For decades, the division of labor in investment research was straightforward: humans generated ideas, framed hypotheses, and took responsibility for risk; models handled computation, backtesting, and execution.
Large language models and AI agents are beginning to blur that boundary. The real question is no longer whether AI will participate in investing, but how far that participation will extend—and what humans will ultimately remain responsible for.
At GIM, we have been exploring these questions through CogAlpha, our agent-based investment research system designed to automate idea generation, factor construction, validation, and iterative improvement. This article shares some of the lessons we have learned along the way.
AI’s advantage is exploring combinations humans overlook
AI’s advantage is not necessarily discovering new principles. In one case, CogAlpha replaced the default sigmoid function with arctan while refining a quantitative factor. The mathematical tool was not new; what was novel was applying it in a place where researchers rarely think to use it. Subsequent testing showed that the modification improved performance.
The lesson was not that AI had invented a new theory. Rather, it was able to search beyond the defaults that human researchers tend to rely on.
Generating factors is becoming easier. Validation remains hard.
Many AI-generated factors perform exceptionally well in sample, only to fail once exposed to unseen data. Evolutionary search helps by producing thousands of candidates and allowing most of them to be discarded. But scale alone is not enough.
The next challenge is automating validation—testing robustness, interpretability, and economic intuition—and gradually feeding those lessons back into the system’s memory so future research becomes progressively more reliable.
The future investment stack may become layered
Quantitative models continue to generate stable signals. AI increasingly takes on the analytical work that once belonged to discretionary portfolio managers. Humans remain responsible for the final investment decision—and ultimately, for accountability.
Prompting is a form of capital allocation
Tokens are like chips on a table. The model determines how to reason, but humans still decide which questions deserve the computation in the first place.
AI is unlikely to reshape investment research in a smooth, linear progression. More likely, it will feel gradual—until suddenly, many parts of the research workflow look fundamentally different.