Part 5 of the SAP Business AI Series | Back to Series Hub

There is a meaningful difference between AI that answers questions and AI that takes action. A copilot responds when you ask. An agent understands an objective, plans the steps required, executes those steps across systems, monitors outcomes, and adjusts when something changes — often without waiting to be prompted at each stage.

This shift from reactive assistance to proactive execution is what the term “Agentic AI” describes, and it represents the most significant evolution in enterprise AI capability since the emergence of large language models. This post explains what Agentic AI means in the SAP context, how Joule Agents work, what infrastructure makes them reliable, and how the new Agent-to-Agent (A2A) protocol enables AI collaboration across platforms.

What Is Agentic AI?

An AI agent is a system designed to act autonomously toward a goal — not just generate a response to a prompt. Instead of waiting for a human to specify every step, an agent receives an objective, breaks it down into a sequence of actions, executes those actions using available tools and data, and adapts its approach based on what it finds along the way.

Think about what a skilled procurement analyst does when asked to evaluate three shortlisted suppliers. They gather data from multiple sources, apply evaluation criteria, run calculations, draft a comparison, flag risks, and present a recommendation — without being told to open each system separately or instructed on how to format the output. An AI agent performs the same multi-step workflow autonomously, at machine speed and without forgetting a step.

The three developments that made agentic AI practically viable in enterprise settings:

1. Computing and algorithm advances — sufficient power to manage complex, multi-step reasoning in real time
2. Enterprise AI adoption — organisations have built the data foundations and integration infrastructure that agents need to operate
3. Large Language Models — the ability to understand natural language goals, plan decomposed steps, and communicate results in human-readable form

Joule Agents: SAP’s Agentic AI Building Blocks

Joule Agents are modular, task-specific AI systems pre-delivered by SAP and embedded in SAP applications. Each agent is expert in a specific business domain and carries out discrete, well-defined functions — creating purchase orders, approving timesheets, analysing talent gaps, forecasting sales pipeline performance.

The relationship between Joule and Joule Agents is like a project manager and a team of specialists. Joule is the orchestrator — the central intelligence that understands what the user wants and coordinates the right agents to deliver it. Joule Agents are the specialists — each one expert in a narrow domain, executing reliably within that domain while Joule handles the cross-functional coordination.

Current examples of pre-delivered Joule Agents include:

• A Purchase Order Creation Agent that takes a procurement request and creates a completed PO in S/4HANA
• A Timesheet Approval Agent that processes approvals based on configured policy rules
• A Talent Intelligence Agent in SuccessFactors that analyses employee skills and recommends development paths
• A Sales Forecast Agent in SAP CX that projects pipeline outcomes and surfaces accounts at risk
• A Dispute Resolution Agent in finance that identifies, categorises, and helps resolve accounts receivable disputes

What Makes Joule Agents Trustworthy: The Data Foundation

An AI agent that takes autonomous action is only as reliable as the data and context it is working from. SAP’s Joule Agents are built on two foundational infrastructure components that ensure their outputs are accurate, relevant, and grounded in business reality:

SAP Knowledge Graph

The SAP Knowledge Graph is the semantic backbone of SAP’s AI architecture. It connects data across SAP and non-SAP systems — finance, supply chain, HR, procurement, customer data — in a way that preserves the relationships between entities, not just the individual data points.

This matters for agents because business processes involve entities that are deeply interconnected. A purchase order is not just a document — it is connected to a supplier, a contract, a delivery schedule, a goods receipt, an invoice, and a payment term. The Knowledge Graph allows an agent to navigate these relationships intelligently, understanding context that a flat data query would miss entirely.

Practically, this means Joule can answer questions like “Do I have any open purchase orders that are at risk of delay?” — because the agent can traverse the graph from PO to delivery status to supplier performance history, rather than running three separate queries and manually correlating the results.

SAP Business Data Cloud

SAP Business Data Cloud unifies SAP transactional data and external data sources into a governed, context-preserved environment. It is the data lake on which the Knowledge Graph is built — ensuring that the information agents access has been validated, harmonised, and aligned to business meaning before agents act on it.

