Skip to main content
Backed by McGill Engine and 21st by CentraleSupélec Learn more
Book a call

Research hubAI in deal work

Agents draft a financial model in about an hour against 18 for an expert. They are far faster, they score lower, and they need review by someone who knows the work.

Published 1 November 2026Evidence to 1 November 202641 sources5 charts

Summary

AI is now common across finance and professional services, but most use is shallow and the value is hard to find in the accounts. Regulators and statistics offices report use by 86% of surveyed UK financial firms and 57% of Belgian professional services firms. Yet only 10% of UK businesses using AI call that use extensive, and 89% of senior executives in a four-country survey report no effect on labour productivity so far. Controlled studies show large gains on bounded tasks, and those gains shrink as the evidence moves from set tasks to firm accounts. On banker deliverables, agents are far faster than experts, score lower, fail in ways that look correct, and need review by someone who knows the work. For an independent firm, the decisions that matter are which tasks to give AI and who checks the result.

Key findings

  1. 0%

    Bankers rated 0% of the best model's outputs "client-ready". In the same benchmark, 21 bankers gave 172 ratings of work from two leading models and none was rated "sendable as-is". On a score calibrated to those ratings, the best model passed 16% of 100 tasks. BankerToolBench, April 2026, by authors including Handshake AI and others. Handshake AI sells AI training data.S24

  2. 89%

    of executives report no effect on labour productivity over three years, although 69% of their firms use AI. Self-reported, nearly 6,000 executives in four countries, November 2025 to January 2026.S21

  3. 93% to 49%

    is the fall in top score from earnings analysis to precedent transactions. Comparables reach 63.9% and financial modelling 57.0%. Benchmark by Vals AI, which sells AI evaluation, 29 October 2026.S26

  4. 19 points

    is how much less likely consultants using GPT-4 were to get a task right when it sat beyond AI's ability. On tasks within its ability they worked 25.1% faster with over 40% higher quality. Field experiment, 758 consultants, 2023 model.S17

  5. 18.3 hours

    against under 1 hour is expert time against agent time on the same ten financial models. Average scores were 77.2% for experts, 70.9% and 61.8% for the two agents. Benchmark by Kensho, S&P Global and MIT, April 2026. Kensho and S&P Global sell data.S28

Analysis

How widely is AI really used in deal and advisory work?

No official series covers independent advisory firms, so the nearest figures come from neighbouring groups. In UK financial services, 86% of full-survey respondents used AI in 2026, up from 75% in 2024, and 73% of smaller firms on the short form did soS1. These firms chose to respond, and the Bank says the results do not describe the whole sector. In Belgium, 57.1% of professional, scientific and technical enterprises with ten or more staff used AI in 2025, against 40.4% across the EUS3. The same broad UK sector stood at 45.8% in June 2026S2. In Canada, 32.4% of professional services businesses and 40.4% of finance and insurance businesses reported use in the second quarter of 2026S4.

Dealmaker surveys disagree. Bain reports 45% of M&A practitioners using AI, up from 21% a year earlierS7, S8. Deloitte reports 90% of US corporate and private equity leadersS9. Most of the gap is definition. Deloitte's 90% counts pilots and limited use, and only 37% use AI across several deal stages, close to the third that Bain calls systematic users.

Depth is thin. Only 10% of UK businesses using AI report extensive use, which implies about 3.5% of all businesses with ten or more staff (our estimate: 34.9% times 10%)S2. In UK financial firms, 9% of respondents have agents in corporate finance for decision support, 3% with constrained execution, and none with more autonomyS1.

Chart 1

Reported AI use has doubled or tripled in each home market in two to three years

Businesses reporting AI use, %. Compare slopes, not levels.

Line chart of reported AI use. Belgian professional services rose from 28.3% in 2023 to 57.1% in 2025. The UK rose from 11.9% in 2023 to 34.9% in 2026, and Canada from 6.1% in 2024 to 19.2% in 2026.

  • Belgium, professional services
  • Belgium, all enterprises
  • UK, all businesses
  • Canada, all businesses
0%
20%
40%
60%57.1%34.5%34.9%19.2%2023202420252026

Each country asks a different question. Canada asks about producing goods or delivering services over 12 months. The UK series is businesses with 10 or more employees and excludes finance and insurance. Belgian figures are enterprises with 10 or more staff, excluding finance. UK points are June, except 2023 (September).

