AI in Finance Pays Off Where a Person Can Check the Work

AI in finance now codes transactions, matches payments and drafts reports. The strongest field study links it to a month-end close 7.5 days faster, while a 2023 benchmark found a leading model failing 81% of questions about company filings. Part of the gap is checkability, and the law leaves the check, and the signature, with a person.

For a founder or finance lead, two questions decide most of it: which outputs a person can check faster than produce, and who answers for the ones nobody checked. The table below gathers the dated evidence, study by study, with the limit of each.

AI in finance: what the dated evidence shows, and what it does not

Use	Evidence (date, sample)	Finding	What it does not show
Coding transactions, closing the books	Choi & Xie (2026): 277 accountants; 79 firms on one AI accounting platform	55% more weekly client support; close 7.5 days faster (working paper)	Cause and effect; firms outside that platform
Reconciliations, first drafts	PCAOB staff outreach (July 2024)	Companies use it for drafts and simpler reconciliations; secondary to customer-facing uses	How widely, or how accurately
Answering questions about filings	FinanceBench (Nov 2023): 150 cases, 2,400 answers reviewed	GPT-4-Turbo with retrieval wrong or silent on 81%	Today’s models
Accounting exams	Wood et al. (Jan 2023); Eulerich et al. (2024)	47.5% against students’ 76.7%; later 85.1% with GPT-4 and tools	Work on a company’s own books
Knowledge work	Dell’Acqua et al. (2026): 758 knowledge workers	12.2% more tasks inside the frontier; 19% less likely correct outside it	Finance tasks as such
Pay and hours of accountants and advisors	Humlum & Vestergaard (2025): Denmark, 11 occupations	Typical time saving 2.8%; effects on earnings and hours above 2% ruled out	Effects after 2024

Source: the studies named, as cited below; figures from each study’s published abstract unless marked.

AI in finance pays where a person can check the answer

The best evidence on the finance function itself comes from accounting. Jung Ho Choi and Chloe Xie combined a survey of 277 professional accountants with data from an AI-enabled accounting platform serving 79 small and medium-sized firms, covering more than 200,000 transaction records (Choi & Xie, 2026). In the working-paper version, adoption came with “a 55% increase in weekly client support” and “a 7.5-day reduction in monthly close time”. About 8.5% of accountant time moved from data entry to client communication and quality assurance, and ledgers became 12% more granular (Choi & Xie, working paper). The published paper reports these as associations, and the firms are one vendor’s customers, so the figures describe what adopters achieved.

The mechanism matters more than the headline. The accountants did not accept every suggestion: in the authors’ words, “expertise increases intervention when AI confidence scores signal uncertainty”. Coding a transaction is work a reviewer can verify against a bank line and a receipt, so the tool drafts and a person checks the doubtful cases. PCAOB staff heard a similar pattern from public companies in July 2024. Where they used generative AI in financial reporting at all, it was for first drafts and for “less complex and repetitive processes, such as preparing account reconciliations” (PCAOB staff, 2024). For a small finance team, the measured prize so far is the one Choi and Xie’s adopters reported: about a week off the close, earned on work whose answers can be looked up.

The same tools fail where checking costs as much as the work

The failures cluster on open questions. FinanceBench, a set of 10,231 questions about public companies, tested 16 model configurations on 150 cases in November 2023 and reviewed 2,400 answers by hand. GPT-4-Turbo with a retrieval system “incorrectly answered or refused to answer 81% of questions” (Islam et al., 2023). The models are three years old and the rate is dated. The shape of the task is the lasting point: checking an answer about a 10-K means finding the page the model should have read, which can take as long as answering.

A preregistered experiment with 758 knowledge workers, run with Boston Consulting Group, measured the same edge. On 18 tasks inside what the authors call a “jagged technology frontier”, participants using GPT-4 completed 12.2% more tasks, 25.1% faster and at higher quality. On one task chosen to sit outside it, they “were 19% less likely to produce correct solutions” than colleagues without AI (Dell’Acqua et al., 2026). Choi and Xie’s framed experiment found the accounting version: AI help raised classification accuracy on average, while leaning on AI recommendations that departed from the consensus could raise the error rate.

