Why structured workflows beat one-shot prompting on hard reasoning
Most people waiting for AI to get good at investing are waiting for the wrong thing. They are waiting for a bigger model, the next release, the version that finally reasons well enough to be trusted with money. The research points somewhere less convenient. The difference between an AI that produces real analysis and one that produces confident filler is rarely the model. It is the process the model is asked to follow.
That claim is easy to make and easy to wave away, so it is worth grounding in what the studies actually measured rather than what the headlines made of them.
What the Chicago study actually found
In 2024, researchers at the University of Chicago's Booth School (Kim, Muhn, and Nikolaev, in "Financial Statement Analysis with Large Language Models") handed a general-purpose model nothing but standardized, anonymized financial statements, stripped of company names and narrative, and asked it to predict whether earnings would rise or fall. With a structured, step-by-step prompt built to imitate how an analyst reasons, the model reached roughly 60% accuracy on the direction of future earnings. That beat the median human analyst, and it landed on par with a machine-learning model built and trained specifically for the task.
The detail that matters is not the score. It is what produced it. The model was not fine-tuned for finance. It was not fed insider context. The unlock was the prompt structure, the instruction to work through the problem the way someone trained in the discipline would, one step at a time. Give the same model the same statements and ask for a verdict cold, and you do not get analyst-grade reasoning. You get a guess wearing the costume of one. The capability was latent the whole time. The structure is what drew it out.
Where the same models fall apart
Now run it the other way. In 2026 a team at Columbia Business School built WorkstreamBench, a test of AI agents on end-to-end financial spreadsheet work, the modeling and scenario tasks that fill an analyst's actual day. The strongest agents handled the easy tasks at reasonable quality. Then they degraded sharply the moment the work required more than a few chained calculations, and they fell short of professional standards on anything genuinely hard.
Read those two findings together and the picture is clear. A capable model on a tight, well-structured task performs like a professional. The same model, turned loose on a long and unstructured one, collapses as soon as the chain of steps grows. The variable is not intelligence. It is whether the work was broken into steps the model could actually hold. This is the same pattern the broader research on chain-of-thought and task decomposition has shown for years: a model reasons far better when a problem is split into an ordered sequence than when it is asked to solve the whole thing in one pass.
What this means for how you use AI
The practical lesson is not flattering to the way most people use these tools. Opening a chat and typing "is this stock a buy" puts the model in exactly the condition where it fails. One pass, no decomposition, no enforced order, a long and unstructured request that asks for the conclusion before any of the steps that would justify it. The model obliges, because obliging is what it does, and it hands back something fluent and unearned.
The condition where it succeeds is the opposite. A defined sequence. One question at a time. The business before the valuation, the bear case after the bull case, each step carrying forward what the last one established. None of that needs a smarter model. It needs a process imposed from the outside, because the model will not impose one on itself. That is the whole reason a structured research workflow beats a brilliant model used carelessly, and it is why the gap between good and bad AI research is mostly a gap in method.
The bigger model will come, and it will help at the margin. It will not rescue a process that was never there. The investors getting real work out of these tools are not the ones with access to some better model. They are the ones who stopped asking the model to think for them and started giving it a structure to think inside. The research has been saying this for two years. The market is slow to hear it because it is a duller answer than waiting for the machine that finally does the job for you. The edge was never the model. It was the scaffolding around it.
Sources:
