Correct to People, Wrong to Machines: The Financial Reporting Control Gap

CFO REGULATORY GOVERNANCE | PUBLIC-COMPANY REPORTING

A correct report produces the wrong answer

Ask an AI service for the latest operating profit of an unnamed public company. It answers $315 million, with the confidence and presentation expected of a current result. The approved financial statements show $420 million; $315 million appears in the same table, but belongs to the prior year. This scene is a hypothetical composite built from recurring AAAnow findings, not a disguised account of one issuer.

Nothing in the visible report is necessarily wrong. The error sits in the relationship recovered from it, where current and comparative figures have separated from the periods that give them meaning.

AAAnow is already finding this outcome in live corporate information channels. Financial PDFs that look correct to finance teams contain reading orders, table structures and character maps that cause machines to recover a different account of the published results. Incorrect answers are not a forecast about future adoption; they are being produced from documents already online.

What disappears beneath a clear page

Human readers rebuild a table without noticing the work involved. Borders, spacing and accounting familiarity reveal which heading governs each number. To finance, the page still looks entirely correct, although the PDF has lost the logical connections among headings, values, units and periods. Extraction may then place headings first and follow them with values from several columns as one continuous linear sequence.

Financial meaning rarely sits inside a single number. Brackets signal negative values, footnotes change qualifications, and columns distinguish reported results from adjusted measures. Continuing operations, currencies, scaling conventions, totals and comparative years all depend upon relationships that a page may display without encoding.

Hypothetical illustration: a visually clear table and a flattened sequence

Possible machine sequence: 2025 | 2024 | Operating profit | 420 | 315

Human readers recognize the intended financial comparison almost immediately. The flattened sequence does not reliably bind $420 million to 2025 or $315 million to 2024. A retrieval system can select the older value, describe it fluently and return it as the current result.

Tagging reduces avoidable ambiguity without governing the complete journey from document to answer. Before a model answers, surrounding systems may crawl and index documents, choose a source, run OCR or another extractor, reconstruct its content and retrieve selected passages. Only then does the model synthesize a response for the user. Different permissions, cached copies, ranking signals and extraction systems help explain why identical questions produce different results.

Fonts create another failure that remains invisible on screen. A displayed glyph still needs character encoding or a Unicode map before software knows its meaning. Failed mapping can remove a digit, corrupt a character or disrupt an entire extracted line.

The evidence changes the control question

When an invisible defect becomes a known risk

Traditional publication controls answer necessary questions about approved figures, authorization and visible presentation. Those checks still matter, yet none demonstrates that a machine can recover the intended period, unit, sign, classification or qualification. Structural assessment adds the control question that visual review was never designed to answer.

Governance must also separate 4 information states because their legal status and control points differ:

1. The official regulatory filing or corresponding structured report.
2. The PDF or other copy published on the corporate website.
3. The information extracted or indexed by a machine.
4. The answer subsequently produced by an AI service.

A correct filing can coexist with a defective website PDF and an inaccurate external answer. Treating those states as interchangeable either overstates responsibility for third-party output or understates control over the corporate material that fed it. Both mistakes prevent management from reaching a proportionate response.

What began as an unrecognized production fault can now be traced. An approved document contains a defective machine layer, extraction separates accounting elements, and an AI service produces a plausible wrong answer. Testing or conflicting results then expose the discrepancy. AAAnow is observing those 4 conditions today, placing the issue within current financial reporting and disclosure governance for public companies.

Once discovery occurs, management must decide ownership, materiality, priority, correction and monitoring. Regulatory, legal, investor and audit consequences become more significant when material misinformation persists across public channels or remains unmanaged after the organization receives evidence.

The executive question is no longer where the production defect began. It is what management did after learning that publication was already misrepresenting financial information.

A defensible record need not bury directors in technical logs. It should preserve the source, test conditions, returned answer, authoritative value, materiality judgment, owner and action. That evidence supports prioritization, retesting and audit committee challenge without turning governance into an exercise in technical archaeology.

United States: existing controls meet a new distribution problem

For US reporting companies, Exchange Act Rules 13a-15 and 15d-15 require disclosure controls and procedures designed to ensure required information is recorded, processed, summarized and reported within SEC time periods. They also support communication to management for timely disclosure decisions. Section 302 certifications require principal executive and financial officers to certify fair presentation in all material respects and evaluate disclosure controls. Although these requirements do not mention AI answers, they provide the framework for responding when published financial meaning is repeatedly recovered incorrectly.

No SEC rule prescribes a machine-readability test for corporate PDFs or requires monitoring of third-party AI services. An inaccurate external answer is not automatically an issuer statement or securities-law breach. The enquiry concerns publication, materiality, control boundaries and management's response after discovery.

Disclosure controls and procedures remain distinct from internal control over financial reporting, or ICFR, even when the same financial information touches both. ICFR concerns reliable financial reporting and statements prepared under applicable accounting principles. Disclosure controls extend across information required within Exchange Act reports and processes supporting timely decisions. A defective website PDF does not establish an ICFR deficiency by itself, although repeated material inaccuracy still warrants investigation, ownership and a documented correction decision.

EDGAR provides the authoritative reference while sharpening the distinction among information states. Official filing documents generally use ASCII or HTML, with limited official PDF categories and an issuer PDF permitted as an unofficial attachment. Inline XBRL combines readable presentation with machine-readable data while preserving reporting context.

Investors may still encounter the issuer-hosted PDF first, while an AI service selects it through a separate index. A correct EDGAR filing helps reconcile the discrepancy without answering whether corporate controls should tolerate a materially different representation elsewhere.