Deloitte supplied the costly version. A report it delivered to Australia’s Department of Employment and Workplace Relations contained fake citations and a made-up quote from a Federal Court judgment. The corrected version disclosed “a generative AI large language model (Azure OpenAI GPT-4o) based tool chain” (The Register, 2025), and Deloitte refunded more than A$97,000 of a contract worth about A$440,000, roughly 22% (our arithmetic; CFO Dive, 2025). A citation costs a model nothing to write and a reviewer real time to confirm.

Read together, these results suggest one question to put to any use: can a person trace the output to a source record (a bank line, an invoice, a page of a filing) in less time than it would take to produce it? Coding and matching pass. Narrative, estimates and uncited answers fail, and for those the review can cost as much as the work it replaces. The question is our synthesis; none of the studies poses it in these words.

Which common uses of AI in finance a person can check against a record

Task	Can a person trace the output to a record?	Where AI fits	Evidence
Coding bank and card transactions	Yes: the bank line and the receipt	Automate; review low-confidence items	Choi & Xie (2026)
Matching payments, reconciliations	Yes: both sides of the match	Automate; review exceptions	PCAOB staff (2024)
Answers about a contract or a 10-K	Only with a page citation	Draft with citations; check the page	FinanceBench (2023)
Variance commentary, board narrative	The figures yes, the causes no	Draft; a person writes the causes	Dell’Acqua et al. (2026)
Estimates and judgments (reserves, revenue timing)	No single record	A person decides and signs	SOX § 302; PCAOB staff
A vendor’s automation rate	Only with a stated definition	Ask for the intervention rate and its definition	SEC order on Presto (2025)

Source: CX Cash synthesis of the evidence cited in this article.

The law puts the check, and the signature, on a person

For a public company the person is named by statute. Under section 302 of the Sarbanes-Oxley Act, the principal executive and financial officers certify each annual and quarterly report, including that the financial statements “fairly present in all material respects the financial condition and results of operations of the issuer” (15 U.S.C. § 7241). The same section makes them responsible for internal controls. Wilfully certifying a report that falls short carries a fine of up to $5 million and up to 20 years in prison (18 U.S.C. § 1350). A private startup signs no such certificate. Its lenders and investors still rely on numbers that a named officer stands behind.

Auditors read their own duties the same way. Audit firms told PCAOB staff that an engagement team member who uses a generative AI tool “is still responsible for the results and documentation of the work”, and that reviewers apply the same diligence as for work done without it. At the time, the firms used it mainly for administrative work and research (PCAOB staff, 2024).

A tribunal has said so outside finance. Air Canada argued before British Columbia’s Civil Resolution Tribunal that it could not be held liable for what its website chatbot told a customer about bereavement fares, and lost in February 2024. The tribunal member was unmoved: “In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission.” The tribunal ordered Air Canada to pay $812.02 (Moffatt v. Air Canada, 2024 BCCRT 149).

A checked model can still fail, and the rules for generative AI are unwritten

Finance has met this problem before, in a spreadsheet. In the first half of 2012 the synthetic credit portfolio of JPMorgan’s Chief Investment Office lost approximately $5.8 billion. The bank’s own task force found that the value-at-risk model behind it “operated through a series of Excel spreadsheets, which had to be completed manually, by a process of copying and pasting data from one spreadsheet to another”. At one step “the spreadsheet divided by their sum instead of their average”, which likely muted volatility by a factor of two (JPMorgan Chase, 2013). Reviewers had called the process “error prone” and approved the model anyway. The report leaves the error’s exact effect unclear, so the formula cannot carry the loss on its own; what the report does show is a review that let it through.

Bank supervisors built their answer around “effective challenge”: critical analysis by objective experts with the expertise, independence and standing to force a change. On 17 April 2026 the Federal Reserve, the OCC and the FDIC replaced the 2011 guidance, SR 11-7, with SR 26-2, and the new text sets aside the tools discussed here: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance” (Federal Reserve, 2026). The regulators with the longest record of supervising models have chosen to wait, and tell banks to govern such tools under their own rules meanwhile.

Engineering described the next problem four decades ago. In a 1983 paper on process control, Lisanne Bainbridge observed that automation leaves people “the tasks which the designer cannot think how to automate”. She went further: “By taking away the easy parts of his task, automation can make the difficult parts of the human operator’s task more difficult” (Bainbridge, 1983). She studied control rooms, and no study cited here measures the effect in a finance team. The mechanism carries over all the same. A team that lets a model code every routine entry keeps only the exceptions, and its juniors stop practising on the easy cases that taught their seniors what a wrong entry looks like. The fit is loose on timing: a plant operator must act in seconds, while a reviewer of the books usually has days, so the risk is a missed error more than a slow reaction.

