AI Agents in ERP: Why SAP, Microsoft and Oracle Are Betting on Autonomous Processes

7 min read

SAP has merged Joule with Microsoft Copilot, Microsoft is transforming Copilot Studio into an autonomous agent, and Oracle has launched 22 agentic applications in a single release. Within three months, all major ERP vendors have integrated AI agents into their systems. Gartner predicts that 40 percent of all enterprise apps will contain task-specific AI agents by the end of 2026. For mid-sized companies, the question is no longer whether, but when their ERP landscape will operate on an agent-based model.

What are AI agents in ERP systems?

Traditional ERP systems are passive databases. They store orders, inventory levels, and financial data, but they don’t act independently. When an order comes in, an employee must check inventory, confirm availability, and trigger shipping. Every step requires a human.

AI agents fundamentally change this principle. They recognize patterns in transaction data, make decisions within defined rules, and execute actions autonomously. An AI agent in the supply chain identifies an impending material shortage, checks alternative suppliers, compares prices, and creates a purchase recommendation. Not in minutes, but in seconds.

The distinction from copilots is crucial: A copilot answers questions and makes suggestions. An agent acts. SAP, Microsoft, and Oracle have completed this paradigm shift within just a few months – ushering in a new phase of process automation.

SAP Joule: From assistant to autonomous process manager

On January 14, 2026, SAP announced the most far-reaching integration in its history: Joule, SAP’s proprietary AI assistant, merges with Microsoft Copilot. In practical terms: A project manager in a Teams meeting can ask Copilot to query current inventory levels for a specific component via Joule – and receives the result in natural language directly in the chat.

But SAP’s strategy goes beyond simple chat integration. New Joule agents for the supply chain independently verify prerequisites for production orders, identify material shortages, and propose solutions. The key point: A Joule agent can commission a Microsoft Copilot agent and vice versa. SAP and Microsoft are building a multi-agent ecosystem where different specialized agents collaborate.

For S/4HANA customers in the mid-market, this means: AI arrives not as a separate product, but as an integral part of the existing system. Those who have already migrated to S/4HANA Cloud receive Joule agents as part of their license model. Those still running the legacy ECC system have one more reason to accelerate their migration.

Microsoft Copilot Studio: Dynamics 365 becomes an AI workplace

In April 2026, Microsoft took the next step: Copilot Studio transforms from a chatbot builder into a platform for reasoning-based AI agents. These agents independently interact with applications, execute complex business processes, and make decisions with minimal human oversight.

Microsoft’s goal is clear: Copilot Studio should no longer just respond, but act autonomously. The updates transform the simple chatbot into a fully autonomous AI worker for business processes.
Based on the Microsoft announcement, April 2026

For Dynamics 365 users, this means: AI agents can independently qualify customer inquiries in CRM, prepare quotes, and schedule follow-up appointments. In the finance module, agents automatically reconcile invoices and only escalate exceptions to human staff. According to a recent analysis, 15 million people already use Microsoft Copilot, though only 3.3 percent pay for the premium version. Agent capabilities are likely to significantly increase this willingness to pay.

The strategic advantage of Microsoft’s approach: Agents work seamlessly across Microsoft 365, Dynamics 365, and Azure. A mid-sized company that already runs on Microsoft gets AI agents without additional infrastructure.

Oracle Fusion: 22 agentic apps in one release

On March 24, 2026, Oracle made the most aggressive move: Fusion Agentic Applications are an entirely new class of enterprise applications that rely not on individual copilots, but on coordinated teams of specialized AI agents. 22 agentic applications cover finance, HR, supply chain, and customer service.

Oracle’s approach differs fundamentally: Instead of bolting AI onto existing software, Oracle has embedded agents natively into the transaction system. The agents directly access enterprise data, workflows, approval hierarchies, and permissions. With AI Agent Studio, companies can additionally build their own agents without traditional application development.

Concrete use cases from the initial release: automated workforce scheduling and payroll optimization, AI-driven supplier sourcing for cost reduction, autonomous accounts receivable collection to improve working capital, and cross-selling agents that identify revenue potential with existing customers.

Oracle’s pace is notable: 22 production-ready agentic applications in a single release send a clear signal to the market. While SAP and Microsoft gradually activate individual agents, Oracle opts for a complete launch. For existing Oracle Fusion Cloud customers, this means immediate access. For the German mid-market, however, the advantage is relative: Oracle is far less prevalent in the DACH mid-market than SAP or Microsoft.

What does this mean for mid-sized companies?

Three developments are converging for mid-sized companies: The skills shortage is intensifying, business process complexity is increasing, and ERP vendors are delivering AI capabilities that go beyond gimmicks for the first time. AI agents in ERP are no longer a future prospect – they are available now.

The concrete benefits emerge in three areas. First: Routine relief. Agents take over tasks like invoice reconciliation, inventory checks, and scheduling that currently tie up qualified staff. Second: Faster decisions. Instead of waiting for the monthly report, agents deliver decision-making data in real time. Third: Fewer errors. Agents work rule-based and consistently – they never forget an approval step or overlook a price deviation.

At the same time, there are legitimate questions. How much autonomy does an agent get? Who is liable when an agent makes a wrong decision? And what happens to employees whose routine tasks disappear? These are questions that managing directors need to answer now, not once the agents are live. The EU AI Act already provides the regulatory framework for this.

A realistic scenario for a 300-employee mid-sized company: The accounts payable agent automatically matches incoming invoices against purchase orders, detects price deviations, and posts directly when they match. The inventory agent monitors critical materials and reorders before minimum stock levels are reached. The HR agent identifies open leave requests, checks cover arrangements, and approves according to policy. Three agents that together save an estimated two to three full-time positions of routine work. Not through layoffs, but by reallocating freed capacity to value-adding tasks.

How to prepare your ERP for AI agents

1. Audit data quality. AI agents are only as good as the data they work with. Clean up master data, remove duplicates, update process documentation. Without clean data, agents deliver wrong results.

2. Ensure cloud readiness. All three vendors deliver AI agents primarily through their cloud platforms. SAP Joule requires S/4HANA Cloud, Microsoft’s agents run on Dynamics 365, and Oracle relies on Fusion Cloud. On-premise systems are left out.

3. Define a pilot process. Don’t automate everything at once. Choose a clearly defined process – for example, invoice matching or inventory monitoring. Measure results, then scale.

4. Establish governance rules. Which decisions can an agent make independently? Above what amount does a human need to approve? Define these rules before rollout, not after.

5. Involve employees. AI agents don’t replace employees – they change their tasks. Administrators become process managers who steer agents and handle exceptions. Communicate this shift early and plan training.

The technology is ready. The question is whether internal prerequisites are in place. Companies that start with a pilot project now gain a 12 to 18-month head start over those waiting for the next licensing round.

Frequently Asked Questions

Title image source: Pexels / Jakub Zerdzicki (px:36598855)