Evidence · strength 1 of 5 · searched to 2 Nov 2026
Does AI make financial analysis better?
Not enough evidence yet to say whether AI makes financial analysis better. 2 of 3 studies found a clear gain in quality and accuracy and 1 a clear loss.
What a central-bank trial and studies of analysts and hedge funds found when finance professionals worked with generative AI: quality, speed and forecast accuracy.
VerdictNot enough evidence yet
No good evidence
ImpactMixed
2 of 3 studies found a clear gain in quality and accuracy and 1 a clear loss.
Evidence strength1 of 5
We have very low confidence in this estimate.
Cost per user a year?
From general assistants to data-platform add-ins; prices vary too much to band.
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Quality (headline)
Mixed- Marsal and Perkowski (2025): +39% Blind-graded quality score (0-10 rubric) on all tasks in the first block of seven (generalist and specialist pooled), between-subjects design, AI access v no access. The generalist-only between-subjects estimate is +36% (2.318/6.421, Table 4 col. 2) (2.304 points (SE 0.205), p < 0.001 (Table 3, col. 2)).
scored 2.3 points higher than those who did not have access in the first seven tasks, representing an average increase of 39 percent relative to the control mean.
- Xue et al. (2026): −0.16 SD Relative earnings-forecast accuracy (scaled by the standard deviation of peer forecast errors), FactSet x Post (t = -2.96; significant at the 1% level (Table 8, col. 2)).
post-integration FACTSET-associated reports exhibit 0.157 standard deviations lower relative forecast accuracy than the comparison group.
- Sheng et al. (2025): +2 percentage points of annualised abnormal return Risk-adjusted portfolio return (alpha) of hedge funds whose trades track ChatGPT-generated signals, interquartile increase in 'GenAI Reliance', post v pre ChatGPT (difference-in-differences); lower end of the stated 2-4% range (FF4-alpha coefficient on GenAI Reliance x Post GPT 2.537, t = 4.46 (Table 3 Panel B, col. 6); range 45-87 bps a quarter across factor models).
An interquartile increase in GenAI Reliance (0.321) is associated with a 45-bps to 87-bps increase in quarterly risk-adjusted returns, depending on factor models used, and equivalent to 2%-4% annual alphas.
Output
Moderate gain- Xue et al. (2026): +26% Distinct information sources per analyst report, FactSet-associated v other reports, after v before the GenAI integration (difference-in-differences) (Sources are in Table 4, not Table 6 (Table 6 is analytical methods). Table 4 reports components only: text sources +2.273 (t = 4.73), figure sources +0.977 (t = 4.87), table sources +1.690 (t = 4.76). The 26% total is the abstract's summary and has no single coefficient).
FACTSET-associated reports become markedly richer--featuring 26% more distinct information sources, 24% broader topical coverage, and 21% more analytical methods
Time
Moderate gain- Marsal and Perkowski (2025): −23% Minutes to complete the bundle of generalist tasks, between-subjects design (control mean 31.2 minutes) (-6.951 minutes (SE 0.990), p < 0.001 (Table 3, col. 4)).
this represents an efficiency of improvement of about 23 percent relative to the control mean.
- Checking time not measured: none of these studies counts the time spent checking the AI’s output separately.
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level2 of 5, from 3 studies graded A, B or C.
- Down one: old models0 of 3 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Flag (no deduction)Mostly working papers or preprints, not yet peer-reviewed (flagged, not deducted).
- Publication biasnot testable with fewer than 10 studies.
- Result1 of 5. We have very low confidence in this estimate.