Adoption figures measure claims faster than outcomes

Words about AI in finance have spread faster than evidence of its effects. EDGAR full-text search finds the phrase “generative AI” in 2 annual reports on Form 10-K filed in 2022 and in 864 filed in 2025 (our count of filings whose main document contains the phrase; SEC EDGAR). In McKinsey’s survey of 102 CFOs, 44% used generative AI for more than five use cases in 2025, up from 7% a year earlier, although the survey counts use cases and gives no method (McKinsey, 2025). The Census Bureau’s business survey, which asks firms about use, put the share of US firms using AI at 5.4% in February 2024 and the rate for finance and insurance at 6.9% (Bonney et al., 2024).

Regulators have started to police the gap between claim and capability. In March 2024 the SEC settled charges against two investment advisers, Delphia and Global Predictions, which had described AI they did not have; they paid $400,000 in penalties, and the SEC’s chair said “Such AI washing hurts investors” (SEC, 2024).

In January 2025 the SEC found that Presto Automation had told investors its drive-thru AI achieved “non-intervention” rates above 95%, when units running its own technology needed human agents in most instances, and in all of them for most of the period. A more advanced version piloted later in 2023 still needed a human agent to enter orders about 70% of the time, so at most about 3 in 10 went through without a person typing them in (our arithmetic; SEC order, 2025). One executive had written internally that “with HITL [humans in the loop], accuracy is not a major concern”.

Outcome data are scarcer and more sober. Anders Humlum and Emilie Vestergaard linked Danish adoption surveys to administrative records for 11 occupations exposed to chatbots, accountants and financial advisors among them. The typical user reported time savings of 2.8%, about 67 minutes of a 40-hour week, and the estimated effects on earnings and recorded hours were precise zeros, ruling out anything larger than 2% two years after ChatGPT’s launch (Humlum & Vestergaard, 2025). Workers said only 3% to 7% of the time saved turned into higher pay (BFI working paper). A board that reads mentions as results risks paying for a capability that turns out to be people.

Will better models make this evidence obsolete?

Speed is the case against this argument. In January 2023, ChatGPT scored 47.5% on 28,085 accounting assessment questions against a student average of 76.7% (Wood et al., 2023). A study published in 2024 found that ChatGPT 3.5 still failed the CPA, CMA, CIA and EA exams, averaging 53.1%, but that GPT-4 with 10-shot prompting and tools such as a calculator passed every section, averaging 85.1%. Its authors concluded “that ChatGPT has sufficient capabilities to disrupt the accounting and auditing industries” (Eulerich et al., 2024). On that pace, FinanceBench’s 81% describes 2023 models, and caution about AI in finance may be guarding against errors that newer tools no longer make.

Eulerich and his co-authors have a point about benchmarks: any accuracy figure in this article will date within a year, and the 2023 failure rates say nothing exact about today’s tools. Their point does less damage to the argument itself, for three reasons. Higher accuracy lowers the number of errors without saying which entries are wrong, so the review remains, as the table below shows. At 99% accuracy, a hypothetical company coding 5,000 transactions a month still has 50 wrong entries a month to find (our arithmetic). The statute and the audit firms assign responsibility to a person whatever the error rate. And in Denmark through 2024, Humlum and Vestergaard’s null results hold even for workers who reported large gains, which suggests the constraint lies in how work is organised.

What a more accurate tool still leaves for review

Coding accuracy	Wrong entries a month	Wrong entries a year
95%	250	3,000
98%	100	1,200
99%	50	600
99.5%	25	300

Hypothetical company coding 5,000 transactions a month; accuracy rates are illustrative, and the counts follow from them. Confidence scores tell a reviewer where to look first, which is the behaviour Choi and Xie observed; they do not remove the review.

Where to start with AI in the books, and what to ask of it

A finance team adopting these tools can take the steps in order.

