The gap between the demo and the go-live
SAP’s agentic stack is moving quickly. At Sapphire in May 2026, SAP positioned more than 50 domain-specific Joule Assistants, each orchestrating a subset of more than 200 specialized agents, and described Joule Studio as “SAP’s AI-first solution for building enterprise agents, applications and agentic workflows”. None of that tells you how to take your first agent from a workshop idea to something running in production against live ERP data. That path has five stages, and the failure points are rarely the technology.
Stage 1: qualify the use case before you build anything
SAP’s own architecture guidance is refreshingly plain here: identify business challenges suitable for AI, evaluate feasibility, and define success metrics or evaluations (“evals”) — in that order. In our delivery practice at IOTEK, we sharpen that into three qualifying questions: Is the process painful enough that someone owns the outcome? Is the data the agent needs actually reachable and reliable? And can you write down, before building, what “correct behavior” looks like — because that definition becomes your test suite later. A use case that fails any of the three is a demo, not a project. (That three-question rubric is IOTEK’s method, not an SAP-published standard.)
Stage 2: choose your build environment — Joule Studio or SAP AI Core
SAP now offers a genuine spectrum. At the accessible end, SAP provides “a growing library of ready-to-use Joule Agents that are embedded across SAP applications for every business function”, and lets customers build their own agents using Joule Studio’s no-code Agent Builder, which creates custom AI agents without coding, using SAP’s business process expertise and data integration. Trade coverage reports that SAP has announced general availability of the Agent Builder capability in Joule Studio, first previewed at SAP TechEd 2025 — note that is SAPinsider’s wording, corroborated by SAP’s own announcements but not a dated SAP GA statement for the full Joule Studio offering.
The new managed Joule Studio, announced in May 2026, is a “fully managed offering that empowers enterprises to build and manage the full life cycle of AI agents, applications, and workflows”, producing agents natively grounded “in live business data, end-to-end processes, and rich business semantics”. SAP is also offering free design-time access through the end of 2026 under fair-use limits — a time-boxed commercial offer, not a statement of general availability, and worth exploiting while it lasts.
For behavior those no-code paths cannot express, the pro-code route runs through SAP AI Core and the Generative AI Hub — the same stack whose Document Grounding module sits in the Orchestration Service. Our rule of thumb at IOTEK: start with the least code that expresses the behavior, and move down the stack only when a requirement forces you to.
Stage 3: grounding is the project, not a feature toggle
The difference between a useful agent and a confident liar is grounding. SAP’s Document Grounding implements retrieval-augmented generation (RAG), using the SAP HANA Vector Engine to retrieve context from relevant documents and generate more accurate responses. At the platform level, SAP Knowledge Graph gives AI agents a structured map of business entities, processes and relationships across a customer’s SAP landscape, inside a Business AI Platform that now unifies BTP, Business Data Cloud and SAP Business AI into a single, governed environment. Practically, that means your first agent project includes a data workstream: deciding which documents, tables and semantics the agent may see, and verifying they are current and correct. Skip it and the agent will be exactly as good as your worst master data.
A note on the Claude roadmap
SAP and Anthropic announced they “will collaborate to embed Claude’s agentic capabilities into the newly announced SAP Business AI Platform”, with agents leveraging Claude connecting to the platform to ground decisions in SAP data and SAP saying Claude “will empower agents to carry out tasks — from closing the books at quarter-end… to rerouting supplier orders mid-shipment”. Every one of those statements is future-tense at source: this is announced direction, not a capability you can switch on today. Architect with it in mind; do not put it on your go-live critical path.
Stage 4: test the agent like a hire, not like a transport
Here is an honest gap: SAP’s public announcements say little about how to test an agent, and we found no fetchable SAP guidance on agent-specific test methodology at the time of writing. What follows is IOTEK’s delivery practice, stated as such. We treat an agent like a probationary hire: build an evaluation set from the success criteria defined in Stage 1 (real historical cases with known-correct outcomes); run the agent against it and score every deviation; red-team it with malformed, ambiguous and adversarial inputs; and require a human sign-off gate on any action that writes to the system of record. The eval set is not throwaway — it becomes your regression suite for every model, prompt or grounding change after go-live.
Stage 5: hypercare for something that behaves probabilistically
Classic SAP hypercare watches transactions and interfaces. Agent hypercare — again, IOTEK practice rather than SAP-published methodology — must additionally watch decision quality: sample and review agent outputs daily in the first weeks, track override and escalation rates, monitor grounding sources for drift, and keep a one-switch path to pause the agent without disrupting the underlying process. Plan the exit criteria before go-live: an agreed accuracy threshold sustained over an agreed period, not a calendar date.
Your first production agent is won or lost in the unglamorous stages — use-case qualification, grounding, evals, hypercare. The build environment is the easy part; SAP has made sure of that. The discipline around it is where consultancies earn their keep, and where we would rather be judged.
Sources
| # | URL | Publisher |
|---|---|---|
| 1 | https://news.sap.com/2026/05/new-joule-studio-enterprise-scale-agentic-development/ | SAP News Center |
| 2 | https://news.sap.com/2026/05/sap-anthropic-to-bring-claude-sap-business-ai-platform/ | SAP News Center |
| 3 | https://news.sap.com/2026/05/sap-sapphire-sap-unveils-autonomous-enterprise/ | SAP News Center |
| 4 | https://learning.sap.com/courses/boosting-ai-driven-business-transformation-with-joule-agents/exploring-out-of-the-box-agents-and-custom-joule-agents | SAP Learning |
| 5 | https://architecture.learning.sap.com/docs/ai-golden-path | SAP Architecture Center |
| 6 | https://help.sap.com/doc/generative-ai-hub-sdk/CLOUD/en-US/_reference/document-grounding.html | SAP Help Portal (SDK reference) |
| 7 | https://sapinsider.org/blogs/joule-studio-agent-builder-hits-general-availability-signaling-shift-in-agentic-ai/ | SAPinsider (Tier C, corroborated) |