With its "AI-Native North Star Architecture" (May 2026, public), SAP has published a blueprint for how AI reshapes enterprise software. The message: SAP is no longer asking whether agentic AI will change enterprise software — it describes how. This architectural direction confirms the path we chose for our own Transfer Pricing platform, and it maps, with unusual precision, where agentic document generation is heading next.
Source: SAP AI-Native North Star Architecture (May 2026, public).
SAP's core claim is deceptively simple: "Data was the moat of the last decade. Context is the moat of the next." Frontier models are becoming a commodity; every vendor can reach the same leading LLMs. What cannot be copied is the context those models reason over — the decision history, the business semantics, the governed data every output is grounded in.
This is not abstract theory. It is precisely the problem transfer pricing documentation has wrestled 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 moving from a rigid past to an adaptive future:
- User-experience layer — from static apps to intent-driven interaction. Joule becomes the central entry point; users no longer navigate, they state their 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 each interaction.
- Platform layer — from hosting applications to running governed enterprise agents, with identity, sandboxing, observability and tenant isolation 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 permissions; a three-tier AI defense (zero-trust foundation, security agents, deterministic oversight agents) wraps the whole.
The Same Principles in Production
SAP's architecture paper is a vision. What makes it credible is seeing the same principles running in production — which is exactly what current SAP implementations show:
- 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 where permission is missing, and the LLM never touches the backend: it is a reasoning layer, not a backend caller. That 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 approval, everything is auditable.
- SAP Company Memory (GA Q2/2026): a system-agnostic (markdown-based) single source of truth that provides and governs business context for 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 is coming to 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 chain, tenant isolation, MCP tools, auditable prompt chain) is now SAP's endorsed architecture — enforced by API policy, not merely recommended.
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 mainstream:
| SAP North Star concept | Our implementation |
|---|---|
| MCP as tool integration | A dedicated MCP server with 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 configuration, no redeploy |
| Governed gateway / trust chain | A JWT bridge: Clerk session → tenant-mapping check → short-lived signed token → agent. Tenant isolation enforced at every step |
| Semantic grounding (RAG) | Uploaded PDF/DOCX/TXT knowledge base, chunked and embedded, ranked by cosine distance via pgvector |
| Data products / master data | Projects modeled per tenant, with target fiscal year and lookback period; master file → chapters → sections with full revision history |
| Business context / prompt governance | A fallback prompt chain (project override → agent configuration → SOUL.md → hardcoded) with explicit source attribution for the audit |
| Decision traces | Every agent persistence writes a revision — an audit trail of what was generated, when, and on what basis |
The architecture is already AI-native in SAP's sense: the model thinks, 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" records what was known, considered and chosen, and what followed — every human correction becomes a structured decision trace. For transfer pricing documentation this is gold: an audit trail that explains why a functional analysis took a given 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 an equal workflow. Master file and local file share context but require different evidentiary bases. The MCP tool model makes the extension trivial — 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 transfer pricing document generation should be multi-criteria: technical completeness, review pass-through rate, revision frequency, client acceptance. These 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 Bottom Line
SAP's North Star is not a product roadmap; it is a validation of an architectural direction. 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, governance by design, grounded in your data.