Key Takeaways

  • OpenAI tested third-party Model Context Protocol (MCP) connectors from prominent data providers and uncovered error rates exceeding 40 percent.
  • Institutional enterprise accounts already held direct data licenses, making bundled data appear economically redundant on first principles.
  • Kim chose to bundle quarterly earnings transcripts directly into ChatGPT for Financial Services to control data quality from end to end.
  • Bundling data increased the software price, but institutional users cited reliable, built-in data access as their favorite feature.

External Connectors Failed at a 40 Percent Rate

When Joseph Kim stepped in to lead ChatGPT for Financial Services at OpenAI, the obvious architecture seemed clear: rely on external connections. The Model Context Protocol was expanding, connectors were appearing in public directories, and financial institutions already spent millions on proprietary market data licenses. Why duplicate data pipelines when customers could simply link their existing accounts?

Conventional platform theory argues for thin middleware. You provide the reasoning engine, and the customer attaches their own external pipes. But Kim tested that assumption against reality. His team ran systematic tests across third-party financial MCPs, including connectors built by established brands. The infrastructure buckled immediately.

“The reality is we went and we tested all these MCPs,” Kim said. “And what we found was the error rates were really really high. Some of the MCPs that we connected, some big names, 40% plus error rate on their MCPs, right?”

In quantitative and investment workflows, a tool that drops four out of ten queries is useless. If an analyst pulls a quarterly earnings call transcript to check guidance on revenue, an API timeout or schema failure does not look like a connector issue. It looks like the AI hallucinated or broke. The model takes the blame for the network boundary.

Direct Control Beats Modular Convenience

Choosing to bundle data created an immediate commercial hurdle. Enterprise buyers are notoriously stingy about paying twice for information they already license. Kim anticipated the skepticism from procurement teams.

“The first question that the first thing people are going to say is, 'Look, there's an MCP connection in the store. Like what is the incremental value that you're providing by bundling that in?'” Kim said.

The answer lay in product accountability. When an MCP connector fails silently, breaks on pagination, or mangles financial tables, the software company cannot fix it without waiting on an external vendor. The only way to guarantee institutional uptime was to own the data layer directly.

“The only way we can go about fixing this at scale is to bring the data in house, like just control that experience end to end,” Kim noted. “And the extra money that you pay on top to bundle that data is worth it, right? And so that was a product decision that we had to make that was very industry specific.”

The financial services edition absorbed the data cost and passed it through to the subscription. Buyers did not revolt. Instead, the reliability of the built-in transcripts became the product's strongest selling point. As Kim put it: “Customers love it. The number one thing they call out is the data. And so it worked.”

What to Do With This

Audit your application's external connectors by running a stress test of 200 concurrent requests against each live integration tomorrow morning. If any third-party connector posts an error rate above 3 percent, pause user interface work and bring that specific dataset directly into your own database. Your users do not distinguish between your software and your data vendor's downtime.