Joule agents: from assisted analytics to autonomous workflows

Joule agents in 2026 operate across three integrated layers, each purpose-built for scalable, governed intelligence:

Key capabilities in November 2026: You build agents in Joule Studio either in the browser-based builder or pro-code, with the Joule Studio CLI and a coding agent in your own IDE. You define an agent's expertise and instructions, choose its LLM, and give it tools: Joule skills, SAP APIs, document grounding and MCP servers—Joule Studio can even generate an MCP server from an existing SAP OData or REST API. Optional human-in-the-loop steps keep people in control. Deployed agents run on the SAP-managed runtime and register automatically with Joule, so users reach them in conversation. Agents you host yourself can connect to Joule through the Agent2Agent (A2A) protocol, and SAP planned Joule's bi-directional A2A capabilities for general availability in Q4 2026.

Joule agents are AI-powered software assistants that automate multi-step business workflows while maintaining human oversight and governance. Unlike chatbots that respond to user queries, agents proactively execute workflows: investigating problems, gathering data from multiple sources, making decisions within defined boundaries, and escalating exceptions to humans when needed.

How Joule agents differ from previous Joule capabilities

Joule as assistant (baseline): Answers questions in natural language—in SAC through just ask and analytical insights—explains, navigates and performs tasks through Joule skills. User-initiated and reactive. Example: a user asks "What drove Q4 revenue variance?" and Joule answers.

Joule Agents: Plan and carry out multi-step work: chain tools, call APIs, take action in SAP and non-SAP systems, and hand decisions to people where needed. Example: an agent investigates a Q4 variance, queries the governed finance data it is allowed to see, checks known drivers, drafts the top findings, and asks a controller to approve before anything is published.

Custom agents and models (advanced): Agents you build in Joule Studio, or host yourself and connect through A2A, combined with models from the generative AI hub—including SAP-RPT for tabular predictions—or your own models served in SAP AI Core. Example: an agent combines an SAP-RPT prediction of late deliveries with a supplier-risk score from your own model and the current plan, and recommends procurement actions for a buyer to approve.

The shift: Previous Joule versions accelerated how fast humans could work. Joule agents take repetitive work off people's desks while keeping humans in the loop for decisions that matter.

Joule agents are ideal for repeating, high-friction business workflows that involve:

• Multi-step data gathering from different systems (Datasphere, SAC, S/4HANA, external APIs)
• Pattern recognition and anomaly detection against governed KPI definitions
• Narrative generation that summarizes findings in business language
• Clear decision boundaries (agent can make routine decisions; escalates exceptions to humans)
• Repeating business cycles (weekly, monthly, quarterly) that would benefit from automation
• Audit trail and governance requirements (all decisions logged; human approval checkpoints built in)

Ideal use cases

Monthly Financial Close Commentary: Agent reads actuals and plan through the APIs your finance data already exposes, compares them against agreed variance thresholds, drafts commentary with an LLM, flags high-variance items for review, and posts the approved narrative where the team works. A controller approves before anything is published. Measure: days spent on commentary per close.

Demand Forecast Scenario Generation: Agent picks up the latest baseline forecast—for example a Smart Predict forecast written to a planning version, or SAP-RPT-TS in SAP IBP—applies a standard set of scenario assumptions (growth, flat, decline, promotion, price), summarizes the supply impact from the data it can access, ranks the scenarios by risk and sends the pack to the planning team. Measure: hours from new actuals to a reviewed scenario pack.

Supply Chain Risk Dashboard: Agent queries supplier performance from governed data, calls external risk APIs (geopolitical, weather, shipping indices), scores each supplier with a model you trust, flags high-risk shipments and recommends alternatives for a buyer to confirm. Measure: hours of manual monitoring per day and disruptions caught early.

