A week after results, the CFO of a listed company asked a public AI service to summarise the year, and the answer showed a profit where the company had reported a loss. The annual report carried the correct figure, audited and signed, but when a machine read the PDF, the brackets marking the figure as negative were lost. The CFO is hypothetical, but the pattern is one we are seeing in published financial PDFs.
We have analysed 4.5m PDF pages and examined how their information is treated within AI-generated results. We have also set source financial PDFs beside the answers AI services gave about the same companies. The incorrect answers we found trace back to flattened tables, poor document structure and characters that failed to extract.
01What the machine receives.
A PDF can record where each character sits on the page without recording which figures belong together. A reader sees a table because the eye groups the columns. A machine reads whatever structure the file carries, and in a flattened table that structure is a single run of text, with the headings first and the values from several columns following in one sequence.
Once the grid is gone, a value can come loose from its year, its column heading, its currency or unit, the brackets that make it negative, the total it belongs to and the footnote that qualifies it. The same break can move a figure from discontinued to continuing operations, or from an adjusted measure to a reported one. W3C guidance on tagged PDF notes that tables converted to PDF can come out with cells wrongly merged or split, even when the source table was built correctly. The fault can arise in conversion, after the people preparing the report have finished their work.
The illustration below is hypothetical and is not drawn from any company.
| £m | 2026 | 2025 |
|---|---|---|
| Revenue | 412.6 | 398.1 |
| Operating profit/(loss) | (18.4) | 27.9 |
| Profit/(loss) for the year | (12.7) | 19.3 |
| As a machine extracts it |
|---|
| £m 2026 2025 Revenue 412.6 398.1 Operating profit/(loss) 18.4 27.9 Profit/(loss) for the year 12.7 19.3 |
Read in that order, the company made a profit of 12.7 for the year. The brackets were lost with the character mapping, and nothing left in the sequence marks the figure as a loss. Without the grid, nothing ties 12.7 to 2026 rather than 2025 either.
Several steps sit between the PDF and the answer: crawling, indexing, selecting a document, downloading it, text extraction or optical character recognition, table reconstruction, retrieval and the model's own synthesis. Each platform runs its own version of these steps, with its own indexes, access permissions and cached copies, so 2 services can read the same report differently.
Fonts add a separate failure that the page does not show. What appears on screen is a drawn shape, and what a machine extracts is the character code mapped to that shape. Where a font fails to load or map, the page still looks right while the extracted text loses characters, corrupts them or drops a whole line. Proper tagging removes avoidable ambiguity, though retrieval and interpretation remain with each platform.
02What we are finding, and why no one reports it.
79% of misinformation or misrepresentation has the organisation's own online position as its root cause. AAAnow Research, Dec. 2023 to Jan. 2026; P&C (human misinformation) 2014 to 2023. Financial PDFs are a concentrated part of that content, for a reason that sharpens the problem.
In our analysis, PDFs carried between 1.44 and 1.8 times greater authority within AI-generated results. The same format is failing structurally: we observed the table-flattening pattern in 23% to 27% of relevant reviewed cases involving accounting data, with the rate varying by market and by the composition of the financial information. Fonts failed to load or map correctly in 11% of reviewed cases. The format carrying the greatest authority within AI-generated results is the format most frequently failing structurally.
In current FTSE 100 annual reports we have documented a flattened PDF, a font that was not embedded, and a file large enough to affect Gemini's access to it. We have also seen Google Search and Google Gemini present materially different information about the same organisation. Access restrictions, source selection and differing processing can each contribute, and we attribute no single case to one cause without direct evidence.
Our published use cases show how this reaches the financial record. Values come loose from their headings and periods, a results table arrives with figures in the wrong columns, and a large annual report is read in part while the answer fills the gap it did not read. Older material does its own damage: when the web page that dated an old annual report is removed, the PDF stays online and answers as the current year.
