The AI just beat the human gatekeepers at the one job they thought was sacred: knowing what a good deal looks like.

The Summary

The Signal

Terminal X built an evaluation agent that grades investment opportunities against a 54-criterion rubric. The new model, GPT-6.1 Sol, agrees with senior partner judgment 88.6% of the time. More importantly, it catches mandate drift — deals that technically pass but violate the fund's investment thesis — 25% more effectively than earlier versions.

This isn't about replacing analysts who build Excel models. Terminal X's Eval Agent is replacing the pattern recognition that comes from seeing 10,000 pitch decks. The kind of judgment that used to take a decade to develop.

"The bottleneck in institutional investing has shifted from information access to decision velocity at scale."

The 54-criterion rubric matters because it makes subjective judgment auditable. Here's what that means in practice:

  • Fund mandates are no longer fuzzy guidelines, they're executable code
  • Junior partners can move faster with AI backup that knows the firm's thesis better than some senior partners
  • Compliance and investment discipline become byproducts, not overhead

The 88.6% concordance rate is the real story. Not because it's perfect, but because it's past the threshold where institutions will trust it to route decisions. Below 85%, you're still checking everything. Above 85%, you're only reviewing the 11.4% where the AI flags uncertainty.

GPT-6.1 Sol is OpenAI's latest reasoning model. The "Sol" designation indicates enhanced chain-of-thought capabilities specifically tuned for multi-step evaluation tasks. In Terminal X's case, that means the model can hold a fund's entire investment mandate in working memory while evaluating whether a Series B SaaS company with 47% gross margins and a cybersecurity angle actually fits the "enterprise infrastructure" bucket or belongs in "security tools."

The Implication

Every institutional investor now faces a choice: use agents like this to get faster and more consistent, or compete against funds that do. The quality bar for "good enough to review" just shot up. Deals that would have made it to partner meetings six months ago won't clear the AI screen.

For founders, this cuts both ways. The agents are more consistent than humans, which means less randomness in who gets funded. But they're also less forgiving of narrative. If your metrics don't map cleanly to the rubric, you're out before a human sees the deck.

Sources

OpenAI Blog