Building an Analytics Agent with Joule Studio
As of 2026-09-25
Takes a Joule Studio agent from M326's fundamentals to one complete, working analytics agent: an agent that answers an ambiguous variance question by calling tools, reconciling their results, and refusing to guess when they do not explain the movement. Covers the agent builder's three components (instructions, tools via MCP-capable BTP destinations, knowledge sources including the SAP Knowledge Graph, LeanIX and Domain Models over a remote MCP server), when multi-agent orchestration earns its complexity, a full worked build of a EMEA-gross-margin agent (tool schemas, instructions, guardrails, edge-case tests), and the AI-Units metering that applies to agent runs — kept strictly separate from generative-AI-hub token/CU metering — including a dated note that SAP's May 2026 free-run promotion has already ended. Closes with a production-readiness checklist and three exercises.
What you will learn
- Explain what an analytics agent adds over a grounded Q&A chatbot, and why the decision loop — not retrieval — is the agent
- Describe the three components of a Joule Studio agent (instructions, tools, knowledge sources) and how tools reach Datasphere or HANA Cloud through an MCP-capable BTP destination
- Design one analytics agent end to end for a named business question: tool schemas, tool-call order, and a written refusal rule for unsupported answers
- State when to split an agent into a multi-agent design versus keep one agent with a short tool list
- Separate, correctly, how generative-AI-hub token/CU metering and Joule agent AI-Unit metering apply, and avoid carrying a lapsed promotion into a cost estimate
- Build a production-readiness checklist for a Joule agent: tool ownership, evaluation set, call logging, fallback behaviour, review cadence
Module overview
Who this is for. You have built a first Joule agent in M326 and you know what MCP is from M327. This module does one thing: it takes those pieces and builds a single, complete analytics agent — an agent whose job is to answer an ambiguous business question ("why did gross margin slip in EMEA last quarter?") by calling real tools, reconciling their answers, and explaining the number in words a finance director can act on. M351 covers how to build the MCP server this agent calls; M352 covers grounding it in documents; this module covers the agent itself — its instructions, its tools, its guardrails, and what a run costs.
Prerequisites
- M326 — Building Joule Agents with Joule Studio (agent fundamentals)
- M327 — MCP & A2A for SAP Systems (protocol concepts)
- Working knowledge of a Datasphere or HANA Cloud data model
- Recommended alongside M351 (the MCP server this agent's tools call) and M352 (document grounding for the second tool)
Outcomes
- Produce a written agent brief (decision, tools, instructions, guardrails) for a real analytics question from your current engagement.
- Ship a tool schema and instruction set that make an agent refuse an unsupported answer instead of guessing.
- Justify, in one paragraph, a single-agent versus multi-agent design decision.
- Build a metering line in a cost estimate that correctly separates AI Units from generative-AI-hub capacity units.
Full module available to members. The full module adds: the decision framework · the end-to-end scenario walkthrough · the KPI scorecard · the anti-patterns · the code blocks · the knowledge check · the diagrams.