Source: Eurostat isoc_eb_ain2, ONS Figure 1 data, Statistics Canada.

See the data
Belgium, professional servicesBelgium, all enterprisesUK, all businessesCanada, all businesses
202328.3%13.8%11.9%Not available
202442.9%24.7%14.8%6.1%
202557.1%34.5%25.1%12.2%
2026Not availableNot available34.9%19.2%

Do controlled gains survive outside the lab?

Yes, but they get smaller as the test moves closer to the accounts.

Evidence typeStudy and sampleResult
Lab, online, set writing tasksNoy and Zhang, 453 professionalsS1640% less time, 18% higher quality
Field experiment, set tasksDell'Acqua et al., 758 consultantsS1725.1% faster, 12.2% more tasks, over 40% higher quality inside AI's range
Field trial, measured by telemetryDillon et al., 7,137 workers, 66 firmsS18Two fewer hours a week on email for users; no detected change in task mix
Firm accounts, instrumental variableAldasoro et al., 12,000+ non-financial firmsS194% higher labour productivity, mostly in medium and large firms
Administrative recordsHumlum and Vestergaard, 25,000 workers, DenmarkS20Effects on earnings and hours above 2% ruled out two years on
Self-reported by executivesYotzov et al., nearly 6,000S2189% report no productivity effect

These rows use different measures and cannot be subtracted. The BIS 4% and the Danish null do not conflict: one measures output per worker in non-financial firms, the other earnings and hours. Stanford's AI Index reaches the same reading, with gains largest in structured work and smaller in deeper reasoningS14.

Self-reports need care. In a randomised trial of 16 experienced developers, tasks took 19% longer with AI while the developers believed they were 20% fasterS22. A later METR study points to a speed-up, but METR warns that selection effects make its data unreliableS22. This was software, not banking, and the tools have since improved.

Where does AI output fail on financial work?

Failure follows task type. Reading is the strong point: on 495 annual reports, the best model scored 88.4% against a human baseline of 82.8%S32. Calculation across a table is weaker. Across ten model set-ups on synthetic private equity spreadsheets, accuracy fell from 89.1% on lookups to 19.6% on complex aggregation, and averaged 48.6% on the largest fileS30.

Building is weakest. In BankerToolBench, by authors including Handshake AI and others, the best model failed nearly half of its rubric criteria and cleared all critical criteria on 2% of tasksS24. Its failing steps were 41% code and formulas, 27% reasoning, 18% retrieval and 13% fabrication, such as hard-coded offer prices shown as derived. APEX-Agents, from Mercor, which sells expert data, found a top pass rate of 27.3% on banking tasksS25. MBABench-2.0, on spreadsheet construction, found a top pass rate of 16% and 59% numerical accuracy on harder tasksS29. That 16% is a separate test from the 16% in finding 1.

The costly errors look right. Answers that stayed the same across eight runs were still wrong 15% to 23% of the time on FinQA, using small open modelsS34. Nearly one correct answer in six in FinFIRST lacked a complete evidence chainS31. In a small test, nine of fourteen runs wrongly flagged 95% to 100% of clean statementsS35, so AI is not yet a dependable checker either.

No single hallucination rate exists. Rates ran from 0.08% to 13.0% on annual reports, against 2.8% for humansS32, and 17% to 33% for paid legal research tools in a 2024 preregistered studyS36. Tasks and definitions differ, which explains much of the spread.

Chart 2

The closer a task gets to building and judging, the lower the best score

Best score of any model on that task type, %

Bars of the best model score by task type, falling from 93.0% on earnings analysis to 63.9% on comparables, 57.0% on financial modelling and 49.3% on precedent transactions.

  • Reading and comparing
  • Deal analytics

Each bar is the category leader, so no single model reached all these scores. A failed core fact scores zero. Not comparable with Finance Agent v2. The grouping is ours.

Source: Vals AI, which sells AI evaluation. Finance Agent v2.1, 29 October 2026, 449 held-out test questions.