Access decisions can hand the answer to someone else. In one case a company blocked AI, its results documents carried no titles or authors, and a hostile party's version of the results was the one AI cited. Two further cases in our library are illustrative, built from observed mechanisms: earlier guidance still cited after the company revised it, and a withdrawn announcement quoted as the current statement.
None of this produces a signal the company would see. Pew Research Center's 2026 survey of 5,119 US adults found that 60% report reading AI summaries at the top of search results, and 4 in 10 use chatbots to search for information. Those figures cover the US adult population as a whole, and they show how much reading now ends at the summary. The misinformed reader stays inside the answer and has no reason to open the document, so no correction signal reaches the issuer.
Pew Research Center, Americans and AI 2026Pew Research Center, Americans and AI 2026We do not enable external links - please copy and paste, you are then certain as to the link being opened
03From production defect to governance question.
For years the defect stayed out of sight for a sound reason. A financial PDF was approved because its figures matched the signed statements on screen, and conventional review had no view of what a machine would read. That history explains how the problem began, and it stops being a defence once the condition can be identified, assessed and measured.
Reading order, table relationships, character extraction and the answers AI services return can now be tested against the official record. The question is no longer only whether somebody produced the PDF incorrectly. It is what management did after learning that the publication process was already causing financial information to be misrepresented. Awareness of the gap proves neither negligence nor breach, but continued ignorance is a weak position once the gap can be evidenced.
The same results now exist in 4 versions. They are the official regulatory filing or structured report, the PDF or other copy on the corporate website, what a machine extracts and indexes from that copy, and the answer an AI service produces. The filing can stay correct while the website copy is misread and the answer carries the error onward. An AI answer is the platform's output, and it becomes a company statement only if the company produces or adopts it.
04Exchange Act Rules 13a-14 and 13a-15.
In the US, Exchange Act Rules 13a-14 and 13a-15 require the CEO and CFO to certify each annual and quarterly report and to maintain disclosure controls and procedures. The SEC drew those controls wider than internal control over financial reporting, covering non-financial as well as financial information, but tied them to what is filed under the Exchange Act. A website PDF and an AI answer sit outside that definition. The filed report on EDGAR, tagged in Inline XBRL, carries the reporting period with each tagged value, which is the relationship a flattened PDF loses.
SEC, Release 33-8124SEC, Release 33-8124We do not enable external links - please copy and paste, you are then certain as to the link being opened, SEC, Inline XBRLSEC, Inline XBRLWe do not enable external links - please copy and paste, you are then certain as to the link being opened
05UK, DTR 4.1.
In the UK, DTR 4.1 requires the annual financial report to be prepared in XHTML, with consolidated IFRS statements marked up in Inline XBRL. The persons responsible must confirm that the financial statements show a true and fair view. The issuer is responsible for all information drawn up and made public under that section, and the report must remain publicly available for at least 10 years. A decade of annual reports therefore stays within reach of AI systems under the rules themselves.
FCA Handbook, DTR 4.1FCA Handbook, DTR 4.1We do not enable external links - please copy and paste, you are then certain as to the link being opened
06UK, Corporate Governance Code 2024 / 29.
The UK Corporate Governance Code 2024 is a governance code rather than a regulation, and its Provision 29 asks boards to declare the effectiveness of their material controls, including reporting controls, as at the balance sheet date. It applies to periods beginning on or after 1 Jan. 2026, so the first declarations arrive in 2027, and the FRC leaves the choice of material controls to each company. Each board now decides whether controls over the published financial record belong among them. The FRC's own HTML edition of its Provision 29 guidance carries a notice that it was converted from PDF by an AI tool without human verification, and that readers should rely on the original PDF.
FRC, Provision 29 MythbusterFRC, Provision 29 MythbusterWe do not enable external links - please copy and paste, you are then certain as to the link being opened
Across the EU, ESMA set its single electronic format to apply from 2020, with annual financial reports in XHTML and IFRS consolidated statements tagged in Inline XBRL, so that financial information can be analysed without manual reprocessing. Regulators in the US, UK and EU have each built a machine-readable version of the record. The corporate website still carries the PDF beside it, and that format carries the greater authority within AI-generated results.
