Key Takeaways

  • OpenAI expects financial AI to shift from entry-level workflows to associate and managing director decisions by 2027.
  • Junior tasks succeeded first because feedback loops are immediate: an analyst spots a broken spreadsheet formula or an ugly presentation slide in seconds.
  • Senior finance tasks create an evaluation horizon problem, because validating an investment thesis can take three to five years of market returns.
  • To train models on strategic judgment, OpenAI bypasses flaky third-party data connectors and gathers direct feedback from top 0.1 percent industry experts.

The Short Feedback Loop of Junior Grunt Work

Most enterprise AI tools started in the basement of investment banks. They targeted the 80-hour workweeks of first-year analysts who spent nights aligning company logos and writing simple spreadsheet formulas.

Joseph Kim, lead of ChatGPT for Financial Services at OpenAI, notes that this target was practical rather than ambitious. “Up until date, the vast majority of applications of AI in finance have been really focused on I would say like the junior analyst's work,” Kim explains. “They really don't like building Excel models. It's a tedious kind of work and it's very detail oriented, it's easy to get wrong. And similarly with slides.”

These junior tasks share a defining property: instant, deterministic verification. If a Python script or an AI agent formats a pitch deck incorrectly, a human reviewer rejects it within five seconds. If a model generates cash flow projections with broken cell references, the error displays immediately. The cost of a bad output is ten minutes of analyst cleanup, making the risk profile low and the benchmark easy to score.

The Multi-Year Delay in Managing Director Tasks

The ceiling on junior automation is approaching fast. Kim projects that frontier models will soon climb the corporate hierarchy. “Going into 2027, we're going to start to move up the stack,” Kim says. “We're going to go from the task that a junior analyst can do to some of the tasks that an associate can do to some of the tasks that an MD would do and move up that chain.”

The problem is that senior work lacks objective scoring. When an AI generates an investment thesis, nobody knows if it is brilliant or disastrous on the day it ships. “As you move up the chain, what I would say is the objectives become much fuzzier,” Kim says. “It goes from this is clearly a good or bad slide, or something that we'd put in front of a client or not, to well, I don't know if this investment idea is good.”

A thesis requires years of operational execution and shifting macro conditions to validate. Because backtests fail to capture future market regimes, models cannot simply grade themselves against historical stock charts. OpenAI's approach requires training models directly on expert judgment. “What it takes is a lot of research focus. It takes a lot of wonderful data. It takes a lot of feedback from experts in the field,” Kim notes. “We really want to get the P99.9 type of feedback to make the model as good as humans.”

For builders, this means generic prompt wrappers around public APIs cannot compete on strategic workflows. If you want to automate senior judgment, you have to build review chains with elite practitioners and bundle clean, proprietary data rather than relying on brittle connectors.

What to Do With This

Audit your product roadmap and tag every automated feature by its verification speed. If a feature produces an output that takes your user more than five minutes to verify, pause development. Strip the feature back to junior-level tasks with instant visual or mathematical validation until you recruit verified domain experts to grade your model's reasoning.