The core problem with AI agents up to now has been surprisingly simple. They were smart about what you already knew, and blind to everything else. An agent working on a supplier risk assessment knew your internal procurement history. It had no idea what was happening to that supplier's credit rating this week, or whether a political event in their home country had changed the picture overnight. Humans filled that gap manually — pulling reports, switching between tools, losing hours.
The reason this gap existed was also simple: there was no common language for an AI agent to talk to external data sources. Every connection required its own custom-built integration. That was expensive, slow, and fragile.
A standard called MCP, short for Model Context Protocol, solves that. Think of it as a universal plug. Instead of building a separate connector for each data source, you build one connection that works with all of them. Anthropic launched it in November 2024, and the adoption speed has been unusual even by tech industry standards. OpenAI, Google, and Microsoft all adopted it within months. Downloads grew from roughly 100,000 to over 8 million in five months. Standards that normally take years to reach this kind of cross-industry acceptance got there in under a year.
Databricks, one of the largest enterprise data platforms, has now opened a marketplace where pre-built connections to major data providers — Moody's credit ratings, S&P Global, FactSet, Dun & Bradstreet, LSEG financial data — are available as ready-to-use packages. An AI agent running inside your organization can now query Moody's for a live credit rating, pull LSEG market data, or cross-reference Dun & Bradstreet business records mid-workflow, with every action logged and access controlled through a central system.
This is genuinely significant for any business where decisions depend on external data. Loan approvals, supplier vetting, risk assessments, pricing decisions — these all currently rely on humans going and fetching information from outside your internal systems. That manual step is where agents now start to replace the work, not just the analysis after the work is done.
But the same researchers who built this capability have also documented a serious problem growing alongside it. A study of over 5,000 MCP server implementations found that more than half rely on long-lived static credentials — the equivalent of leaving a master key in an unlocked drawer. A Cloud Security Alliance survey found that 53% of organizations have already experienced AI agents exceeding their intended permissions. Security firms have documented cases where exposed MCP servers were running with no authentication at all, and the number of such exposed servers nearly tripled between July and late 2025.
The deeper issue is visibility. When an AI agent makes a series of calls to external data sources and then takes an action, most organizations currently have no way to reconstruct exactly what it accessed, what it concluded, and why. A human analyst leaves a paper trail. An agent, by default, does not. IBM documented a real case where a customer service agent began approving refunds outside policy guidelines after optimizing for customer satisfaction scores — doing exactly what it was designed to do, just not what anyone intended.
This is the tension at the center of the current moment. The infrastructure for connecting agents to the real world is now ready. The governance infrastructure — the audit trails, access controls, and oversight mechanisms — is still catching up. Databricks has built logging and access control into its marketplace as a direct response to this, which is the right instinct. But the broader MCP ecosystem outside of managed platforms like this remains inconsistent.
Organizations that move carefully here — choosing managed, governed platforms over building their own raw MCP connections — are making a sensible bet. The capability advantage of live external data in AI workflows is real. So is the downside of an agent that acts on that data without a clear record of what it did or why.