ESMA, single electronic formatESMA, single electronic formatWe do not enable external links - please copy and paste, you are then certain as to the link being opened
None of these rules requires a company to monitor what AI services say about it, and we know of no enforcement action built on an AI answer. Directors' and officers' (D&O) liability runs instead through older principles meeting a gap that can now be measured. Under section 174 of the Companies Act 2006, a UK director's care, skill and diligence is judged against what the role demands and against the knowledge, skill and experience that director actually has. In Delaware, Marchand v Barnhill (2019) confirmed that directors must make a good-faith effort to put board-level monitoring and reporting in place for risks critical to the business.
Companies Act 2006, section 174Companies Act 2006, section 174We do not enable external links - please copy and paste, you are then certain as to the link being opened, Jones Day on Marchand v BarnhillJones Day on Marchand v BarnhillWe do not enable external links - please copy and paste, you are then certain as to the link being opened
Marchand concerned food safety at a company with a single product, and we know of no court that has treated machine-read financial information as a risk of that kind. Both tests still turn on what directors knew and what they did. A recorded baseline, recorded corrections and later checks show both. Our use cases record the same question surfacing at D&O renewal, where evidenced control bears on risk quality, terms and limits, though no outcome is automatic.
The insurance market is reading the same signals. Aon's Mar. 2026 D&O insight expects more derivative actions alleging inadequate board oversight of AI governance and controls, and advises companies to ensure the accuracy of their public statements. Gallagher's 2026 survey of more than 35 public company D&O underwriters found 88% expecting D&O risk to rise. The underwriters are looking beyond AI washing to AI's wider effect on exposure, and they name the CEO and board as the roles most critical to managing it.
Aon, AI litigation and D&O exposuresAon, AI litigation and D&O exposuresWe do not enable external links - please copy and paste, you are then certain as to the link being opened, Gallagher, D&O risks for 2027Gallagher, D&O risks for 2027We do not enable external links - please copy and paste, you are then certain as to the link being opened
Materiality decides how much weight each discrepancy should carry. A misread figure in a report from 9 years ago sits in a different place from a profit shown for a current loss in the week after results. Where the misinformation concerns material financial information, persists across public channels, conflicts with the official disclosure or stays uncorrected after discovery, the consequences can reach regulators, investors and the audit committee. The degree depends on the jurisdiction and the facts.
Our use cases show the commercial cost arriving first, in corrections that have to reverse an established position, with each repetition of the wrong figure reinforcing it. That cost falls on investor relations and on confidence in management.
07The CFO and the CIO.
The CFO owns the integrity of the financial meaning: the reporting process, disclosure controls, materiality and the consequences of a wrong number in the market. A CIO could fairly see the annual report as finance's document, produced by finance and its printers and hosted on the website.
The divergence between Google Search and Gemini changes that view. Which version of the company an AI service reads depends on access rules, crawler permissions, file size, fonts and the processing chain, and each of those sits in the technology estate. The oversized FTSE 100 file affecting Gemini's access is a technology cause, and so is the company's own block on AI behind the hostile-results case.
The CFO cannot correct those causes, and the CIO cannot judge which figures are material. The accountability is shared, and neither half transfers to the other. For the board and audit committee, the point is ownership and proportion. A defective PDF is not automatically a material control deficiency, and the committee needs to see which exceptions are material, who owns them and when they close.
08Inside the company.
The same documents now move inside the company, into internal AI and agentic systems that ingest them automatically. We are seeing poorly structured PDF estates fed into these deployments as they stand. Where a table has been flattened, a character lost or a relationship broken, the organisation is supplying its own systems with defective source information.
Agentic systems retrieve, summarise and consolidate, so an error made at extraction is repeated and compounded at each step until its origin has gone. We see this contributing to inaccurate outputs and to deployments that fail and are then written off as proof that AI projects fail. One of our public-sector use cases shows the pattern beyond finance: a school district fed a new AI-enabled service for families with outdated policy documents, and families and staff acted on withdrawn policies and deadlines. Speed of adoption links these cases, because the AI was deployed on the documents as they stood.