See the data
Value
Earnings analysis93.0%
General quantitative87.8%
Disclosure analysis83.3%
General qualitative79.2%
Adjustments74.7%
Market analysis71.0%
Comparables63.9%
Financial modelling57.0%
Precedent transactions49.3%

Chart 3

Agents finished in about an hour against 18 for an expert, and scored below the expert on seven or eight of ten models

Expert-graded score by task, %

Dot chart of ten financial model tasks. Experts averaged 77.2%, GPT-5.4 70.9% and Claude Opus 4.6 61.8%. Experts scored higher than GPT-5.4 on seven tasks and higher than Claude Opus 4.6 on eight.

  • Expert
  • GPT-5.4
  • Claude Opus 4.6
18.3
Expert, average hours
0.97
GPT-5.4, average hours
0.71
Claude Opus 4.6, average hours
30%50%70%90%

Graders judged one of ten expert models better rebuilt than fixed, against two of ten for GPT-5.4 and six of ten for Claude Opus 4.6. GPT-5.4 was about 19 times faster (our calculation: 18.3 divided by 0.97). These are April 2026 models.

Source: FrontierFinance, Kensho, S&P Global and MIT, 7 April 2026, Table 3. Kensho and S&P Global sell data.

See the data
ExpertGPT-5.4Claude Opus 4.6
M&A, Cisco-Splunk77.5%69.9%72.9%
M&A, J&J-Shockwave73.0%80.6%75.6%
3-statement, Merck67.6%73.8%36.5%
3-statement, Rackspace77.4%49.2%67.6%
LBO, Electronic Arts72.2%77.1%76.1%
LBO, Walgreens74.4%61.6%35.3%
Lender, Carnival79.8%75.8%71.5%
Lender, MGM76.0%71.6%62.5%
DCF, Autozone88.6%78.7%58.7%
DCF, Thor85.0%70.5%61.1%
Average77.2%70.9%61.8%

Why has value lagged adoption?

Individual gains have not reached the accounts. In McKinsey's 2026 survey, 80% say AI improved their own productivity, while 37% attribute any EBIT effect to it, unchanged on the yearS10. BCG finds 5% of firms achieving value at scale and 60% little or noneS13. Twelve of 50 large banks report a realised or projected returnS12, and 18% of professional services respondents know their firm tracks return on AI toolsS11. Consultancies sell this work, so read these as signals rather than measures.

Two causes have support. Saved time is absorbed: 85% of Danish chatbot users moved it to other tasksS20. Review is costly: of 129 current and former bankers, 78% cite too many errors, 66% cite heavy manual review, and 12% cite priceS24.

Where does banker time go?

Bankers, mostly analysts, reported 72 hours a week in Wall Street Oasis's 2024 survey, down from 82 in 2021 (531 unverified respondents)S38. The widely quoted 98-hour figure comes from 13 Goldman analysts in 2021S37. eFinancialCareers puts boutiques at 60.3 hours across all functions, and these are large namesS39.

In the APEX survey, run by Mercor among people on its talent marketplace, analysts put 47.2% of time on core work, and financial modelling is 23.5% of that, or about 11% of the week (our estimate: 47.2% times 23.5%). Meetings, messages and admin take 38.6%S25. Bankers in a separate survey put over half the week on deal execution and pitch workS24. A drafting tool touches only part of this.

Chart 4

Self-reported banker hours fell ten a week from 2021 to 2024 but stayed above 70

Average hours worked per week

Line chart of self-reported weekly banker hours: 82 in 2021, 78 in 2022, 75 in 2023 and 72 in 2024.

60
70
80
90
80 hours: weekly cap JPMorgan reportedly set in 202482722021202220232024

Self-reported and unverified. The 2024 sample is 531 and skews to analysts. Sample sizes for earlier years were not stated in the source.

Source: Wall Street Oasis, 2024 Investment Banking Working Conditions Survey, January 2025. Cap as reported by the New York Post, March 2026.

See the data
Average weekly hours
202182
202278
202375
202472

Chart 5

Analysts say under half their time goes to core work, and meetings, messages and admin take over a third

Share of working time, %

Stacked bars of working time. Investment banking analysts report 47.2% on core work, management consultants 43.5% and lawyers 50.0%. Meetings, communications and admin take 38.6% for analysts.

  • Core
  • Meetings
  • Comms
  • Admin
  • Learning
  • Non-productive
Investment banking analyst
Management consultant
Lawyer

Self-reported. 227 respondents from the Mercor talent marketplace with 10.8 years' average experience, so this is not a junior banker sample. The investment banking row has 58 respondents and sums to 100.1 from rounding.