Behind the average
- Lower- v higher-skilled employees (Marsal and Perkowski (2025)): Lower-skilled staff gained most in quality on generalist tasks; higher-skilled staff gained most in efficiency; the equalising effect did not hold on specialist tasks
- Participants slower with AI (within-person comparison) (Marsal and Perkowski (2025)): 13 percent of participants were less efficient when they had the tool; 80 percent were more efficient
- Analysts covering more firms than the median (Xue et al. (2026)): Forecast accuracy fell by 0.174 SD (t = -2.40) for analysts with many firms, against +0.044 SD (not significant) for those with few; difference p = 0.029 (one-tailed)
- Analysts on busy days (revising two or more firms the same day) (Xue et al. (2026)): Accuracy fell by 0.205 SD (t = -2.97) on busy days against -0.066 SD (not significant) otherwise
- Asset managers that are not hedge funds (Sheng et al. (2025)): No clear benefit: weak evidence on FF4 alpha, and raw returns fell with adoption
- Small, young and passive hedge funds (Sheng et al. (2025)): No significant return gain; gains confined to large, older and more active funds (top v bottom quintile differences significant for size, age, turnover)
How, where and why it works
Why it works
1 of 4. AI drafts analysis and pulls in more sources; one study found more sources but less accurate forecasts.
For whom and where
2 of 4. Analysts covering many firms or working on busy days lost the most accuracy; lower-skilled central-bank staff gained most in quality.
What it takes
1 of 4. One trial used a single day of realistic tasks; the others infer AI use indirectly.
The full cost
0 of 4. No study counts cost.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Use AI to gather and summarise sources, then check every figure against the filing.
- Keep your own forecast before you look at the AI's.
If you are rolling it out
- Track forecast accuracy before and after rollout, against a comparison group.
- Check your financial regulator's rules on AI use.
What it takes
- Two of the three studies infer AI use indirectly rather than observing it.
What can go wrong
- Sell-side analysts' earnings forecasts became less accurate after an AI tool launched, most for analysts covering many firms.
- 13% of central-bank staff were slower with the tool in one trial.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| A Task-Based Approach to Generative AI: Evidence from a Field Experiment in Central Banking Ales Marsal (National Bank of Slovakia), Patryk Perkowski (Yeshiva University), 2025. National Bank of Slovakia Working Paper 13/2025 (working paper) | B | RCT, realistic task | 101 central bank employees completed both stages (160 completed the stage-one survey); 14 tasks each, AI access randomised by task block (within- and between-subject); between-subject estimates use the first seven tasks | OpenAI GPT-4o (May 2024) | none |
| Generative AI for Analysts Jian Xue (Tsinghua University), Qian Zhang (Xiamen University), Wu Zhu (Tsinghua University), 2026. arXiv:2512.19705v2 [q-fin.ST], September 2026 (preprint) | C | other quasi-experiment, deployed work | 46,853 report-firm pairs by US sell-side analysts, January 2022 to October 2024 (41,391 in the main regressions) | FactSet Mercury (LLMs integrated with FactSet's proprietary financial data) (Dec 2023) | none |
| Generative AI and Asset Management Jinfei Sheng (UC Irvine), Zheng Sun (UC Irvine), Baozhong Yang (Georgia State University), Alan Zhang (Iowa State University), 2025. Working paper, August 2025 version (working paper) | C | other quasi-experiment, deployed work | 644 hedge fund companies, 11,921 company-quarters (13F holdings), 2016 Q1 to 2024 Q3; survey of 45 funds (33 from the sample) in 2025 | ChatGPT (GPT-3.5), used by the authors to generate AI signals from earnings-call transcripts; the funds' own tools are not observed (Mar 2022) | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Studies we found but do not count
- Leveraging ChatGPT for Enhancing the Internal Audit Process: A Real-World Example from a Large Multinational Company (Emett et al. (2024), grade D). A case study: four team members' own estimates, with no comparison (section 4, rule 3).
These studies are real and often cited, but they fall outside our rules for this rating. What counts as a study.
Does the rating hold up?
- Run without company-run studiesNo change: no such studies.
- Run without working papers and preprints0 studies: strength 0, verdict no good evidence.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. Studies cover models released Mar 2022 to May 2024. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. This page is not yet in search engines: health, legal and finance pages wait for a named domain reviewer, and none is appointed yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.