FOXNUT

Statistics

Can AI do financial statements? Only part of the work

AI can do part of a set of financial statements. Extracting the figures is largely solved. Deciding what each figure is has not been.

By the Foxnut team · Updated

Can AI do financial statements?

On the published evidence, AI can do part of a set of financial statements. Extracting the figures is largely solved, and deciding what each figure is has not been. On the full US GAAP taxonomy of 17,688 concepts, the best model tested pulls a number and its context out of a filing at an F1 of 0.6932, and then maps it to the correct accounting concept 18.89% of the time.

We are Foxnut Studios, a senior-only practice in Bengaluru and Paris that builds AI systems into commercial operations and hands them over. You will find a link to talk to us at the end of this article. Let’s look at the evidence first.

That gap is the whole process. Producing statements is expensive: the reporting regulator estimates 1,695.2 hours and $334,438.81 of cost per US annual filing. The expensive part is judgement that is audited, published and, when wrong, corrected in public.

Each record below carries the figure, who was counted, over what period, and the publication it was read from, so you can check any single row without trusting this page. What a build in this area has to hand over at the end is what a reporting system’s owner needs to keep it honest.

1,695.2 hours

Estimated burden per Form 10-K response, the annual filing that carries a US registrant's audited financial statements, across roughly 6,740 responses a year, for a total annual reporting burden of 11,425,648 hours

Method Paperwork Reduction Act extension notice for Exchange Act Form 10-K, published in the Federal Register on 31 July 2026 and read on govinfo.govSample Every Form 10-K filer. The regulator's own burden estimate for the collection, not a survey of preparersPeriod Estimate published 31 July 2026Source US Securities and Exchange Commission, Federal Register, 31 July 2026

$334,438.81

Estimated cost burden per Form 10-K response, for a total annual cost burden of $2,254,117,579 across the same 6,740 responses

Method The same Paperwork Reduction Act notice, read on govinfo.gov. The notice states the per-response and total figures and does not break them into internal and external labourSample Every Form 10-K filerPeriod Estimate published 31 July 2026Source US Securities and Exchange Commission, Federal Register, 31 July 2026

222

Annual reports and accounts reviewed by the UK reporting regulator in its 2024/25 cycle, down from 243 in 2023/24 and 263 in 2022/23, of which 38% were FTSE 350 companies

Method Monitoring statistics published in the Annual Review of Corporate Reporting 2024/25, read in the report at page 8Sample UK companies selected for corporate reporting review - FTSE 350 companies on a five-year cycle, plus risk-based and thematic selectionsPeriod 2024/25 monitoring cycle, published September 2025Source Financial Reporting Council, Annual Review of Corporate Reporting 2024/25

37%

Share of those reviews that produced a substantive letter asking the company for further information or explanation: 26% within the FTSE 350 and 44% outside it, against 47% overall the previous year

Method The same monitoring statistics, page 8. A substantive question is asked only where a material breach of a reporting requirement appears possibleSample The 222 reviews performed in the 2024/25 cyclePeriod 2024/25 monitoring cycleSource Financial Reporting Council, Annual Review of Corporate Reporting 2024/25

18

Companies required to refer to the review in their next annual report, which the regulator describes as typical where a company restates comparative information in the primary financial statements. One of the 18 affected profit, against five of 26 the previous year

Method The same monitoring statistics, page 8Sample The 222 reviews performed in the 2024/25 cyclePeriod 2024/25 monitoring cycleSource Financial Reporting Council, Annual Review of Corporate Reporting 2024/25

10%

Share of reviews raising impairment of assets, the most frequently raised issue for the third consecutive year, ahead of cash flow statements and financial instruments at 9% each, then presentation of financial statements and revenue at 5% each

Method Top-ten issues table in the same report, page 9. Restatement and a required reference arose in the cash flow statement, financial instrument, presentation and income tax categoriesSample The 222 reviews performed in the 2024/25 cyclePeriod 2024/25 cycle, with 2023/24 comparativesSource Financial Reporting Council, Annual Review of Corporate Reporting 2024/25

0.1889

Best reported accuracy at mapping an extracted financial figure to the correct concept in the full 17,688-concept US GAAP taxonomy, against an F1 of 0.6932 for extracting the figure in the first place. Both are DeepSeek-V3, the strongest model on each subtask

Method FinTagging benchmark, which splits XBRL tagging into numerical extraction and taxonomy concept alignment; results tables read in the HTML version of the paperSample Filing data tagged against the complete official US GAAP taxonomy rather than a truncated subset; the concept-alignment set holds 261,457 query-answer pairsPeriod Submitted 27 May 2025, last revised 17 May 2026Source Wang and others, FinTagging, arXiv:2505.20650

