AI agents in ERP systems: what SAP, Microsoft and Oracle are changing in 2026

In six months, the ERP vendors’ story changed in kind. SAP announced in January 2026 that Joule, its AI assistant, was merging with Microsoft Copilot. Oracle answered in March with 22 agent applications shipped at once. Microsoft turned Copilot Studio into a platform for autonomous agents in April. Three announcements, one message: AI will no longer be a product beside your ERP, it will live inside it.

“40% of enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025.” Source: Gartner, August 2025

Why are AI agents in ERP systems arriving so fast?

Because the three major vendors shipped their agentic layers almost simultaneously, and Gartner predicts that 40% of enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025. The pace of the last few months is unheard of: SAP Joule merged with Microsoft Copilot in January 2026, Oracle shipped 22 agent applications in March, Copilot Studio became a platform for autonomous agents in April, as recapped by My Business Future.

The delivery model matters as much as the technology. Joule agents come with the S/4HANA Cloud licence. Microsoft agents ship through Dynamics 365 and Copilot Studio. Oracle’s Fusion Agentic Applications activate for existing Fusion Cloud customers. In other words, many organizations will receive AI agents without having asked for them. The question lands on the desk of IT and finance leaders before any budget is voted.

What is an AI agent in an ERP, concretely?

A copilot answers your questions; an agent acts: it watches transactional data, makes decisions inside defined rules and executes actions in the system. To situate the two families of technology involved, see our article on the difference between traditional and generative AI.

Examples from the first releases: a supply agent that spots an imminent stockout, checks alternate suppliers and drafts a purchase proposal; a finance agent that matches invoices to purchase orders and only escalates discrepancies; an HR agent that applies replacement rules before approving a leave request. Nothing spectacular. That is precisely the point: agents target the repetitive work that ties up qualified people today.

What are the risks of turning on agents over poorly controlled processes?

An autonomous agent wired to stale master data or a fuzzy process accelerates errors instead of fixing them. Three questions to settle before switching anything on: which decisions the agent can make alone, above which amount a human confirms, and who answers when the agent gets it wrong. Approval hierarchies and audit trails already exist in your ERP; agents must inherit them, not bypass them.

There is also a commercial angle to keep in mind. Joule agents require S/4HANA Cloud: ECC and on-premise customers do not get them, which adds very real migration pressure. The vendor sells the agent, and the certified integrator sells the migration. To separate the useful from the forced, an independent perspective helps; we have already explained why an integrator certified by a vendor cannot advise you neutrally.

How do you prepare your ERP for AI agents without rebuilding everything?

By treating the arrival of agents as a process project: a bounded scope, clean master data, written governance rules and a measured pilot before any scale-up. The approach fits in five points:

• inventory candidate processes and pick one bounded case, such as invoice matching or inventory monitoring;
• validate master data quality: duplicates, inactive suppliers, up-to-date approval rules;
• write the governance rules before the rollout, not after;
• measure a before-and-after: exception rates, delays, hours recovered;
• train the teams: managers become pilots of agents and their exceptions.

At PlanAxion, that is exactly what a scoping workshop does: align stakeholders, inventory use cases, prioritize them by value, complexity and data maturity, then leave with a defensible roadmap.

Which process should you start with?

AI agents in ERP systems are no longer conference talk: they are in the licences you already pay for, or in the ones you will be offered at the next renewal. The choice that remains in your hands is which process you hand them first. Pick one that hurts every month, measure it, govern it, then decide the next step with numbers.

Frequently asked questions

What is the difference between a copilot and an AI agent in an ERP?

A copilot answers questions and suggests actions: it waits for a request. An AI agent acts on its own inside defined rules: it detects a situation, makes a decision and executes the action in the ERP, then escalates uncertain cases to a human.

Do you need to migrate to the cloud to get AI agents in your ERP?

Often, yes. SAP Joule agents require S/4HANA Cloud, Microsoft agents ship through Dynamics 365, and Oracle delivers them in Fusion Cloud. On-premise systems are generally left out. The migration should still be decided on your business case, not under licensing pressure alone.

Will AI agents replace finance or procurement teams?

No. The first use cases target repetitive work: invoice matching, inventory monitoring, routine approvals. Teams shift to exceptions, agent oversight and analysis. Human judgment remains required, especially for decisions above approval thresholds.

Where should you start with AI agents in your ERP?

With a bounded, measurable pilot on a single process, with validated master data and written governance rules. Measure a before-and-after, then scale if the numbers hold. A scoping workshop is often enough to prioritize the use cases.