At the Scheer IDS „ALL IN AI“ event (Scheer Austria GmbH, Vienna), the message was unmistakable: SAP is no longer asking whether AI will reshape enterprise software — it is publishing the blueprint — and it is subsidizing the transition. SAP co-funds the adoption of its new Business AI technology stack — experience with comparable packages shows funding of up to 50% (up to €100k) — and Scheer IDS is a certified SAP implementation partner who evaluates, informs, and implements.
The combination of SAP's „AI-Native North Star Architecture“ (May 2026) and Scheer IDS' real-world PoCs (S/4HANA production assistant, ABAP modernization with MCP servers) confirms the architectural direction we chose for our own Transfer Pricing platform. And it maps, with unusual precision, where agentic document generation is heading next.
Source: Scheer IDS — ALL IN AI: Verantwortungsvoll beschleunigen (presentation, 2026) · Scheer IDS Austria: https://www.scheer-ids.com/de/unternehmen/scheer-austria · SAP AI-Native North Star Architecture (May 2026, Public)
SAP's central claim is deceptively simple: "Data was the moat of the last decade. Context is the moat of the next." Frontier models are commoditizing; every vendor can access the same leading LLMs. What cannot be copied is the context those models reason over — the decision history, the business semantics, the governed data that grounds every output.
This is not abstract theory. It is the exact problem Transfer Pricing documentation has struggled with since BEPS Action 13: how to turn five years of financial data, intercompany agreements, and functional analyses into a defensible, auditable narrative — without losing the "why" behind the numbers.
From AI-First to AI-Native: The Four Layers
SAP describes four architectural layers, each shifting from a rigid past to an adaptive future:
- User Experience Layer — from static apps to intent-driven engagement. Joule becomes the central entry point; users stop navigating and start stating intent.
- Process Layer — from predefined workflows to agentic orchestration. Applications become capability providers that agents discover and invoke through a governed gateway.
- Foundation Layer — from system of record to system of context. SAP Business Data Cloud and SAP Knowledge Graph provide semantically grounded data; context engineering assembles the right slice of enterprise information for every interaction.
- Platform Layer — from hosting applications to running governed enterprise agents, with identity, sandboxing, observability, and tenant isolation built in from day one.
Two open protocols do the heavy lifting: MCP (Model Context Protocol) for tool integration — agents discover and invoke capabilities across systems — and A2A (Agent2Agent) for multi-agent coordination. Security treats agents as first-class principals with scoped authorizations; a three-tier AI defense (zero-trust foundation, security agents, deterministic supervisory agents) wraps it all.
Why Scheer IDS Matters Here
SAP's architecture paper is a vision. What makes it credible is seeing the same principles running in production — which is exactly what Scheer IDS demonstrated at their ALL IN AI event:
- Production Intelligence Cockpit: A German-language AI assistant for shift supervisors on S/4HANA on-premise — no RAG, no ETL, no vector database. The agent calls published OData APIs, SAP's own authorization model enforces a 403 without permission, and the LLM never touches the backend: it is a reasoning layer, not a backend caller. This is SAP API Policy v4 §2.2.2 in practice — agentic access runs through the governed Agent Gateway.
- ABAP modernization with an MCP server: Three layers — a read-only Fiori analysis app, an Eclipse extension for developers, and a policy-driven MCP server as the controlled interface for any system change. "Co-pilot, not autopilot": every write requires release, everything is auditable.
- SAP Company Memory (GA Q2/2026): a system-agnostic (markdown-based) single source of truth that seeds and governs business context across all agents — the operationalized version of the system of context.
- SAP × Anthropic: Claude models are available through the SAP Generative AI Hub, and Claude Code lands in Joule Studio 2.0 — but always behind the Agent Gateway, never as direct model access to SAP systems.
For a document automation platform, this is the strongest possible external validation: the same governance-by-design pattern we built (JWT trust path, tenant isolation, MCP tools, auditable prompt chain) is now SAP's endorsed architecture — enforced by API policy, not just recommended.