81%

Share of questions that GPT-4-Turbo with a retrieval system answered incorrectly or declined to answer, on a benchmark of 10,231 questions about publicly traded companies with answers and evidence strings attached

Method FinanceBench, 16 model configurations evaluated with manual review of the outputsSample 10,231 questions drawn from public company filingsPeriod Submitted 20 November 2023Source Islam and others, FinanceBench, arXiv:2311.11944

95% to 100%

False-positive rate on clean statements in nine of fourteen complete model runs asked whether a set of financial statements is numerically consistent. One run reported 0% false positives; on a rounded variant of the same set that model's recall was 79.0%, against 100% on the unrounded one

Method FinVerBench, built from SEC 10-K XBRL filings with four categories of injected error - arithmetic, cross-statement linkage, year-over-year and magnitude. A preprint, not peer reviewed, and a single-author studySample 43 S&P 500 companies; a 105-instance observable subset of 43 clean and 62 error-injected casesPeriod Submitted 28 May 2026Source Panda, FinVerBench, arXiv:2605.29586

Every number on this page, with what it measures, its period and its source
NumberWhat it measuresPeriodSource
1,695.2 hoursEstimated burden per Form 10-K response, the annual filing that carries a US registrant's audited financial statements, across roughly 6,740 responses a year, for a total annual reporting burden of 11,425,648 hoursEstimate published 31 July 2026US Securities and Exchange Commission, Federal Register, 31 July 2026
$334,438.81Estimated cost burden per Form 10-K response, for a total annual cost burden of $2,254,117,579 across the same 6,740 responsesEstimate published 31 July 2026US Securities and Exchange Commission, Federal Register, 31 July 2026
222Annual reports and accounts reviewed by the UK reporting regulator in its 2024/25 cycle, down from 243 in 2023/24 and 263 in 2022/23, of which 38% were FTSE 350 companies2024/25 monitoring cycle, published September 2025Financial Reporting Council, Annual Review of Corporate Reporting 2024/25
37%Share of those reviews that produced a substantive letter asking the company for further information or explanation: 26% within the FTSE 350 and 44% outside it, against 47% overall the previous year2024/25 monitoring cycleFinancial Reporting Council, Annual Review of Corporate Reporting 2024/25
18Companies required to refer to the review in their next annual report, which the regulator describes as typical where a company restates comparative information in the primary financial statements. One of the 18 affected profit, against five of 26 the previous year2024/25 monitoring cycleFinancial Reporting Council, Annual Review of Corporate Reporting 2024/25
10%Share of reviews raising impairment of assets, the most frequently raised issue for the third consecutive year, ahead of cash flow statements and financial instruments at 9% each, then presentation of financial statements and revenue at 5% each2024/25 cycle, with 2023/24 comparativesFinancial Reporting Council, Annual Review of Corporate Reporting 2024/25
0.1889Best reported accuracy at mapping an extracted financial figure to the correct concept in the full 17,688-concept US GAAP taxonomy, against an F1 of 0.6932 for extracting the figure in the first place. Both are DeepSeek-V3, the strongest model on each subtaskSubmitted 27 May 2025, last revised 17 May 2026Wang and others, FinTagging, arXiv:2505.20650
81%Share of questions that GPT-4-Turbo with a retrieval system answered incorrectly or declined to answer, on a benchmark of 10,231 questions about publicly traded companies with answers and evidence strings attachedSubmitted 20 November 2023Islam and others, FinanceBench, arXiv:2311.11944
95% to 100%False-positive rate on clean statements in nine of fourteen complete model runs asked whether a set of financial statements is numerically consistent. One run reported 0% false positives; on a rounded variant of the same set that model's recall was 79.0%, against 100% on the unrounded oneSubmitted 28 May 2026Panda, FinVerBench, arXiv:2605.29586

Is financial reporting worth automating?

Part of it is. Six structural dimensions decide whether a process is worth automating, and financial reporting reads at an extreme on three of them. The regulatory exposure is the highest in this family, the errors that matter sit in judgement and not in volume, and the software market is mature at the layer that formats the output and demonstrably immature at the layer that classifies it.

Financial reporting: process profile
DimensionWhat it reads on financial reportingSource
Exception varianceConcentrated in judgement, not in volume. The issues that draw a regulator's substantive question are impairment of assets (10% of reviews), cash flow statements and financial instruments (9% each), then presentation of financial statements and revenue at 5% each. Routine postings are not what goes wrong; estimates and classifications are, and they are the part of the work that changes every period.Financial Reporting Council, Annual Review of Corporate Reporting 2024/25
VolumeSet by the calendar, not by a queue. A registrant files once a year, at an estimated 1,695.2 burden hours and $334,438.81 of cost per response across roughly 6,740 filers. Payback is therefore a question about a recurring cycle of known size and a fixed deadline, not about throughput - which is the opposite of the invoice-style processes in this family.US Securities and Exchange Commission, Federal Register, 31 July 2026
Cost of an errorPaid in public, by the preparer, a year later. A material error becomes a restatement of comparative information in the primary statements plus a required reference to the regulator's review in the next annual report: 18 such references in the 2024/25 cycle, one of which affected profit, against 26 the year before.Financial Reporting Council, Annual Review of Corporate Reporting 2024/25
ReversibilityReversible on the page, not in the record. The correction mechanism exists and is used, but it runs through the following year's accounts and is legible to every reader of them. Automated consistency checking is not yet the safety net it sounds like: nine of fourteen model runs judging clean statements flagged 95% to 100% of them as erroneous.Panda, FinVerBench, arXiv:2605.29586
Regulatory exposureThe highest in this family, and it is statutory rather than contractual. The statements must give a true and fair view under section 393 of the Companies Act 2006 and IAS 1.15, they are audited, and a national regulator reviews a sample each year: 222 reviews in 2024/25, 37% of which produced a substantive letter to the company.Financial Reporting Council, Annual Review of Corporate Reporting 2024/25
Vendor market maturityMature at the filing layer, immature at the judgement layer. The output is a mandated machine-readable artefact tagged against a taxonomy of 17,688 US GAAP concepts, and the tooling around tagging and filing is long established. Best reported accuracy at choosing the right concept for an extracted figure is 0.1889, against an F1 of 0.6932 for extracting it.Wang and others, FinTagging, arXiv:2505.20650

Two of those rows carry the argument. The vendor-maturity row says the market is mature about the wrong half of the problem: formatting, tagging and filing a statement is a solved commodity, and classifying a transaction correctly is where the published accuracy collapses.

The exception-variance row says the same thing from the regulator’s side. The issues raised most often are impairment, financial instruments and revenue, all of which are estimates and classification decisions and not arithmetic.

The volume row is where financial reporting parts company with every other process in this family. An invoice queue rewards automation in proportion to how many invoices arrive. A reporting cycle arrives twelve or four or one times a year, at a size known in advance, against a filing deadline.

That makes the payback question a labour question about a known peak, not a throughput question. It is why the honest first measurement is how many hours the close and the reporting pack actually consume, and not how many entries pass through them.

Is budget variance reporting the same process?

No. Budget variance reporting looks like the same work, with the same ledger, the same period and the same finance team, and it sits on the opposite side of every dimension above. Nothing in a variance pack is filed, audited or attested. There is no taxonomy to map an account to and no external reader who can be misled, so a misclassification costs an internal argument and not a restatement. A wrong variance explanation is corrected in the next meeting, not in next year’s comparatives.

So the two halves of what people call financial reporting have almost inverted automation profiles. The statutory half is constrained by classification accuracy and regulatory exposure, which is exactly what the benchmark evidence says machines are worst at. The management half is constrained by whether anyone trusts the explanation attached to the number, which is a different and considerably softer problem.

Anyone asked to automate financial reporting should establish which half is meant before quoting anything, because a variance-commentary tool and a statutory reporting system share a data source and almost nothing else.

Budget variance also has no measurable search demand in the data we harvested, in any question frame. That says something about how the work is discussed and not about whether it matters, and it is one reason this section sits here and not on a page of its own.

What do most claims about AI and financial statements get wrong?

The first error is treating financial statement analysis and financial statement production as one capability. Reading a set of accounts and forming a view is a research task with a tolerant error profile. Producing a set is a classification task with an audited one, and the two have opposite failure costs.

Even the analysis result is sobering. On questions about public company filings, with the evidence available to a retrieval system, the strongest configuration tested answered 81% of them wrongly or declined to answer.

The second error is repeating figures whose source has stopped standing behind them. The most circulated academic claim in this area is that a large language model outperforms human analysts at predicting earnings direction from anonymised statements. The paper behind it was withdrawn on 20 February 2025, after a co-author identified inconsistencies in the data and analyses while attempting to replicate earlier work. It is in the sources below for that reason, and none of its numbers appear on this page.

Our rule is stricter than checking whether a claim is popular. A figure that cannot be traced to a live primary publication with a date is not a statistic, however often it is quoted, and the widely repeated project-failure percentages in this space fail that test entirely.

The third error is reading a low benchmark score as a permanent verdict. The concept-mapping number will move, and the extraction number already has. What is less likely to move quickly is the gap between them, because extraction is a text problem with abundant training signal, and concept mapping is a judgement problem against a 17,688-item taxonomy where the right answer often depends on facts that are not in the document at all.

