution backbone
- •End-to-end process coverage is a durable advantage
- 6:53 – 12:14
CTO priorities: going ‘all-in’ on AI across product and engineering
Philipp describes SAP’s internal adoption (agentic coding) and customer-facing AI products focused on measurable outcomes. He highlights Joule for Consulting as a high-ROI example that accelerates migrations and modernization in complex SAP landscapes.
- •Company-wide use of agentic coding for developer productivity
- •Outcome-led AI products (e.g., Joule for Consulting) drive adoption
- •30% effort reduction cited for consulting-heavy transformation work
- •Agents already live in specific products (e.g., travel/expense workflows)
- 12:14 – 17:06
The real engineering challenge: scaling AI from demos to enterprise reality
Philipp argues the hardest part isn’t building an LLM demo—it’s getting correct, context-aware behavior at massive scale. As document counts, policies, and integrations grow, systems must personalize responses, manage context limits, and orchestrate thousands of tools and APIs reliably.
- •POCs work on 10 docs; enterprise requires 1,000+ and complex access rules
- •Context depends on master data (location, payroll, tax rules, entitlements)
- •Tool scale becomes extreme: SAP cites ~20,000 APIs
- •Unified cross-domain experience (finance + HR + supply chain) raises complexity
- 17:06 – 20:42
Verifiability, evals, and ‘agent mining’ as a data flywheel
To make agents trustworthy, SAP leans on verification and evaluation strategies similar to what made coding agents effective: you can test outcomes. Philipp describes capturing decision traces—“agent mining”—to turn tribal knowledge into data, detect anomalies, and continuously improve SOPs and agent behavior.
- •Coding agents succeed partly because outputs are verifiable (tests/compilation)
- •Enterprise agents need explicit evals and boundary conditions (security/privacy)
- •System-of-record data provides initial ground truth, but is not sufficient
- •Agent mining captures decision traces to refine SOPs and build a learning loop
- 20:42 – 22:11
Tool calling vs. computer-use agents: what belongs where
Philipp expects most enterprise automation to rely on tool calling via structured integrations rather than UI-driving ‘computer use,’ due to reliability and operational preferences. However, computer-use approaches remain valuable when APIs are missing or legacy systems persist.
- •Tool calling is likely the default for scalable, headless background execution
- •UI automation is improving but can be slow and brittle
- •Legacy and gaps in API coverage keep computer-use relevant
- •Pragmatic hybrid approach: structured integrations first, UI fallback when needed
- 22:11 – 24:58
Where agents deliver value first: unstructured work and data-to-insight workflows
Early wins cluster around unstructured domains—service/support, sales, consulting, document-heavy knowledge work—where LLMs excel. As orchestration improves, SAP also targets ‘chat with your data’ experiences that generate analyses on demand and can be pinned into lightweight, evolving dashboards.
- •Unstructured workloads (documents, tickets, knowledge bases) show fastest ROI
- •Orchestration challenges include ambiguity (e.g., ‘order’ has multiple meanings)
- •Natural-language analytics reduces dependency on analysts for ad hoc questions
- •Knowledge graph bridges language and structured enterprise data
- 24:58 – 29:07
Why LLMs aren’t enough for predictive/tabular analytics (and SAP’s RPT approach)
Philipp argues enterprise decision-making needs robust prediction (demand, cash flow, payment behavior), which classic LLM token-generation is not optimized for. He describes SAP’s research on relational/tabular transformers (RPT) to reduce reliance on bespoke model-building and scale predictive ML across many countries and contexts.
- •Planning requires prediction: classification/regression/time series use cases
- •LLMs excel at unstructured tasks but struggle as primary predictive engines
- •Traditional ML (e.g., XGBoost/AutoML) works but doesn’t scale operationally
- •RPT (Relational Pre-trained Transformers) targets scalable predictions on relational data
- 29:07
Enterprise adoption barriers and the shift to consumption/outcome-based pricing
Adoption is constrained by fragmented data landscapes, integration difficulty at scale, and enterprise-grade security requirements. On monetization, Philipp expects a gradual move from seat-based licensing to hybrid and more consumptive models, with outcome-based pricing becoming more feasible as trust and verifiability improve.
- •Data fragmentation (systems sprawl, M&A) limits AI impact
- •Security hardening is mandatory; ‘GitHub-ready’ isn’t enterprise-ready
- •Customers want cost predictability and still calibrate trust in AI outputs
- •Pricing evolves: seat-based → hybrid → more consumptive, eventually outcome-based
Why SAP has endured through multiple technology transitions
SAP’s durability is framed as an outcome-driven story: customers consistently want reliable business outcomes, even as underlying tech shifts (mainframe → client-server → internet → mobile → AI). Standard software scaled better than repeated bespoke implementations, and that economic logic still holds.
- •Standardization beat re-implementing similar systems customer-by-customer
- •Tech stacks evolve, but customer demand for outcomes persists
- •AI is the next transition—requiring re-engineering, not surface-level add-ons
- •Enterprise adoption lags innovation; closing that gap becomes strategic
Three-layer re-engineering: UI, processes, and data foundation
SAP’s AI transformation is presented as a parallel to the cloud shift: you don’t just “host” existing software differently—you redesign how it works. Philipp breaks the change into three levels: generative UI, agent-augmented business processes, and a harmonized data layer to power reliable AI.
- •Generative UI replaces static screens with dynamic, task-driven experiences
- •Agents reshape rigid workflows into adaptive ‘outcome as a service’ execution
- •Data layer modernization enables combining SAP and non-SAP data safely
- •Harmonized semantics are critical: ‘AI is only as powerful as the data’