Together, the Knowledge Graph and Business Data Cloud create the conditions for agents to act with the same confidence that a knowledgeable human expert would — not because they are guessing well, but because they have access to the right information at the right level of context.

Building Custom Agents with Joule Studio

Pre-delivered agents cover common business scenarios well. But every organisation has processes, policies, and workflows that do not map to a standard SAP template. Joule Studio, part of SAP Build, provides the development environment for building custom Joule Skills and AI agents without requiring deep coding expertise.

The distinction between Skills and Agents is worth understanding:

• Joule Skills handle deterministic, rule-based tasks — fetch a stock level, submit a ticket response, look up a policy document. They follow a predictable execution path every time.
• Joule Agents handle more complex, adaptive workflows that require multi-step reasoning — logistics planning that accounts for multiple constraints, supplier negotiation that adjusts based on market data. They make decisions along the way rather than following a fixed script.

Joule Studio provides governance, versioning, and data masking as part of the development lifecycle — ensuring that custom agents meet the same compliance and security standards as SAP’s pre-delivered capabilities.

The AI Agent Hub: Managing the Agent Ecosystem

As organisations deploy more AI agents — from SAP, from third parties, and built in-house — managing them becomes a governance challenge. SAP’s AI Agent Hub, within SAP LeanIX, provides a centralised platform for the full agent lifecycle.

The AI Agent Hub enables organisations to:

• Discover available agents across internal and external sources
• Understand which business processes each agent supports
• Ensure agents can work together effectively
• Assess the cost and benefit of deploying each agent
• Identify automation gaps where new agents would create value
• Govern agent usage through policies and oversight mechanisms
• Evaluate workforce impact, helping HR teams plan upskilling

Agent-to-Agent (A2A) Protocol: AI Systems That Collaborate

As AI agent deployments mature, a new challenge emerges: how do agents from different vendors, built on different platforms, collaborate on shared tasks? An SAP Joule Agent handling a customer escalation might need to delegate a fulfilment action to a third-party logistics agent. Without a shared communication standard, this requires custom integration work for every combination.

The Agent-to-Agent (A2A) protocol, co-developed by Google Cloud and SAP, addresses this directly. It is an open standard that enables secure, scalable collaboration between AI agents regardless of who built them or where they run.

How A2A Works

A2A communication involves two roles:

• Client Agent — the agent that initiates the interaction by formulating and sending a task
• Remote Agent — the agent that receives the task and executes it

This interaction is structured around four capabilities:

1. Capability Discovery — agents advertise their skills in a standardised Agent Card format (JSON), so other agents know what they can do
2. Task Management — A2A is task-oriented. The client sends a task; the remote agent executes and reports back
3. Collaboration and Messaging — agents exchange structured messages as they work toward task completion
4. User Experience Negotiation — agents communicate what content types they can handle (text, images, forms, structured data)

A2A vs MCP: Understanding the Difference

Both protocols are complementary rather than competing. MCP handles the connection between an agent and a system. A2A handles the coordination between agents. Together, they form the communication infrastructure for a multi-agent enterprise ecosystem.

A Cross-Functional Agent in Action: CFO Use Case

Consider a CFO asking Joule: “What is hindering our cash collection, and how do we improve it?” This is not a simple data retrieval question — it requires cross-functional analysis.

Joule orchestrates a sequence of agents: a finance agent analyses open receivables and dispute patterns; a customer experience agent identifies accounts with poor service history; a supply chain agent flags delivery delays that correlate with disputed invoices; a risk agent profiles customer payment behaviour over time. Each agent contributes its domain expertise, and Joule synthesises the results into a coherent analysis with actionable recommendations — all in response to a single natural language question from the CFO.

Key Takeaways

• Agentic AI takes action autonomously toward a goal — it does not just respond to prompts
• Joule Agents are modular, pre-delivered AI specialists embedded in SAP applications
• The SAP Knowledge Graph and Business Data Cloud provide the contextual foundation that makes agent outputs trustworthy
• Joule Studio enables organisations to build custom agents without heavy coding
• The A2A protocol enables AI agents from different vendors to collaborate securely at scale
• MCP and A2A are complementary standards — one for agent-to-system, one for agent-to-agent communication

Next in the series: Post 6 — SAP AI Across Business Functions →