Source: APEX-Agents, Mercor, Appendix B, Table 7, 23 February 2026. Mercor sells expert data.

See the data
CoreMeetingsCommsAdminLearningNon-productive
Investment banking analyst47.214.713.510.48.36.0
Management consultant43.518.714.710.17.15.9
Lawyer50.010.212.811.111.44.5

What does this mean for an independent firm?

  • Sort work by the evidence. Give AI reading, extraction and first-pass summaries. Treat comparables, precedents, adjustments and models as drafts.
  • Keep calculation in the spreadsheet. Use live formulas and tie each figure to a source. Hard-coded and unsupported numbers are the recurring faults.
  • Price review into the job. Graders judged two of ten GPT-5.4 models and six of ten Claude Opus 4.6 models better rebuilt than fixed, against one of ten from expertsS28. Measure hours to a sendable document, not to a first draft.
  • Put the reviewer above the tool. Across 106 experiments, human and AI pairs beat humans alone but fell short of the better of the two, and gains came where the human outperformed the AIS23.
  • Test on your own past deals. Most benchmarks use US public filings, not private companies or French and Dutch documents.
  • Do not assume the averages apply. BIS gains sit in medium and large firmsS19, and smaller organisations are flat at 22% scaling agents against 40% at large onesS10.

Limits

  • No boutique data. We found no study, survey or official series on independent advisory firms in Belgium, the UK or Canada. Every adoption figure here is a neighbouring group.
  • No trial of bankers with AI. Benchmarks test agents alone on set tasks. Nobody has published a controlled test of bankers using AI on live deals, so the net time saved after review is unknown.
  • US public company bias. Most benchmarks use SEC filings. Private company data rooms and non-English documents are mostly untested.
  • Publisher interest. Handshake AI and Mercor sell AI training data, Vals sells evaluation, Kensho and S&P Global sell data, Microsoft makes the tool in S18, and consultancies sell AI work. We used their data with that in mind.
  • Fast decay. Benchmark scores age within months. The April 2026 models in S24 and S28 have been replaced. In 2023, a leading model got 19% of basic filing questions right with shared retrievalS33. Tests differ, so the rise cannot be measured exactly.
  • Weak time and hours data. All hours figures are self-reported. None comes from timesheets, and all skew to large banks and North America.
  • Not comparable across countries. Each statistics office asks a different question, and the UK series leaves out finance.
  • Left out because unverified or unaudited. Bank-reported hours saved, vendor time-saving claims, Bain's case examples, and press reports of large cuts to analyst classes. Finance Agent v2 strict scores were dropped because its page gives both 50.88% and 55.79%. The 2025 Belgian rate for legal, accounting and consulting firms is not published in the Eurostat series.
  • Summaries. We read Yotzov et al. through the NBER Digest summary, not the full working paper. The 80-hour cap comes from a press report of earlier coverage, not a bank document.
  • Final check. On 1 November 2026 we reopened BankerToolBench, FrontierFinance, the NBER Digest and the Vals AI board. The 0%, 16%, 18.3 hours, 89%, 93% and 49% figures matched.

Sources

All sources below were opened, and read in full or in the form stated in Limits. Dates are publication dates unless marked.

Regulators and official statistics

  1. S1Bank of England and FCA, Artificial intelligence in UK financial services 2026, 26 Oct 2026. 196 responses, 105 on the full survey, 95 smaller firms. Taken: adoption, agent autonomy by business area, short-form results.
  2. S2Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 Jul 2026. Business Insights and Conditions Survey wave 159, 38,637 responses, 26.7% response rate, finance and insurance excluded. Taken: Figures 1 and 3 data files, extent of use. Professional services 45.8% is 17.3 plus 28.5 from Figure 3.
  3. S3Eurostat, dataset isoc_eb_ain2, updated 15 Jun 2026. Enterprises with 10 or more staff. Taken: Belgium and EU rates. Cross-checked with ec.europa.eu
  4. S4Statistics Canada, Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026, 11 Jun 2026. Stratified sample of 21,105, 9,251 responses, fieldwork 1 Apr to 6 May 2026. Taken: use by sector and year.
  5. S5Statistics Canada, Analysis on planned use of artificial intelligence by businesses in Canada, third quarter of 2026. Fieldwork 2 Jul to 6 Aug 2026. Taken: 54.8% of professional services businesses plan to use AI in the next 12 months. Background only.
  6. S6OSFI and FCAC, AI Uses and Risks at Federally Regulated Financial Institutions, 24 Sep 2024. Voluntary questionnaire, December 2023. Taken: about 50% use in 2023, 70% expected by 2026 (a forecast). Background only.