09Practical answers already exist.
Our PDF Automation identifies, catalogues and profiles PDF estates, creates structured HTML versions of prioritised public PDFs, and focuses remediation of the original PDFs on the documents carrying the greatest value or exposure. Secure Tag is designed so AI systems can identify authoritative corporate information, supporting source verification and discovery, and it can control how legacy or historic content is treated. Both run on a platform of 187 agents performing more than 360,000 checks. They are ways to put governance decisions into operation, and neither guarantees a correct AI answer, removes regulatory exposure or provides legal compliance.
10A proportionate response.
Start with an inventory of financially significant public documents, classified by materiality and expected external use, keeping official filings distinct from website reproductions. Test those documents for reading order, table relationships, character extraction and font mapping, and reconcile values across the structured filing, any HTML pages and the PDF copies.
Put representative financial questions to the AI and search services that matter to your investors, and repeat them after each results date and filing. Review crawler permissions and AI access restrictions jointly with the CIO, and publish authoritative structured alternatives where a PDF cannot carry the meaning on its own.
Introduce a publication acceptance control signed by both the CFO and the CIO. Keep an exception register with accountable owners and target dates, set a correction and escalation route for material discrepancies, and preserve the evidence of how each risk was assessed and addressed. Report significant exceptions to the audit committee or the appropriate board forum.
The work is proportionate: it means understanding the exposure, setting priorities, recording decisions and controlling the financially significant documents first, with the historical estate following in order of risk.
11Before the next audit committee.
Ahead of that meeting, open a live AI service and ask it 2 questions about your latest results: what the company reported for the year, and how that compares with the year before. Then set both answers beside the signed accounts. If they match, the board has a baseline worth recording. If they differ, the gap can now be seen, and what happens next is the record directors will be asked about.
12About the author.
Lawrence Shaw is the founder of AAAnow. He has managed the Boeing/RR 777 EMCS, launched an ISP in 1999 and an early e-commerce platform in 2002. He was integral to EU Compliance (DP) regulation-drafting across 5 countries, and has consulted in automated regulatory compliance, risk management and process methodology for both Deloitte and IBM. His e-service delivery work spans Vodafone, eBay, Merck and Oracle, including the World's single largest client privacy deployment.
In 2023 he developed the ARTYMUS framework for AI content governance. He now applies 25+ years of compliance and risk discipline to AI readiness: the foundations, governance and visibility that determine how organisations are read by generative AI.
10 questions the CFO should ask.
Which of our financially significant public documents exist as PDFs, and who owns each one?
The reply should be a really list ranked by materiality, with a named owner, and with filings kept separate from website copies.
When did anyone last test how a machine reads our annual report and results announcement?
Look for a dated test of reading order, tables and character extraction, with the results attached.
Do the figures in our website PDFs reconcile with our structured filing?
Expect a reconciliation record listing each exception, including any that were accepted.
What do the main AI services say about our latest results today?
The answers should be dated, set against the signed figures and repeated after each results date.
Which AI crawlers can reach our investor pages, and who decided that?
A joint CFO and CIO record of the settings, with the reasons behind them, is the answer to look for.
Can superseded reports, guidance or announcements still be reached without a date or status on them?
Each item needs a recorded decision: label it, replace it, redirect it or retire it.
Who accepts a financial document for publication, beyond confirming that it looks right?
A named control covering machine reading, signed by finance and technology together, closes this gap.
Which internal AI systems ingest these PDFs, and were the documents checked before they went in?
The answer should name the systems and show the checks, or record that none were made.
What happens, and who acts, when a material AI discrepancy is found?
A written route with owners, an escalation point and a correction step should already exist.
What does the audit committee receive on this, and has our auditor said what its work covers?
A standing exception report with owners and dates, and the audit scope confirmed in writing.
Source and link register.
