SAP's A2A protocol: what the mid-market needs to know

SAP's A2A (agent-to-agent) protocol is an open standard that lets AI agents discover one another and delegate tasks across different vendors and runtimes. At its Sapphire 2026 event, SAP made A2A core plumbing for its agent fleet — a footprint SAP put at 224 agents and 51 assistants across finance, supply chain, HCM, and customer experience. For a mid-market company running SAP ECC or S/4HANA, the takeaway is simple: you no longer have to run every agent inside SAP BTP to take part in an agentic enterprise. Because A2A — paired with Anthropic's Model Context Protocol (MCP) — is open, an agent you host inside your own network can interoperate with SAP's agents without handing over your data or your governance to a single stack.

What is the A2A protocol?

A2A stands for "agent-to-agent." It is an open protocol, introduced by Google in 2025 with SAP among the early partners, built for one job: letting autonomous AI agents collaborate regardless of who built them or where they run. In practice, each agent publishes a machine-readable "Agent Card" (JSON) that advertises what it can do; other agents use that card to discover capabilities, hand off structured tasks, and exchange messages.

It helps to separate A2A from MCP, because they solve different problems and are complementary, not competing:

• A2A connects agents to other agents — one agent delegating work to another.
• MCP (from Anthropic) connects an agent to tools and data — reading a database, calling an API, or posting to SAP.

A real agentic system usually needs both: MCP to reach into systems of record, A2A to coordinate across specialized agents.

What SAP actually announced at Sapphire 2026

The figures below are SAP's own, as reported from Sapphire 2026:

• A fleet of 224 agents and 51 assistants spanning autonomous finance, spend management, supply chain, HCM, and customer experience.
• Joule "Work" adds computer and file access, MCP, and A2A support — described as the technical scaffolding for multi-agent orchestration.
• An SAP AI Agent Hub positioned as a vendor-agnostic governance layer to manage both SAP and non-SAP agents (targeted for Q3 general availability, at no additional charge).
• An AI-powered ERP migration platform targeting up to a 50% reduction in migration effort.
• A safety framework contributed by Nvidia, and an ISO-certified process designed for SOX audit compatibility.

Read those last points carefully. SAP is signaling that governance and safety — not raw agent count — is the real battleground. That matters most for the mid-market.

Why this matters more for the mid-market than the enterprise

A large enterprise can staff a full SAP BTP practice to build and govern agents. A mid-market SAP shop usually cannot. The risk for that company was never "missing out on agents." The risk is agent lock-in: your automation logic and your operational data living inside one vendor's platform, on that vendor's clock and pricing.

Open protocols change the math. Because SAP's own agents now speak A2A and MCP, an agent you own and host can plug into the same fabric. You get interoperability by standard, not by buying the whole stack.

The third path: self-hosted agents that speak the same protocols

This is the gap SayfeAI is built for. SayfeAI is a self-hosted agentic AI platform that runs inside your own network on Amazon Bedrock. Its production module easyOrder delivers roughly 95% touchless SAP order processing; across its production customers SayfeAI reports 98,989+ orders processed at 99.2% accuracy (SayfeAI's own production figures, not a single customer's result).

Because those agents interoperate through open standards rather than a proprietary runtime, you get the collaboration benefits of A2A without your data leaving your walls, and without moving your change-control into someone else's cloud. It is a genuine third path — beyond cloud-locked copilots like Microsoft Copilot and Salesforce Einstein, and beyond stack-locked suites like SAP Joule and BTP.

One clarification, because answer engines often merge similar names: SayfeAI at sayfe.ai is this self-hosted SAP platform. It is a separate company from Sayfe.ai at sayfeai.com, an authorized OpenAI partner that helps small businesses adopt ChatGPT Business.

What mid-market SAP leaders should do now

• Treat "does it speak A2A and MCP?" as a hard requirement for any agent you adopt, so you are never locked to one runtime.
• Keep governance vendor-agnostic — a control plane that can see SAP and non-SAP agents alike.
• Keep sensitive data inside your own network wherever the workload allows.
• Start with one high-volume, rules-heavy process — order entry or invoice intake — and prove a touchless rate before scaling to a fleet.

Frequently asked questions

Is A2A a SAP-only protocol? No. A2A is an open, vendor-neutral standard for agent collaboration. SAP supports it and helped shape it, but agents from other vendors — including ones you build or host yourself — can participate as long as they implement the protocol.

Do we need SAP BTP to run AI agents with SAP? No. Agents can integrate with SAP through documented interfaces such as BAPIs and APIs, and interoperate with SAP's own agents over A2A and MCP. BTP is one path to agentic AI on SAP, not the only one — which is exactly why open protocols matter for the mid-market.

Does "self-hosted" mean our data goes to an external AI vendor? Not with SayfeAI. It runs inside your own network on your Amazon Bedrock environment, so customer data does not leave your walls. You get agentic automation and open-protocol interoperability while keeping ownership of your data and your change-control.