SAP’s expansion from 40 to over 200 Joule AI agents between 2025 and 2026 signals how fast the enterprise agent market is compressing what used to be multi-year roadmaps into single-cycle decisions. Three forces are driving the surge: data fabric and pipeline infrastructure from vendors like Salesforce and Snowflake, MCP servers (the connective tissue letting agents hand tasks to one another across platforms), and proprietary agent-development tooling that vendors are building for themselves first. CIOs tracking AI agent deployment options<\/a> now face a selection problem, not a supply problem.<\/p>\n The organizations most exposed here aren’t the laggards, they’re the ones that ran fast pilots across multiple platforms without a selection framework. When SAP, Salesforce, and Appian are each offering dozens of overlapping agents, the procurement question stops being “which vendor has agents” and starts being “which agents are we actually accountable for.” If your enterprise doesn’t have a named owner for that accountability, the default answer is nobody, and shadow AI fills the gap quietly.<\/p>\n The MCP layer deserves specific attention. MCP servers, which act like a shared language letting AI agents from different vendors pass instructions and context between themselves, are where vendor lock-in will actually crystallize. Every major platform announcing MCP support sounds like interoperability, but the agent that sits at the top of a workflow, orchestrating the others, becomes the chokepoint. The vendor whose agent owns that orchestration role owns the relationship. That’s not a neutral infrastructure choice; it’s a strategic one dressed up as plumbing. The piece, written for a CIO-focused publication with a clear interest in positioning IT leadership as indispensable to these decisions, doesn’t dwell on this dynamic, but the numbers it cites make the case anyway.<\/p>\n The vendors building their own agent-development tools and deploying them internally first aren’t being generous when they later release those tools to customers. They’re shipping agents that already reflect the vendor’s workflow assumptions, optimized for their own integration partners. CIOs who rely exclusively on vendor-built agents will find their process automation shaped by someone else’s architectural opinions. That’s worth weighing at the next renewal, not after the agents are embedded.<\/p>\n MCP, or Model Context Protocol, is a standardized communication layer that lets AI agents built by different vendors exchange instructions, share data, and hand off tasks within a larger automated workflow, roughly analogous to how USB created a common plug standard so devices from any manufacturer could connect to any computer. In enterprise AI, MCP servers determine which agents can “talk” to which, making them the quiet infrastructure decision underneath every multi-agent deployment.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: MCP servers<\/h2>\n