Adoption and value surveys

  1. S7Bain and Company, M&A Capability for a New Era, 27 Jan 2026. More than 300 M&A executives. Taken: 45% use, about one-third systematic. Bain's case examples were not used.
  2. S8Bain and Company, Generative AI in M&A, 4 Feb 2025. n=307. Taken: 21% use, 16% in 2023.
  3. S9Deloitte, 2026 Generative AI in M&A Pulse Study. and PDF deloitte.com 500 US private equity, portfolio company and corporate M&A leaders, online, 14 to 27 May 2026. Taken: 37%, 33% and 20% use bands, which sum to 90%.
  4. S10McKinsey, The state of AI in 2026, 25 Aug 2026. 1,719 online respondents in 97 nations, 4 May to 8 Jun 2026, weighted by GDP. Taken: 80% individual gain, 37% EBIT effect, 40% against 22% scaling agents.
  5. S11Thomson Reuters Institute, 2026 AI in Professional Services Report, 9 Feb 2026. More than 1,500 professionals in legal, tax, accounting, risk and government. Taken: 18% ROI tracking.
  6. S12Evident, AI Index for Banks: Key Findings, 6 Oct 2026. 50 large banks, outside-in, public data. Taken: 12 banks report a return, up from 8.
  7. S13BCG, The Widening AI Value Gap, Sep 2025. n=1,250 firms. Taken: 5% and 60%.
  8. S14Stanford HAI, AI Index Report 2026, Chapter 4: Economy, 2026. Review of other studies. Taken: gains are largest in structured work. Its 88% adoption figure equals McKinsey's 2025 resultS15, so we did not count it separately.
  9. S15McKinsey, The state of AI in 2025, Nov 2025. 1,993 participants, 25 Jun to 29 Jul 2025. Taken: 88% regular use, used only to check S14.

Controlled studies and measured effects

  1. S16Noy and Zhang, Science 381(6654), 14 Jul 2023. Preregistered online experiment, 453 college-educated professionals. Taken: 40% and 18%.
  2. S17Dell'Acqua et al., Organization Science, online 11 Mar 2026. and working paper mitsloan.mit.edu Field experiment, 758 BCG consultants, GPT-4, 18 tasks inside the frontier. Taken: 12.2%, 25.1%, over 40%, 19 points.
  3. S18Dillon, Jaffe, Immorlica and Stanton, Shifting Work Patterns with Generative AI, v4, 13 Nov 2025. Randomised field experiment, 7,137 workers, 66 firms, telemetry, six months. Some authors worked for Microsoft, which makes the tool. Earlier versions reported larger email savings, and we used v4. Taken: two hours, no task-mix shift.
  4. S19Aldasoro et al., BIS Working Paper 1325, 23 Jan 2026. EIBIS and ORBIS data, more than 12,000 non-financial EU and US firms, US peer adoption as instrument. Taken: 4%, medium and large firm concentration.
  5. S20Humlum and Vestergaard, NBER Working Paper 33777, revised Mar 2026. and PDF nber.org Surveys of 25,000 workers in 7,000 workplaces, 11 occupations, linked to Danish registers to Dec 2024. Taken: 2% bound, 85% reallocation.
  6. S21Yotzov et al., Firm Data on AI, NBER Working Paper 34836, summary in NBER Digest, 1 May 2026. Nearly 6,000 CEOs, CFOs and senior finance managers in the US, UK, Germany and Australia. Taken: 69%, 89%.
  7. S22METR, 10 Jul 2025. and 24 Feb 2026, metr.org Randomised trial, 16 developers, 246 issues; follow-up of 57 developers. Taken: 19% longer, 20% believed faster, selection warning. Interval on the first result: +2% to +39%.
  8. S23Vaccaro, Almaatouq and Malone, Nature Human Behaviour, 28 Oct 2024. Preregistered meta-analysis, 106 experiments, 370 effect sizes, studies from 2020 to June 2023. Taken: Hedges' g of minus 0.23 against the better of the two; plus 0.46 where humans beat AI.

