Enterprise software has always been built around the assumption that there will be a human sitting at the keyboard. Every design decision, every workflow, every pricing model has been optimized for people interacting with interfaces. That assumption is collapsing faster than most vendors anticipated. Vendors who don’t adapt won’t just lose market share. They’ll become infrastructure.
AI agents don’t care about your UI. They don’t appreciate an elegant dashboard or a well-designed navigation menu. What they need is data access, API depth, and integration reliability. That’s a fundamental shift in what enterprise software is actually for. And the numbers suggest it’s happening faster than most vendors anticipated.
The moment is now
When Anthropic launched the Model Context Protocol (MCP) in November 2024, it recorded roughly 100,000 downloads in its first month, a solid start for a new integration standard. When OpenAI adopted MCP in March 2025, monthly downloads jumped to 22 million within weeks. A more than two-hundred-fold increase in a matter of days. That’s not the adoption curve of an interesting experiment. That’s a standard taking hold.
Several vendors moved quickly. Salesforce launched the Agentforce Sales app inside ChatGPT in open beta in December 2025. Block, Pfizer, and Cloudflare all started running agent-driven workflows in production. The theoretical became operational, and the competitive dynamics of enterprise software reset while some vendors were still debating whether agents were real.
Own the agent or feed it
The most instructive thing happening right now isn’t any single vendor move. It’s the contrast between how different vendors are positioning for an agent-driven world.
Salesforce built Agentforce in 2024 as a proprietary agent harness: running inside the Salesforce ecosystem, optimized for sales and customer service, priced as a premium product. The logic was clean. Own the orchestration layer, control the experience, deepen customer commitment. Then in 2025, they pivoted toward MCP. An Agentforce MCP client entered beta. Salesforce-hosted MCP servers went into beta at Dreamforce in October 2025, with general availability in 2026. The Agentforce Sales app went live inside ChatGPT in December 2025. This is textbook hedging: they want customers in Agentforce, but they know external agents are coming regardless, so they’re making sure Salesforce is in the workflow either way.
Contrast that with a different posture visible across several ERP and back-office vendors: positioning the system of record as the agent destination, not the agent source. The pitch is openness. Rather than racing to own the AI layer, these vendors are investing in clean APIs and MCP compatibility so external agents can call their systems without friction, whatever model or orchestration layer the customer chooses. They’re not betting on owning the agent. They’re betting on being indispensable to it.
Two coherent bets on the same future. One side believes they can own the orchestration layer. The other believes the durable value sits in being the cleanest, most accessible system of record. Both could be right. What matters is that neither is waiting to find out.
From sandbox to production
Adoption at the enterprise level matters because it removes the theoretical from the conversation.
Block built an internal AI agent called Goose, running on MCP, that orchestrates workflows across internal systems and connects data warehouses to internal service registries. Cloudflare built an internally governed MCP platform that lets teams expose internal resources to agents without creating security exposure, with governance built at the protocol layer rather than bolted on afterward. Enterprises across financial services and life sciences are using MCP servers for controlled agent access to sensitive data under the same architecture. This pattern keeps surfacing: agents at the orchestration layer, enterprise systems as the data and execution layer, humans reviewing outputs rather than executing workflows.
These aren’t proofs of concept. They’re production deployments at sophisticated organizations. That pattern is consistent enough to call it a new architecture for enterprise computing, not just a trend.
The UX moat is gone
If agents become the primary orchestration layer (and the current trajectory points there), competitive advantage in enterprise software shifts from interface quality to data quality and API comprehensiveness.
A cross-application agent coordinating work across CRM, ERP, and HCM systems isn’t choosing platforms based on how they look. It’s choosing them based on how reliably they expose their capabilities and how trustworthy their data is. That’s existential for vendors whose moats are built on UX. And it forces a rethink of pricing, go-to-market, and product strategy from the ground up. If the primary user of your platform might be an AI rather than a human, who are you selling to? How do you price when consumption is driven by agent calls rather than seat count? These aren’t hypothetical future problems. Forward-looking vendors are working through them right now.
Make the call before the market does
Vendors best positioned for this shift are investing in API depth, clean data models, and MCP compatibility, and making a clear call on whether they want to own the orchestration layer or be the best-in-class system agents call. The vendors most exposed are still competing primarily on interface, without a coherent answer to the agent question.
The deeper divide: which vendors become the nerve center of the agent era, and which become the infrastructure those agents run on? Both are viable businesses. But they require completely different product strategies, pricing models, and customer relationships.
The window to make that choice deliberately is closing. Vendors who wait for the market to decide will find the decision has already been made.