Does AI-enriched analyst reports improve forecast accuracy?

"Generative AI for Analysts" measures what happens to sell-side research when a domain-specific AI platform is bundled into an existing data product. The authors treat the December 2023 introduction of MERCURY, FactSet's GenAI platform, as a plausibly exogenous change in AI access. Across 46,853 report-firm pairs issued by U.S. analysts, FactSet-associated reports after the integration show 26% more distinct information sources, 24% broader topical coverage and 21% more analytical methods, and the abstract adds that they became more timely. The decision-quality result runs the other way: "relative forecast accuracy declines when analysts face greater information-processing demands." A machine-learning benchmark fed the same observable inputs shows no analogous decline, which the authors read as "a human processing constraint rather than poorer underlying information." Every richness figure is the authors' own measurement. They used GPT-4o-mini to code each report's text, so the numbers are coded from the reports, not reported by FactSet or by the analysts.

The mechanism the excerpt gives is attention rather than capability. The authors write that AI "does not magically resolve attention constraints"; when workload and information demands are high, "AI-generated information compounds their cognitive burden." The same tool that enriches reports under routine conditions "can become a source of distraction when attention is strained." The paper also says the strain reaches investors, who show weaker market reactions to AI-assisted analyst reports. The treatment is defined narrowly too. The authors observe FactSet citations, not analysts' prompts or clicks, so the estimate is the "incremental effect of exposure to FACTSET after its GenAI integration" rather than the effect of verified MERCURY use. They note this makes the estimate conservative if control reports also use other GenAI tools.

Against the nearest notes, this excerpt is a case where the output that looks better is not the output that predicts better. Can self-ratings replace objective performance scores for AI competence? argues that self-report cannot stand in for performance; here the coded richness measures rise while the objective forecast-accuracy measure falls, which is the same warning in a different form. Where have workers actually delegated tasks to AI? also places AI use in information-intensive work, but it measures adoption from agent skills, whereas this excerpt measures what happens to output once a tool sits on proprietary data. The bottleneck argument echoes What makes accountable judgment scarce when AI cognition is cheap?: when first-pass information is cheap, the scarce input is the analyst's processing capacity. Does generative AI shift knowledge workers away from communication? reports an individual-documentation tilt from workplace trace data; this excerpt measures the reports themselves.

The excerpt does not establish the decision-quality result on its own. The forecast-accuracy specification, the benchmark design and the placebo tests that make a platform-wide trend unlikely appear only as claims in the abstract and introduction; the excerpt ends before any accuracy table. The investor-reaction finding is asserted in the related-work discussion without its evidence. The design covers one platform and one population of analysts, and "relative" in "relative forecast accuracy" is never defined in the text provided. So the strength the evidence allows is narrower than the title's wording. The study supports that report richness and forecast accuracy diverge under heavier demands, and it makes human processing a candidate explanation, because the benchmark did not decline. The "human constraint" reading is an inference from that contrast, which the authors state as "pointing to." It does not show that analysts would be more accurate with less information.