Materiality should determine the intensity of management's response. A wrong period attached to operating profit deserves different treatment from a damaged decorative table, while recurrence after discovery deserves different treatment from an isolated historic defect. Finance leaders should establish where the discrepancy appears, whether the source is defective, how widely the answer travels and whether correction, communication, escalation or control redesign is required.

United Kingdom comparison

The UK regime offers a separate comparison because structured reporting and board-control expectations are explicit. FCA DTR 4.1 requires in-scope issuers to publish annual financial reports containing audited statements, a management report and responsibility statements. Relevant reports use XHTML with Inline XBRL marking for consolidated accounts, while the issuer remains responsible for information drawn up and made public under that section.

FRC work shows why rendered accuracy cannot finish the control discussion. Its 2025/26 review found accounting, scaling, consistency and website-availability errors despite correct human presentation. Those findings confirm regulatory scrutiny of machine-readable meaning.

The UK Corporate Governance Code 2024 provides a governance comparison rather than a universal statutory requirement. From financial years beginning on or after 1 January 2026, Provision 29 asks boards to monitor and review material financial, operational, reporting and compliance controls, then declare on their effectiveness. Its comply-or-explain approach emphasizes evidence, identified weaknesses and remediation after a defect becomes measurable.

European comparison

Within the European Union, ESEF requires in-scope annual financial reports to use XHTML and Inline XBRL for consolidated IFRS statements. ESMA's 2025 enforcement priorities emphasize management responsibility, audit committee oversight, consistency and controls, with enforcement where material misstatements arise. Those requirements govern the official structured report, not external AI answers.

European reporting rules already treat machine-readable accounting meaning as a regulatory concern. Corporate information can lose that meaning again when it moves from a controlled structured filing into a visually approved PDF distribution copy.

Financial impact arrives before enforcement

A regulator need not open a case before the company incurs a cost. An incorrect revenue, profit, leverage or cash figure can distort investor screening, alter analyst questions and enter media coverage. Finance and technology teams then absorb investigation, legal review, correction and stakeholder communication.

Confidence weakens when management cannot identify the correct public answer. Investors and counterparties question the company's reporting control while the wrong number circulates.

The CFO governs the impact while the CIO fixes delivery

A CIO can reasonably view a financial PDF as content owned by finance or investor relations. That boundary fails once workflows, font handling, storage, crawler permissions and extraction monitoring determine what machines recover. Those are technology delivery conditions, even when the content remains financial.

The technical delivery problem is for the CIO to solve. The financial impact is for the CFO to understand, assess and govern.

Google Search and Gemini are both Google services, but they do not share one information pipeline. Search can index an issuer's HTML or link to its PDF. Gemini applies separate access, source-selection and reconstruction steps before answering an investor's financial question.

AAAnow has observed crawler permissions preventing AI access to information available through Google Search, an oversized PDF affecting Gemini access, and flattened tables detaching values from their headings, periods or units. The approved figures remain unchanged, but the route used to recover them has changed. That is how the same organization can be represented differently across two services from the same provider.

For the CFO, two credible versions of reported performance are circulating at once. The CIO must establish which source each service reached, whether access controls intervened and whether extraction preserved the financial relationships. Neither executive can close that control issue alone.

Joint acceptance should test more than whether the approved number appears on the page. It should confirm the number remains connected to its period, unit, sign and accounting context after extraction across external services. Board oversight should focus upon material exceptions, ownership, remediation and residual exposure without treating each technical defect as a material control deficiency.

Internal AI inherits the same defect

The external problem already has an internal counterpart, with AAAnow observing organizations ingesting PDF estates containing flattened tables, missing characters and broken contextual relationships into AI and agentic systems. Defective information enters the retrieval layer, where agents summarize, combine and reuse it across later tasks, allowing errors to repeat and compound. Inaccurate outputs and failed deployments can begin with documents that passed human review years earlier. FRC research into AI within corporate reporting reinforces the need for data quality, accountability, validation and human oversight; the wider internal consequence belongs within the second article.

Practical responses already exist

AAAnow PDF Automation identifies, catalogues and profiles PDF estates, creates structured HTML versions of prioritized public documents and focuses original-file remediation upon content carrying the greatest financial value or exposure. Secure Tag helps AI systems identify authoritative corporate information, supports source verification and discovery, and controls how legacy or historic content is treated. Both capabilities operationalize decisions about priority, authority and remediation within a governed publishing process. They do not guarantee correct AI answers, remove regulatory exposure or create automatic legal compliance.

Governance should follow financial significance

Repairing an entire historical PDF estate is neither the starting point nor the measure of control quality. Management first needs to know which public documents carry material financial meaning, how machines recover them and where incorrect answers are already appearing.

A proportionate program should cover the following controls, applied according to financial significance and observed exposure:

• Inventory financially significant public documents, separating official filings from website reproductions.
• Classify likely external use, materiality and financial consequences before assigning remediation priority.
• Test reading order, table relationships, character extraction and font mapping against authoritative values.
• Reconcile structured filings, HTML pages and PDF copies after results and filing dates.
• Ask representative financial questions across relevant AI and search services, recording returned sources.
• Review crawler permissions and access restrictions alongside authoritative structured alternatives.
• Introduce a CFO-CIO publication acceptance control covering visible and machine-recovered meaning.
• Maintain an exception register with owners, deadlines, correction thresholds and escalation routes.
• Preserve evidence of assessment, decisions and remediation for audit committee challenge.

Historical content can then be prioritized rather than treated as one undifferentiated liability. Material documents receive testing and correction first, accepted exceptions remain visible, and control owners can explain why resources followed financial significance. