The platform behind 90%+ load rates and agents that run in production.

The DEBCOR AI Platform is a five-layer agent architecture purpose-built for SAP landscapes. Orchestration Agents route work, Intelligence Agents reason over your data, Worker Agents execute tasks (data cleansing, order validation, code classification), Governance Agents enforce segregation of duties and authorisation policy before any action runs, and Auditing Agents produce tamper-evident trails for every step. Underneath, it runs on LangGraph for agent orchestration, n8n for workflow design, and Anthropic Claude as the reasoning engine — connected to SAP via MCP, Integration Suite, and the standard BTP services. It has been running in production with SAP clients for roughly 18 months before SAP announced the same stack at Sapphire 2026, and three patents are pending across the suite. It ships as discrete products (AI Data Engine, AI Data Cleansing Tool, Order Validation Engine, Memoria, Payment Card Engine, SAP Joule Agent) deployed inside engagements — not as a standalone SaaS subscription.

SAP Joule is the AI layer woven into SAP's applications — copilots and, with Joule Studio 2.0, custom agents grounded in SAP's Knowledge Graph and Company Memory. The DEBCOR AI Platform is complementary, not competitive. Where Joule is best used: standardised scenarios SAP has built native agents for, RISE customers activating entitlements they have already paid for, and any organisation that wants the SAP-managed runtime. Where the DEBCOR platform is best used: scenarios Joule does not cover natively yet, regulated environments that need explicit governance and auditing as architecture (the two trust layers), non-RISE landscapes (ECC, public cloud, hybrid), or any case where the answer needs the Company Intelligence Layer's client-specific learning. In practice DEBCOR delivers both — Joule Studio 2.0 agents for the use cases SAP has paved, and LangGraph/n8n/Claude agents via MCP for the rest — under one architectural and operating model.

Most agent demos are a model with API access. That works for chat. It fails for SAP. A production SAP agent has to enforce segregation of duties, authorisation scoping, approval thresholds, and produce evidence for SOX, PCI, GDPR, and the regulator of the day — and those controls cannot be optional or bolted on after the agent is shipping work into the system. By making Governance (Layer 4) and Auditing (Layer 5) explicit agents in the architecture, every proposed action by an Orchestration, Intelligence, or Worker agent passes through them before it executes, and every step is logged in a tamper-evident trail. This is the architectural difference between 'an AI we deployed' and 'an AI we can defend.' It is also why CIOs at regulated organisations choose this platform over generic agentic frameworks that treat governance as middleware.

Yes — and it is one of the highest-value entry points. RISE contracts already include Joule entitlements, and most organisations do not fully deploy what they are paying for. The standard sequence: audit the entitlement, activate Joule against your actual SAP data via the Company Intelligence Layer (grounding, custom agent registration, governance policy), and deliver the first custom Joule Studio 2.0 agents in 60–90 days. For scenarios Joule does not cover natively, the same architecture extends via LangGraph and Anthropic Claude over MCP — without modifying SAP core. The result is a single agent estate that uses Joule where SAP has invested the most, and the external stack where Joule has not arrived yet — all governed and audited by the same platform.

No — and this is a design choice, not a marketing position. Reasoning happens in the Generative AI Hub on BTP (with SAP's own grounding, templating, data masking, and I/O filtering controls), or against on-premise or private S/4HANA via MCP with destination-service security and authorisation scoping. Sensitive fields are masked at the boundary by the Governance Agent layer before any model sees them. Audit logs of every read, every reasoning step, and every action stay within the customer's tenancy. For regulated environments — pharmaceutical, financial services, aerospace and defense — this matters more than every other selling point combined. The architecture was designed so that 'no, the data does not leave' is the truthful answer, not a sentence written to win a security review.

Generic RAG (retrieval-augmented generation) attaches a vector index to a model so it can answer using documents instead of training data. That is table stakes. The Company Intelligence Layer is structural: it captures the patterns specific to a client's SAP system — their actual chart of accounts, their pricing condition keys, their custom IDoc types, the failure modes that recur in their landscape, the manual workarounds their team has learned over years — and grounds every agent in that institutional memory. The longer a managed services engagement runs, the smarter the layer becomes. By quarter four, the agents are catching exceptions Joule's generic models miss because the client's specific pattern looks normal to a model trained on the population, but anomalous given that client's history. It is the difference between AI that knows SAP and AI that knows your SAP.