Benchmarks on financial work (all lab tests of AI, not of people using AI)

  1. S24Lau et al., BankerToolBench, 13 Apr 2026. Authors including Handshake AI, which sells AI training data, and others. 100 tasks, 9 models, 502 bankers, 21 bankers gave 172 ratings; surveys of 129 and 193 bankers. Taken: 0%, 16%, 2%, failure modes, survey results.
  2. S25Vidgen et al., APEX-Agents, v3, 23 Feb 2026. Mercor, which sells expert data. 480 tasks, 160 in banking; survey of 227. Taken: 27.3%, time tables.
  3. S26Vals AI, Finance Agent v2.1, updated 29 Oct 2026. 449 held-out questions at second or third-year analyst level, 35 systems, three-model jury. Taken: category leaders.
  4. S27Vals AI, Finance Agent v2 (archived), 29 Oct 2026. Used to check method and to record that scores are not comparable with v2.1.
  5. S28Krumdick et al., FrontierFinance, 7 Apr 2026. Kensho, S&P Global, MIT. 25 tasks, 10 expert-graded with human baselines. Taken: Table 3.
  6. S29MBABench-2.0, Oct 2026. 101 spreadsheet construction tasks, more than 20 agents. Taken: 16%, 59%.
  7. S30FinSheet-Bench, Mar 2026. Synthetic private equity data, 24 files, 10 configurations, text input without code tools. Taken: 89.1%, 19.6%, 48.6%.
  8. S31Wang et al., FinFIRST, 21 Sep 2026. Ant Group's Inclusion AI, Peking University, CICC. 123 tasks, 15 configurations. Taken: 17.48% unsupported-correct rate (201 of 1,150).
  9. S32Can Open-Weight Models Compete on Financial Text Comprehension?, Aug 2026. 2,967 question sets, 495 annual reports, 20 models. Taken: 88.4%, 82.8%, hallucination rates.
  10. S33Islam et al., FinanceBench, Nov 2023. Patronus AI. 150 cases, human review. Taken: 19% correct with shared retrieval. Background for limits.
  11. S34Wang, Confidently Wrong, 13 Jul 2026. Single author; three small open models on FinQA. Taken: 15% to 23%.
  12. S35Panda, FinVerBench, 28 May 2026. Single author; 105 instances, 14 runs, one prompt design. Taken: false positive rates.
  13. S36Magesh et al., Journal of Empirical Legal Studies 22(2), June 2025; preprint May 2024. Preregistered evaluation of three commercial tools. Taken: 17% to 33%.

Hours and labour

  1. S37Goldman Sachs first-year analysts, Working Conditions Survey, Feb 2021. 13 respondents, self-reported, leaked deck. The headline says over 95 hours and the table says a mean of 98 since January. Press reports used 95.
  2. S38Wall Street Oasis, 2024 Investment Banking Working Conditions Survey, Jan 2025. 531 respondents, unverified, 8 Oct 2024 to 14 Jan 2025, 70.2% analysts. Taken: hours trend.
  3. S39eFinancialCareers, The banks with the best and worst working hours, 7 May 2025. More than 2,500 professionals, all functions. Boutiques include Evercore, Lazard and Moelis. Taken: 60.3 hours. The same article cites Wall Street Oasis at up to 83.8 for juniors, so the two are not comparable.
  4. S40New York Post, 20 Mar 2026, reporting earlier Wall Street Journal coverage. Press report. Taken: 80-hour cap.
  5. S41Revelio Labs, Banks Are Losing Interest in Junior Talent in the AI Era, 13 Oct 2026. US workforce data. Taken: junior bank headcount 4.7% below October 2022, with a later and much smaller fall in investment banking. Revelio reads this as a shift to digital banking, not AI replacing juniors. Background only.

Published 1 November 2026. Evidence to 1 November 2026.

To cite: Verdalyze, “Bankers rated 0% of the best model's outputs "client-ready"”, 1 November 2026, verdalyze.com/research/ai-in-deal-work

Book a call