Joe O'Donnell, a former Tiger Global short portfolio manager, argues domain judgment is the decisive ingredient in building AI for public markets investors.
Joe O'Donnell says firms should default to building their own AI investment systems, since codifying an investment process takes years of fine-tuning, data work and analyst review.
Joe O'Donnell predicts routine financial modelling will be solved within twenty four months, while Canary builds its own transaction dataset and partners with vendors where possible.
Why Judgment, Not Generic AI, Drives Investing Results
Good AI for public markets depends on the judgment of the best investors, not on generic models or vendors claiming AI for financial services. Most investors lose value and returns come from the top few percent, so a system has to encode that skill through proprietary data, fine-tuned models and years of manual expert work. Funds that treat AI as a chatbot subscription leave alpha on the table.
Takeaways
Domain judgment decides whether investing AI works. Joe O'Donnell says the average investor destroys value and results come from the top few percent. A Buffett or Druckenmiller would beat a thousand juniors. Good AI therefore needs builders who did the job at an extremely high level for years.
Generic models lack the last accuracy percentiles. Canary fine-tunes its own models for performance, not cost, because off-the-shelf models lack the last few percentiles of accuracy. Some new models add nothing or degrade its financial work. Generic tools are best used for information gathering.
Canary stacks proprietary data, intelligence and vertical tools. Canary runs hundreds of microservices over proprietary niche datasets that yield a few valuable insights a year from millions of data points. Vertical tools sit on top, such as an AI forensic accountant covering about 10,000 companies.
Building in-house is possible but slow. Automation gets 90 percent of the way easily, and the last 10 percent takes five times as long. Fine-tuning needs hundreds of expert-labelled examples and thousands of hours. Joe defaults to doing it yourself unless an outside model is proven.
AI judgment stays lossy without underlying layers. AI judgment is lossy without data and intelligence layers beneath it, and reading an earnings call or transcript remains the least lossy method. Canary's idea generation agents turn a list of 14 companies into zero to three finished reports.
Routine financial modelling will be solved within two years. Joe would be surprised if AI forecasting of financial models were not solved within twenty four months. Brett notes AI balances balance sheets for banking but almost never forecasts. Much bespoke analysis will be abstracted away.
Firms will organize around judgment, not analyst pyramids. Joe expects firms built around people with good judgment, cheaper fund launches and more small hedge funds at scales like $25m, whose economics are poor today. He suggests alpha may stay high because so many investors misuse AI.
Starting AI adoption at a fund
Decide which judgment tasks stay in-house before buying any tools.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Integrate your data and learn the limits of the models.
Hire someone who can teach the real capabilities and limitations of the models.
Do not start with a naive enterprise chatbot subscription.
Expect enterprise versions of models such as Claude to become the operating system, with Canary plugging in through MCP.