Generative AI for Analysts
By leveraging the 2023 integration of GenAI into FACTSET as a natural experiment, this study reveals that while the technology significantly enriches the breadth and timeliness of financial analysts' reports, it simultaneously exposes human attention as a new bottleneck, leading to reduced forecast accuracy under high information-processing demands despite the availability of superior underlying data.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of finance, professional analysts act as essential guides for investors. Their job is to take a massive, chaotic flood of financial data—earnings reports, market trends, and economic news—and distill it into clear, actionable advice. For decades, this work has been defined by a simple, human limitation: there is only so much information a person can read, understand, and synthesize in a single day. Even the most skilled analyst has a finite amount of attention and time. When the volume of data grows faster than the human mind can process it, the quality of the final advice can suffer. This is the central puzzle that a new study tackles: what happens when a powerful new tool arrives that can instantly gather and organize far more information than any human ever could?
The researchers, led by Jian Xue, Qian Zhang, and Wu Zhu, investigated this question by looking at a specific moment in late 2023. A major financial data company called FactSet introduced a new artificial intelligence system named Mercury. This tool was designed to help analysts find information faster, create charts automatically, and draft reports. Because FactSet made this tool available to its existing users without extra cost, the researchers could observe a natural experiment. They compared the research reports written by analysts using FactSet before the AI arrived with those written after. They did not just look at whether the reports were longer; they used advanced language models to read every word, table, and chart in thousands of reports to count exactly how many different sources of information were used, how many different topics were covered, and how many different methods of analysis were applied.
The results revealed a striking transformation in the work itself. After the AI tool was introduced, reports associated with FactSet became significantly richer. They included 26 percent more distinct sources of information, covered 24 percent broader range of topics, and utilized 21 percent more analytical methods. The reports also arrived faster, often published within three days of a company's earnings announcement. The AI had successfully removed the bottleneck of finding and organizing data. Analysts were no longer struggling to locate the right numbers; they were presenting a much wider and deeper view of the companies they covered.
However, the story takes a surprising turn when the researchers looked at the accuracy of the predictions. Despite the reports being richer and faster, the relative accuracy of the forecasts actually declined. The analysts using the AI tool were less precise in their earnings predictions than their peers who were not using the tool. This finding challenges the common assumption that more information always leads to better decisions. The researchers dug deeper to understand why this happened. They found that the drop in accuracy was not uniform; it was concentrated among analysts who were already under the most pressure. Those covering a large number of companies, working in small teams with little support, or trying to update forecasts for multiple firms on the same day saw the biggest decline in performance.
The study suggests that the problem was not the quality of the information the AI provided, but the human capacity to process it. When the AI tool handed analysts a much larger pile of data, it also handed them a much heavier cognitive load. The analysts had to verify, prioritize, and make sense of conflicting signals that they had never had to juggle before. For those with limited time and attention, this overload led to confusion and errors. To prove this, the researchers built a computer model that processed the exact same information found in the reports. Unlike the human analysts, the computer did not get tired or distracted. When the machine processed the new, richer data, its accuracy did not drop. This confirmed that the information itself was still good; the bottleneck had simply shifted from finding the data to understanding it.
The researchers also ruled out several other possibilities. They checked to see if the AI was generating nonsense or "hallucinating" facts, but found that the data in the reports remained just as reliable as before. They checked if the reports were just becoming longer and more repetitive, but found no increase in useless text. They even tested if the analysts were rushing the work to get it out faster, but the decline in accuracy was not limited to the fastest reports. The evidence pointed consistently to one conclusion: the tool had expanded the supply of information faster than the human mind could absorb it.
This finding offers a nuanced view of artificial intelligence in the workplace. It suggests that while AI can dramatically increase productivity by handling the heavy lifting of data collection, it does not automatically improve decision-making. In fact, if the human user is already stretched thin, adding more information can make things worse. The study implies that for AI to truly help, organizations may need to redesign how their employees work. Instead of just giving analysts more tools, they might need to give them more time, smaller workloads, or better systems to help them sort through the flood of new information. The value of the technology depends not just on what the machine can generate, but on the human capacity to make sense of it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.