Sales Pipeline Weekly Executive Summary: Agent reads the CRM pipeline, scores deal risk with SAP-RPT using historical wins and losses as context, identifies deals trending away from forecast, surfaces the top blockers and recommended actions, and drafts the weekly summary. Measure: hours spent building the weekly report and forecast accuracy over time.

SAP does not ship a catalog of analytics agents such as a "variance investigation agent." It delivers Joule Agents inside its applications—finance, supply chain, procurement, HR and customer experience. SAP's Q1 2026 release highlights reported more than 30 specialized agents and over 2,500 Joule skills, and the SAP Autonomous Suite adds more than 50 Joule Assistants that orchestrate over 200 agents as they roll out. Analytics workflows such as variance commentary are usually agents you build yourself in Joule Studio. The cards below show real SAP-delivered agents with the status SAP reported when it announced them, followed by the analytics patterns you build on your own.

Customization: Joule Studio lets you build and extend agents, but how far you can change an SAP-delivered agent depends on that agent, so check its documentation. Where an agent does not fit your process, build your own and keep the SAP agent for the standard case.

Joule Studio has two paths. In the browser-based builder, business analysts and developers define an agent's instructions, tools and human-in-the-loop interactions without any local setup. In the pro-code path, developers use the Joule Studio CLI with a coding agent connected through MCP, in their own IDE with Git and CI/CD. Both deploy to the same SAP-managed runtime. The agent builder has been generally available since December 2025 (that original edition in SAP Build is now called the classic edition), and the new Joule Studio on the SAP Business AI Platform has been rolling out in phases since October 2026.

Studio Capabilities

• Agent definition: Expertise, instructions and context, written in natural language, tell the agent what to do and how to act.
• Model choice: Select the LLM that powers the agent's reasoning, with an optional backup provider; the new Joule Studio offers more than 70 AI models.
• Tools: Joule skills (rule-based, built in the skill builder), SAP APIs, MCP servers—including ones Joule Studio generates from SAP OData or REST APIs—and document grounding on your own content.
• Human-in-the-loop: Define where users confirm, choose or approve before the agent acts.
• Testing: Test agents in a private environment before release, then evaluate their performance continuously.
• Lifecycle and governance: Identity and access controls including SAP Cloud Identity Services, managed runtime and lifecycle controls, and visibility into agent behavior, decisions and failures.
• Workflows: n8n workflows that orchestrate agents across enterprise systems are open to early adopters, with broader availability to follow.

Studio Workflow: Building a Custom Agent in 5 Steps

1. Define the use case: Write a simple problem statement. E.g., "Every Monday, analyze employee absences, identify unexpected spikes, and alert HR managers."
2. Write the instructions: Describe the agent's expertise and the steps it should take—retrieve the week's absences, compare them with the usual pattern, draft an alert, and ask for confirmation before sending—and choose the LLM.
3. Give it tools: Add a Joule skill or an MCP server that reads absence data through an approved API—for example an MCP server generated from an existing OData service—with filters for the last seven days and approved leave excluded.
4. Set rules and checkpoints: State the threshold in the instructions or in a skill—for example, flag a team when absences exceed its usual average by two standard deviations—require HR confirmation before anything is sent, and escalate to the head of HR above an agreed share of the workforce.
5. Test and deploy: Test the agent in your private environment on past weeks and check that its outputs make sense, then deploy it to the managed runtime, where it registers with Joule. Monitor its behavior and failures, and decide how it starts each Monday: by a user, from a workflow, or on a schedule your release supports.

Advanced: pro-code agents

For logic beyond the browser-based builder, use the pro-code path:

• Coding-agent development: The Joule Studio CLI works with a coding agent such as Claude Code or OpenCode, connected through MCP, in your own IDE.
• Models as tools: Call SAP-RPT or other models in the generative AI hub, models you serve in SAP AI Core, or external ML services such as Vertex AI or SageMaker through their APIs.
• Bring your own agent: Agents you build and deploy in your own BTP subaccount can connect to Joule through A2A.

Example: A procurement agent reads the demand baseline from the plan, calls SAP-RPT to predict late deliveries, calls an external supplier-risk API, and drafts recommendations for a buyer to approve. The instructions, tools and approval step live in Joule Studio; the predictive pieces are API calls.

Organizations often compare Joule agents to Microsoft Copilot for Power BI and Gemini in BigQuery. This section clarifies the differences and when each makes sense.

Agent2Agent (A2A) is an open protocol that lets agents from different vendors call each other. SAP's Q1 2026 release highlights already said its agent-to-agent protocol lets SAP agents work across SAP and non-SAP systems; in May 2026 SAP said agents built in Joule Studio will support A2A natively and that Joule's bi-directional A2A capabilities would be generally available in Q4 2026. At TechEd 2026, SAP added a governed agent gateway through which third-party agents invoke Joule Agents. Inside Joule, Joule Assistants coordinate Joule Agents, and Joule Work orchestrates the steps that assistants and agents need.

How to design for A2A: Treat each agent as a service with a clear contract: what it accepts, what it returns, and which decisions it may take. Agents you host in your own BTP subaccount connect to Joule through A2A, and external agents reach Joule Agents through the agent gateway. Keep each agent small and reusable rather than building one monolithic agent.

Example design (supply-to-cash): An orchestrating agent collects the demand baseline from planning, asks a supply agent to check material and capacity constraints, and asks a pricing agent you built for recommendations. It detects conflicts—high demand against constrained capacity—and escalates to a manager before an integrated plan is published. Each agent keeps its own permissions and escalation rules.

A2A benefits: Separation of concerns (each agent is simple and maintainable). Reusability (a supply-check agent serves multiple parent workflows). Parallel execution (multiple agents run simultaneously, reducing end-to-end time). Governance clarity (each agent has defined inputs/outputs and escalation paths). Easier testing and debugging (test agents independently before composition).

Governance in A2A: Define the thresholds at which a sub-agent's recommendation needs a human—for example, escalate to the demand planner when recent forecast error exceeds an agreed limit—and register every agent and MCP server in SAP AI Agent Hub so you know what exists, who owns it and what it can reach. Log agent-to-agent calls so decisions can be reconstructed.

The Model Context Protocol (MCP) is an open standard for exposing tools and data to AI agents. In Joule Studio, an MCP server is a tool source: you register MCP servers at design time and attach them to agents, and Joule Studio can generate an MCP server from an existing SAP OData or REST API and deploy it to the managed runtime. Joule Work uses open protocols including MCP to orchestrate work, and SAP AI Agent Hub discovers, inventories and governs MCP servers alongside agents.

What MCP enables: Tools described once and reused by many agents; new capabilities added by registering an MCP server instead of changing every agent; consistent access to SAP and non-SAP systems; and one place to see which agents can reach which tools. The flip side is governance: every MCP server is a door into a system, so register it, scope its permissions and give it an owner.

Practical starting points (November 2026): MCP servers generated from the SAP OData and REST APIs you already use; third-party MCP servers, such as Google's for BigQuery and other databases; and MCP for coding agents in the Joule Studio pro-code path.

Agents operate within governance boundaries with clear escalation paths. Three escalation levels ensure human oversight:

Level 1: Routine (Agent Autonomous): Low-risk, high-volume decisions with clear rules. Example: Approve purchase orders <$1K from existing approved vendors. Agent executes autonomously, logs decision, daily audit report generated.

Level 2: Guided (Agent Recommends + Human Approves): Medium-risk decisions where context matters. Example: Recommend sourcing changes based on spend analysis and supplier risk. Agent gathers data, generates recommendation with rationale. Human approves before execution.

Level 3: Strategic (Human Decides): High-risk, low-volume decisions requiring judgment. Example: Pricing changes >5%, new supplier onboarding, workforce restructuring. Agent provides analysis and recommendation to executive. Executive makes decision and documents rationale.