SAP Funds the New Stack: Up to 50%, up to €100k
Here is the part that changes the business case: SAP is subsidizing the adoption of exactly this stack. Through its Partner-Led Adoption Packages, SAP co-funds the jump into the Business AI Platform — and Scheer IDS, as certified implementation partner, evaluates, informs, and implements the funded projects.
| Package | SAP funding | Scope |
|---|---|---|
| Launch | €15k | AI Assistant/Agent activation, Joule Studio entry, guided discovery & enablement |
| Low-Touch | €25k | First custom agent or pro-code app, pre-production MVP on the SAP Business AI Platform, landscape architecture & roadmap |
| Mid-Touch | €50k | Custom agent and pro-code app, BAIP landscape architecture, MVP roadmap, productive rollout with Agent Hub |
| Enterprise | €100k | Joule Studio customer agent, pro-code app, Joule Skills, SAP Agent Hub, unified metering, customer BAIP |
Experience with comparable packages shows funding of up to 50% — for the Enterprise package, that means SAP co-funds up to €100k. The economics of an AI-native rollout change fundamentally: the first agentic solution is not an experiment you pay for yourself — it is a co-funded, governed entry into the platform that SAP itself has declared its architectural future.
The delivery approach is standardized across all packages: discovery → statement of work → resource schedule → development → UAT/validation → enablement → support. And every package runs through certified SAP partners — in Austria, Scheer IDS (Scheer Austria GmbH, Vienna) is the partner that evaluates eligibility and informs on the funded path.
Practical note: funding is tied to SAP-endorsed architecture — agentic access runs through the Agent Gateway, models through the Generative AI Hub. The subsidized path is also the compliant path.
What This Validates in Our Platform
We built our Transfer Pricing Master File platform (tp.finestselection.net) around exactly these principles — before the SAP paper made them fashionable:
| SAP North Star concept | Our implementation |
|---|---|
| MCP as tool integration | A dedicated MCP server exposing five JWT-protected tools: section context hydration, system-prompt resolution, semantic knowledge search, revision persistence, and usage logging |
| Agent orchestration | An ADK agent (FastAPI) that verifies identity, calls MCP tools, and streams LLM output — with model/provider switching via environment config, no redeploy |
| Governed gateway / trust path | A JWT bridge: Clerk session → tenant ownership check → short-lived signed token → agent. Tenant isolation enforced at every hop |
| Semantic grounding (RAG) | Uploaded PDF/DOCX/TXT knowledge base, chunked and embedded, ranked by cosine distance over pgvector |
| Data products / master data | Projects modeled per client with target fiscal year and lookback period; Master File → Chapters → Sections with full revision audit trail |
| Business context / prompt governance | A fallback prompt chain (project override → agent config → SOUL.md → hardcoded) with explicit source labels for auditing |
| Decision traces | Every agent persistence writes a revision — an audit trail of what was generated, when, and on which context |
The architecture is already AI-native in SAP's sense: the model reasons, the harness governs.
Where SAP Points Us Next
The North Star paper is also a roadmap for what we have not built yet:
1. Decision traces as first-class assets. SAP's "context graph" captures what was known, considered, chosen, and what resulted — each human correction becomes a structured decision trace. For TP documentation, this is gold: an audit trail that explains why a functional analysis took a certain position, not just that it was saved. Our revision system is the foundation; exporting traces as structured, queryable evidence is the next step.
2. Local File as a sibling workflow. Master Files and Local Files share context but demand different evidence bases. The MCP tool model makes this trivial to extend — a get_local_file_context tool, its own knowledge base, its own prompt chain — without touching the agent runtime.
3. Multi-agent orchestration. SAP's A2A pattern suggests separating a research agent (knowledge retrieval, precedent search) from a drafting agent (section generation) — each with its own identity, scoped permissions, and evaluation. Single-agent systems hit quality ceilings; governed multi-agent systems compound.
4. Agent evaluation as a feedback loop. SAP warns (Goodhart's Law) against optimizing a single metric. The evaluation target for TP document generation should be multi-objective: technical completeness, review-through rate, revision frequency, client acceptance. Those measurements become the learning signal.
5. Fiscal-year-native data modeling. SAP's data-product thinking applies directly: financial data is not a snapshot but a time series. Our platform already models projects by target year with lookback periods; the next iteration treats fiscal years as a first-class data dimension — so agents reason over five years of trend, not one year of numbers.
The Takeaway
SAP's North Star is not a product roadmap; it is a validation of an architectural direction — and a funded one at that. With SAP co-funding up to 50% (experience with comparable packages; up to €100k) through the Partner-Led Adoption Packages and Scheer IDS as certified implementation partner (evaluating, informing, and implementing in Austria), the paper confirms what we have seen in practice: context engineering, governed agent runtimes, and MCP-based integration are the durable foundation for document automation — and for Transfer Pricing documentation specifically, they convert the "system of record" (numbers) into a "system of context" (numbers plus the reasoning behind them).
That is the difference between documentation that survives an audit and documentation that merely exists.
Finest Selection GmbH — AI-driven Transfer Pricing documentation. Master File and Local File automation, governed by design, grounded in your data.