The short answer: there are three routes to SAP AI agents, and you should test them in order. First, check whether SAP already delivers a Joule Agent for the process. If one fits, use it. Second, if the work lives mainly in SAP and no delivered agent fits, build the agent with Joule Studio. Third, build a custom agent in code when the data sits largely outside SAP, or when you need full control over tools and testing. You can also combine the routes, as the case below shows. This guide sets out the four tests we use to choose, and what has to be in place before any agent reaches production.
A note on names first. SAP now uses the name Joule Studio for two products. “Joule Studio, classic edition” is the tool in SAP Build. By contrast, you reach the new Joule Studio through Joule Work. It builds agents, apps, workflows and extensions of SAP’s own agents, and it includes a skill builder. In this guide, “Joule Studio” means the new product.
The decision in one table
Each test below narrows the choice. Read the table across, then read the test that decides your case.
| Test | Delivered Joule Agent | Joule Studio | Custom agent |
|---|---|---|---|
| Does a delivered agent exist? | Yes, and it covers the process | No, or it needs an extension | No, and the process is not SAP-led |
| Where does the data live? | In SAP | Mainly in SAP | Largely outside SAP, or across many systems |
| How much control over tools and evaluation? | SAP’s design | Within Joule Studio’s current scope | Full control in your own code |
| Who runs it after go-live? | Your team, within what SAP delivers | A team that knows Joule Studio | Engineers who can maintain code |
Test 1: does a delivered agent already exist?
SAP delivers Joule Agents for its own processes. However, availability varies from agent to agent, and some have restricted availability. So, check the status of the agent you need, not a headline count. Our catalogue of Joule Agents is a good place to start. If a delivered agent covers the process, you avoid building and maintaining your own. Moreover, the new Joule Studio can extend SAP’s agents, so a near fit can still be the starting point.
Test 2: where does the data live?
If the agent needs SAP data and SAP processes only, Joule Studio is the natural fit. However, check its reach first. At general availability, the new Joule Studio cannot use external MCP servers yet. As a result, an agent that must reach non-SAP systems today usually points to a custom build. Alternatively, Joule can call an external agent that does the non-SAP work, which we cover in the last section.
Test 3: how much control do you need over tools and evaluation?
Some agents need custom tools, a specific model or a test harness that runs in your own pipeline. At general availability, the new Joule Studio does not include the pro-code extension CLI. Therefore, heavy pro-code needs point to a custom agent. On SAP BTP, a custom agent can use the generative AI hub in SAP AI Core, which SAP offers only with the extended service plan. If you are still choosing between retrieval, tools, workflows and agents, read our white paper on RAG, tools, workflows and agents first.
Test 4: who runs it after go-live?
After all, every agent needs an owner after release. Someone has to watch usage, cost and errors, keep the evaluation set current and be able to switch the agent off. In an SAP Joule implementation, this test matters as much as the other three. If your team will run the agent, choose the route whose tools your team can maintain. If nobody in-house will own it, plan the run service before you build, not after.
How the agent reaches SAP data, and where authorisations apply
Every route follows the same basic path, so the design questions repeat. The list below is a text diagram of that path.
- Request. A user asks the agent for an answer or an action.
- Tool choice. The agent picks a tool: an SAP API, an MCP server or another agent.
- Identity. The tool calls the SAP system under an identity.
- Authorisation check. The SAP system applies that identity’s authorisations before it returns data or posts anything.
- Result. The agent answers, or it drafts a write for a person to approve.
In practice, step 3 is where most of the design work sits. We allow one of two identities, and we review the choice before every release:
- The user’s own SAP authorisations. The agent can see and do only what that user may see and do. As a result, it cannot widen anyone’s access.
- A scoped technical user. One identity serves the agent for every user. For that reason, scope it to the minimum the use case needs, and make sure your logs tie each action back to the person who asked.
In short, the SAP authorisation check is the control. An instruction in a prompt is not.
Governance: registering agents in SAP AI Agent Hub, and what still needs testing
SAP AI Agent Hub is now generally available. SAP describes it as a way to govern every agent, LLM and MCP server across SAP and non-SAP environments. In addition, customers can register their own agents, as agent cards through an API. They can also register their MCP servers through discovery integrations. So a custom agent does not have to sit outside your SAP governance view.
Registration brings an agent into that governance view. Two questions still have to be tested on your side: does the agent give the right answer on your cases, and is its access scoped correctly? We answer both with our own release checks, which we describe below. For how we design these controls, see AI governance, evaluation and security.
What an AI engineer owns that a demo does not
For example, a demo shows an agent handling a good case. However, a production release needs more, and an AI forward-deployed engineer owns each part of it:
- The evaluation set. We measure quality against labelled real cases from your business before go-live.
- Tool permissions. The agent acts only with the user’s own SAP authorisations or a scoped technical user.
- The first production release. At first, a person approves every write or posting to SAP. We also monitor usage, cost and errors, and you can switch the agent off at any time.
Our AI forward-deployed engineers page explains how these checks work as release gates, and what you keep at hand-over. When the work needs core SAP build as well, such as ABAP, SAP BTP extensions or integrations, our SAP forward-deployed engineers work alongside them.
A worked example: Joule Studio and SAP BTP together
For a Saudi petrochemicals manufacturer, Altivate delivered AI built with Joule Studio and SAP BTP, using the generative AI hub. It covers finance, procurement and supply chain, and sales and customer service. It is now running in production across finance, procurement and customer service.
The lesson for this guide is the combination. The work did not rely on a single route. Instead, apply the four tests to each process, not once for the whole programme, because the answer can change from one process to the next.
Where the SAP path ends and a custom route starts
Two current limits shape the boundary. First, Joule Studio is now rolling out to general availability in phases. Second, at general availability, external MCP servers and the pro-code extension CLI are not yet available. Plan around both if your agent needs non-SAP tools now.
Beyond that, the Agent2Agent (A2A) protocol connects the two worlds. Joule can already call external agents, which SAP calls “Bring Your Own Agent”. SAP has also announced bi-directional A2A support, with general availability targeted for Q4 2026. Until it ships, do not build a plan that depends on it. For more detail, read what the Agent2Agent protocol means for SAP and how SAP Joule and A2A enable multi-agent enterprises. For the wider picture, see AI for SAP.
Frequently asked questions
What are SAP AI agents?
In SAP’s terms, Joule Agents are the agents SAP delivers. In addition, you can build your own agents, with Joule Studio or in code. For the delivered ones, see our catalogue of Joule Agents.
What is the difference between Joule Studio and Joule Studio, classic edition?
They are two products. First, the classic edition is the tool in SAP Build. On the other hand, you reach the new Joule Studio through Joule Work, and it is now rolling out to general availability in phases.
Can a Joule Studio agent use non-SAP data?
Not through external MCP servers at general availability. Instead, Joule can call an external agent over A2A, or you can build a custom agent for that part of the work.
How long does a first production release take?
With our teams, an engagement runs 8–12 weeks from kickoff to the first production release. A team is two or three engineers who work inside yours, for example an AI engineer and an SAP integration engineer.
Is SAP Business AI Platform replacing SAP BTP?
No. SAP announced in May 2026 that SAP Business AI Platform unifies SAP BTP, SAP Business Data Cloud and SAP Business AI. It does not replace BTP, and SAP’s help documentation still uses the BTP name.
Scope an embedded team
First, bring one process and the data it touches. We will run the four tests with you and outline what a first production release would take.