Our own figures are observed evidence from our analysis and are not externally published research, so they carry no public link here.
- U.S. Securities and Exchange Commission, Final Rule: Certification of Disclosure in Companies' Quarterly and Annual Reports (Release 33-8124).Effective 29 Aug. 2002. United States, regulator rule. Defines disclosure controls and procedures and the CEO and CFO certification.sec.gov/files/rules/final/33-8124.htmsec.govWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- U.S. Securities and Exchange Commission, Inline XBRL.Last updated 17 Jan. 2025. United States, regulator publication. Sets out Inline XBRL requirements and the period information carried with each tagged value.sec.gov/data-research/structured-data/inline-xbrlsec.govWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Financial Conduct Authority, DTR 4.1 Annual financial report.Last updated 28 Jul. 2023. United Kingdom, binding rules. Covers the XHTML and Inline XBRL format, responsibility statements, issuer responsibility and 10-year availability.handbook.fca.org.uk/handbook/dtr4/dtr4s1handbook.fca.org.ukWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Financial Reporting Council, Provision 29 Mythbuster: Preparing to report.29 Jan. 2026. United Kingdom, governance code guidance. Gives Provision 29 scope, effective date and first reporting year, and carries the notice on its AI-converted HTML edition.frc.org.uk/docs/9097/html/frc.org.ukWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- European Securities and Markets Authority, New rules make EU issuers' annual financial reports machine-readable.18 Dec. 2017. European Union, regulator publication. Introduces the single electronic format, with IFRS statements tagged in Inline XBRL inside XHTML reports.esma.europa.eu/press-news/esma-news/new-rules-make-eu-issuers%E2%80%99-annual-financial-reports-machine-readableesma.europa.euWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- legislation.gov.uk, Companies Act 2006, section 174: Duty to exercise reasonable care, skill and diligence.Revised text current to 6 Oct. 2026. United Kingdom, binding law. Sets the objective and subjective tests for a director's care, skill and diligence.legislation.gov.uk/ukpga/2006/46/section/174legislation.gov.ukWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Jones Day, Delaware Supreme Court Reinforces Directors' Oversight Obligations on Mission-Critical Subjects.Aug. 2019. United States (Delaware), professional interpretation of case law. Summarises Marchand v Barnhill, its limits and the value of board records.jonesday.com/en/insights/2019/08/delaware-court-reinforces-directors-oversightjonesday.comWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Aon, Artificial Intelligence Litigation: Emerging D&O Exposures.Mar. 2026. International, insurance market commentary. Expects more derivative actions alleging inadequate board oversight of AI governance and controls.aon.com/risk-services/financial-services-group/insight_171_artificial-intelligence-litigation-emeaon.comWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Gallagher, The D&O Risks Underwriters Will Be Watching in 2027.28 Sep. 2026. United States, insurance market survey. Survey of more than 35 public company D&O underwriters on risk direction, AI exposure and the roles that manage it.ajg.com/news-and-insights/the-d-and-o-risks-underwriters-will-be-watching-in-2027/ajg.comWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- W3C Web Accessibility Initiative, Technique PDF20: Using Adobe Acrobat Pro's Table Editor to repair mistagged tables.Updated 25 Sep. 2025. International, technical standard guidance. Notes that tables converted to PDF can lose their cell structure even when built correctly.w3.org/WAI/WCAG21/Techniques/pdf/PDF20.htmlw3.orgWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact.17 Jun. 2026. United States, independent research. Survey of 5,119 US adults in Feb. 2026 on AI summaries and chatbot use for searching.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/pewresearch.orgWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Our use case library, The use cases: what happened, and the impact.Updated 7 Sep. 2026. Our published use cases. Source of the financial reporting cases, with illustrative cases labelled.aaanow.ai/use-cases/
- City of Harlow Springs benchmark report (HSP-2512_05OCT), AI Readiness result, sample report.Assessed 2 Oct. 2026. Internal sample data, with no public link. Used in the article for the page 12 case only: it supplied the school district case, and its sample figures were not used.