What can this page not tell you?

It cannot tell your organisation what its own reporting cycle costs. The Form 10-K figures are a regulator’s estimate for a filing that most companies never make, and we quote them as the only per-statement cost figure published by a named body with a date, not as a benchmark for a private company’s close.

Important note: nothing here is accounting advice. We take no position on how any transaction should be recognised, measured or presented. That is the standard-setter’s territory and the auditor’s, and this page has no standing in it.

It also cannot answer whether your finance function is ready to automate any of this, which turns on whether the underlying ledger data is complete, documented and consistently coded. That is a data-quality question and not a reporting one, and the wider readiness assessment sits alongside it.

How were these figures compiled?

Every record was read from its primary source in August 2026: the Securities and Exchange Commission’s Paperwork Reduction Act notice for Form 10-K, the Financial Reporting Council’s annual review of corporate reporting, read page by page, and three papers on arXiv. The Federal Register figures were read twice to confirm the per-response and total values. The notice does not split the cost burden into internal and external labour, so we claim no such split.

We dropped three sources and did not approximate them, because they could not be reached: APQC’s close-cycle benchmarks, which would have given a days-to-close figure; the US Bureau of Labor Statistics handbook entry for accountants and auditors, which would have given a labour-cost figure; and the Commission’s own PDF of Form 10-K, whose burden block we read from the Federal Register notice instead. No figure from any of them appears here in paraphrase, and none came from a search result summarising them.

One source is cited as a caution and not as evidence: the withdrawn paper on financial statement analysis. Two of the three benchmark papers are peer-reviewed venue submissions and one, FinVerBench, is a single-author preprint, which its record states.

Should you automate financial statements?

The evidence supports automating the production of financial statements at the extraction and assembly layer. It does not yet support automating the classification layer, which is where the value and the risk both sit.

A system that pulls figures from source records, assembles the pack, checks it against the previous period and formats it for filing is doing work machines demonstrably do well, and it removes a real and recurring cost: the regulator’s own estimate is 1,695.2 hours per annual filing. A system that decides which account a transaction belongs in, or whether an asset is impaired, is working where the best published accuracy is 0.1889 and where the regulator’s most frequent challenges land.

The reason is structural and not temporary. Three of the six dimensions point the same way: the exceptions are judgement calls and not volume, the errors are corrected in public a year later, and the exposure is statutory. A process with that profile does not reward removing the reviewer. It rewards giving the reviewer less to do and a clearer record of what was decided, which makes the audit trail the centre of any such build, more than the model or the interface.

For a finance team, the useful first question is not which product to buy. It is which half of the work is actually being automated, how many hours the last close consumed and where they went, and whether the ledger data is coded consistently enough for a machine to classify from at all. A team that can answer those three can buy on evidence. A team that cannot is being asked to automate a judgement process it has not measured, and measuring it first is cheaper than any of the alternatives.

We would start you there too, and we would tell you if the evidence does not support the build you came to ask about. The senior people you speak to are the people who would do the work, and we launched our own products before we advised anyone on theirs. A free audit comes first, and what we do for clients is one click away. Tell us what your last close cost and we will help you find where the hours went.

Sources

  1. US Securities and Exchange Commission, 'Agency Information Collection Activities; Submission for OMB Review; Comment Request; Extension: Exchange Act Form 10-K', Federal Register document 2026-15557, published 31 July 2026 - read on govinfo.gov Retrieved
  2. Financial Reporting Council, 'Annual Review of Corporate Reporting 2024/25', September 2025 - monitoring statistics at pages 8 to 9, reporting-requirement framing at page 7 Retrieved
  3. Wang, Qian, Peng and others, 'FinTagging: Benchmarking LLMs for Extracting and Structuring Financial Information', arXiv:2505.20650, submitted 27 May 2025 and last revised 17 May 2026 - results tables read in the HTML version of v5 Retrieved
  4. Islam, Kannappan, Kiela, Qian, Scherrer and Vidgen, 'FinanceBench: A New Benchmark for Financial Question Answering', arXiv:2311.11944, submitted 20 November 2023 Retrieved
  5. Panda, 'FinVerBench: Benchmark Validity and Calibration in Large Language Model Financial Statement Verification', arXiv:2605.29586, submitted 28 May 2026 - a preprint, not peer reviewed Retrieved
  6. Kim, Muhn and Nikolaev, 'Financial Statement Analysis with Large Language Models', arXiv:2407.17866, submitted 25 July 2024 and WITHDRAWN 20 February 2025 after a co-author identified inconsistencies in the data and analyses Retrieved

Foxnut Studios takes on work like this from Bengaluru and Paris. If you want to see how we do it before you talk to anyone, here is what an AI engagement costs before you ask.