1. Measure close time and correction rates before any tool goes in. A 7.5-day gain is one study’s association; the team’s own baseline is the test that counts.
2. Start where an output traces to a record: transaction coding, payment matching, reconciliations and first drafts. Review everything below the tool’s confidence threshold and sample a share above it.
3. Ask the vendor for its intervention rate and how it counts an intervention. Presto’s reported rates counted restaurant staff and left out its off-site human agents.
4. Require a citation for every AI answer about a contract, a filing or a policy, and open the cited page before the answer travels.
5. Rotate junior staff through some coding and reconciliation by hand, so the people reviewing exceptions have seen the routine cases.
6. Keep a named person on every number that leaves the company: the board pack, the lender’s covenant report, the tax return. The month-end close checklist below is a place to record which steps a tool performs and who reviews each.

A director or investor reading the board pack needs three answers from management. The first is which numbers were drafted by AI and who checked them against source records. The second is the measured rate behind any claim the company makes about its AI, with its definition and sample. The SEC asked as much of Presto and Delphia. The third is who provides effective challenge, in SR 26-2’s sense, for any model that feeds reported numbers: someone with the expertise, the independence and the standing to force a change.

Cheap checking decides the return

AI in finance earns its keep where a person can trace its output to a record faster than producing it, which today means coding, matching, reconciling and drafting. Where the check is as slow as the work, it adds a draft and a liability. The statute, the audit firms and a tribunal have each declined to let software carry the responsibility, and the bank regulators have declined, for now, to write rules for it. The practical question for a finance team is how to make checking cheap. How far newer models move the line between checkable and uncheckable work is still open, since every accuracy figure cited here belongs to the models tested at the time.

References

• Bainbridge, L. (1983). Ironies of automation. Automatica 19(6), 775–779.
• Bonney, K. et al. (2024). Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey. US Census Bureau, CES Working Paper 24-16.
• CFO Dive (2025). Deloitte refunds over $60K for report with AI errors, Australian government says. 21 October.
• Choi, J. H. and Xie, C. (2026). Human + AI in accounting: early evidence from the field. Journal of Accounting Research; working-paper version, Stanford Graduate School of Business.
• Civil Resolution Tribunal of British Columbia (2024). Moffatt v. Air Canada, 2024 BCCRT 149.
• Dell’Acqua, F. et al. (2026). Navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science.
• Eulerich, M., Sanatizadeh, A., Vakilzadeh, H. and Wood, D. A. (2024). Is it all hype? ChatGPT’s performance and disruptive potential in the accounting and auditing industries. Review of Accounting Studies.
• Federal Reserve, OCC and FDIC (2026). SR 26-2, Revised Guidance on Model Risk Management, 17 April.
• Humlum, A. and Vestergaard, E. (2025). Still waters, rapid currents: early labor market transformation under generative AI. NBER Working Paper 33777 (first circulated as “Large Language Models, Small Labor Market Effects”; BFI Working Paper 2025-56).
• Islam, P. et al. (2023). FinanceBench: a new benchmark for financial question answering. arXiv 2311.11944.
• JPMorgan Chase & Co. (2013). Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses, 16 January.
• McKinsey & Company (2025). AI in finance: driving automation and business value (Sukharevsky, West, Catania, Grande and Tricoli).
• PCAOB staff (2024). Spotlight: Staff Update on Outreach Activities Related to the Integration of Generative Artificial Intelligence in Audits and Financial Reporting, July.
• SEC (2024). Press release 2024-36, SEC charges two investment advisers with making false and misleading statements about their use of artificial intelligence, 18 March.
• SEC (2025). In the Matter of Presto Automation Inc., Securities Act Release No. 11352, 14 January.
• SEC EDGAR full-text search, form 10-K, phrase “generative AI”, filing years 2022 to 2025 (queried 3 October 2026).
• The Register (2025). Deloitte refunds Australian government over AI in report. 6 October.
• US Code: 15 U.S.C. § 7241 (Sarbanes-Oxley Act § 302); 18 U.S.C. § 1350 (Sarbanes-Oxley Act § 906).
• Wood, D. A. et al. (2023). The ChatGPT artificial intelligence chatbot: how well does it answer accounting assessment questions? Issues in Accounting Education 38(4), 81–108.

Study figures come from each paper’s published abstract unless the text says otherwise, the Deloitte figures from press reports, and every benchmark result belongs to the models tested at the time. Shares and conversions given without a source of their own, such as minutes saved a week or the share of Deloitte’s fee refunded, are our calculation from the cited figures, as is the count of unique 10-K filings on EDGAR. The company coding 5,000 transactions a month in the accuracy table is hypothetical.