The Agents Have Started Talking to Each Other. Microsoft Copilot and SAP Joule Can Now Hand Work Back and Forth Across Your ERP.

For two years the story in enterprise AI was one agent reaching into one system. The newer, bigger story is agents from different vendors reaching for each other. At SAP Sapphire, Microsoft and SAP announced agent-to-agent (A2A) integration between Microsoft 365 Copilot and SAP Joule, letting the two coordinate across a single workflow instead of sitting in separate windows.

What was announced

Per Microsoft's own Azure blog, the integration delivers connected AI experiences using agent-to-agent capabilities, so Copilot and Joule can execute agentic flows with the combined context of Microsoft Work IQ and SAP's Knowledge Graph. Microsoft's example is deliberately ordinary: an employee preparing for a performance review starts in Microsoft 365 Copilot inside Word, draws on SAP-delivered Joule skills for systems like SAP S/4HANA or SuccessFactors, and schedules the one-on-one with their manager, all in the same place. Joule, for its part, is built to speak both the open Agent2Agent protocol and Model Context Protocol to draw on tools and third-party agents.

Why this is more than a Copilot that reaches into an ERP

Yesterday's model was a single assistant with a connector, like the Dynamics 365 plugin that lets Copilot Cowork read and write Finance and Operations. A2A is a step beyond that. It is one agent handing a sub-task to another agent that lives in a different vendor's world, getting an answer or an action back, and carrying on. Where MCP standardised how an agent reaches a system, A2A standardises how one agent reaches another. The ERP stops being a place you visit and becomes a service other agents call.

The part that should give you pause

This is genuinely useful, and it inherits a problem. When a Copilot agent can hand a task to a Joule agent that acts on SAP data, the question of what each agent is allowed to touch no longer lives inside one vendor's permission model. It spans two. The lesson from this month's AI agents breaking out of their test environments is not abstract here. Federating agents across systems multiplies both the value and the blast radius. Cross-vendor handoffs need cross-vendor guardrails, and those are younger than the demos.

The layer this still does not touch

Here is the strategic read, and it is the same one as with Copilot Cowork. Microsoft and SAP are solving how agents connect and hand off. Neither is solving what decision the agent should make once it is in your data. Knowing that a Copilot agent can ask a Joule agent to pull a supplier record is not the same as knowing which supplier to move up, which SKU to slot closer, or how to shorten a pick path. That decision layer rides on top of all this connective tissue, across every vendor's agent, and it is exactly the layer an optimization engine like Cognilium is built to be. The agents can talk now. What they should say is still the open question.