Guide

SAP Joule Studio consultant: building, connecting and governing agents

As of 2026-10-07

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

A SAP Joule Studio consultant designs, builds and governs custom agents for Joule: scoping the use case, configuring the agent builder, wiring tools through MCP and A2A, placing approval points where an action cannot be undone, and handing the result to an operations owner. The product has its own guide; this page covers the person you hire. On 7 October 2026 none of the 1,657 live postings in the Analytics Legends corpus named Joule Studio, while 95 named Joule, so the skill is usually sold under a wider title.

The role in four jobs, and what it is not

The work has four parts: choose what to build, build it, connect it to the systems it acts on, and control what it may do. A consultant who only configures prebuilt agents does a different job: SAP's delivered agents are activated, not built. The Joule Studio consultant starts where that catalogue stops, when no standard agent exists for the process in the customer's edition and region, or the gap is limited to tools, instructions or data sources.

The role also differs from the platform engineer's: AI Core and the Generative AI Hub are where a pro-code agent calls its models, whereas the Joule Studio consultant decides whether the agent is worth building and owns how it behaves in a conversation.

Two studios, two skill sets

Joule Studio now exists in two forms. The classic edition sits in SAP Build: the skill builder became generally available in Q2 2025 and the agent builder in Q4 2025. The new studio, announced on 13 May 2026, opens from Joule Work and adds a command-line tool, a managed runtime and an Agent Gateway. As at 23 September 2026 we found no SAP statement declaring it generally available; it was in Early Adopter Care. The two editions carry separate SKUs, roles and provisioning paths, so one subscription does not cover both.

The first question of every engagement is therefore which edition, and whether the client accepts programme conditions. Classic suits a low-code team; the new studio suits professional developers who want continuous delivery. Design-time access is free to the end of 2026, but SAP's continuous-improvement terms let it process configurations and execution logs, which matters before regulated data is involved.

Building: the agent builder and Build Process Automation

An agent in the classic builder is defined by its expertise, instructions, a planning model and a response model, a budget of thinking steps, and a small toolset: skills, document grounding, a calculator, subagents and MCP servers. The step budget runs from 1 to 100. A larger budget or a backup model trades latency and AI Units for resilience, so it belongs on agents whose failure is expensive.

Write actions are where reviews go wrong. SAP's reference pattern for an action that changes data is a process automation invoked from inside a skill, not a bare REST call from the agent, because stateful, multi-step actions need retry and approval behaviour. Agents are designed in Build Process Automation but run in a Joule instance that is often another subaccount; destinations such as MCP credentials must exist in both, and a missing one surfaces as a generic tenant error at deployment.

Treat multi-agent designs with suspicion: classic subagents run sequentially, up to 5 direct and 25 in total, so latency adds up. The default is one agent with three to five tools.

Connecting: MCP and A2A, hop by hop

The working rule: a person reaches the agent through the interface, the agent reaches SAP data and actions through MCP, and a foreign agent through A2A. Wrapping a partner agent as an MCP tool is acceptable only for short, synchronous work.

The limits are concrete. Joule expects an A2A agent server to answer within 60 seconds; longer work needs push notifications, which need identity trust between both sides. For code-based agents SAP's samples document direct A2A (one minute), an asynchronous REST interface (five minutes) and a streaming client service beyond that. The classic builder accepts MCP servers only over streamable HTTPS with client credentials and without principal propagation, so a server that needs per-user isolation cannot connect there at all. SAP's API policy limits agentic use of SAP APIs to endorsed routes such as the Integration Suite gateway, so scope this before designing tools.

Governing: approval points, guardrails, inventory

Three control points are easily confused: confirmation on a tool inside a turn, agent-initiated pauses for missing input, and lifecycle gates such as the approval before production deployment. Verify all three. Match the checkpoint to reversibility: postings, payments and HR data get a confirmation before execution; reversible, high-volume drafting can run with sampling afterwards. Where the AI Act's human-oversight article applies, a bare yes-or-no prompt lets no reviewer understand or override anything.

Guardrails sit at environment level, with masking and filtering at each step. SAP AI Agent Hub, built on SAP LeanIX, inventories agents, MCP servers and models; it governs, but does not run or trace them.

Decision table: which route for which case

Process already covered by a standard SAP agent — Route: activate or extend · Check: scope item, release status, edition and region · Risk: building what SAP ships.

No standard agent, low-code team — Route: classic agent builder · Check: environments, audit log, MCP constraints · Risk: a tool needing per-user identity cannot connect.

Pro-code agent or continuous delivery — Route: new studio command line, or AI Core plus A2A · Check: Early Adopter Care status, 60-second limit · Risk: go-live date set before general availability.

Write action to a core system — Route: process automation invoked from a skill · Check: approval step, retry behaviour · Risk: a bare call with no approval.

Several domains, separate scopes — Route: supervisor with subagents or A2A peers · Check: owner per specialist · Risk: sequential latency, blurred accountability.

Decisions about people or credit — Route: agent with a mandatory human decision, legal classification first · Check: provider or deployer status · Risk: a high-risk system built as a pilot.

Cost and market: a worked example

SAP meters delivered agents at 0.02 AI Units per agent action in 2026; assistants carry no separate charge. As an illustration with assumed volumes, an agent taking 3 actions per run and running 10,000 times a month consumes 30,000 actions, or 600 AI Units. SAP's own example prices an AI Unit at EUR 7, an example rather than a list price, which would give EUR 4,200 a month before negotiation. Custom-built agents are billed differently: runtime goes through the BTP agreement. Consumption is SAP's fee, paid by the client to SAP; never blend it into a day rate.

In the corpus of 7 October 2026, 95 of 1,657 live postings name Joule: 87 permanent, 6 freelance and 2 contract, with Germany 35 and the United States 17.

What we cannot assert

No SAP statement we hold declares the new Joule Studio generally available, and no source publishes a day rate for the role. Corpus counts depend on how skills are tagged in postings and mix SAP-specific and generic AI wording; they measure advertised demand, not total demand. The AI Unit calculation uses assumed volumes.

Frequently asked

Is Joule Studio the same thing as SAP Build?

The classic edition is a component of SAP Build on Build Process Automation. The new studio opens from Joule Work and has its own SKU, roles and provisioning, so check which one the client holds.

Does a Joule Studio consultant need to write code?

Not in the classic builder, which is low-code; the new studio adds a command-line route. Either way MCP, A2A and approval design must be understood.

Which certification applies?

C_AIG_2604 tests the Generative AI Hub (3 hours, 76 % pass mark). We hold no source for a Joule Studio exam.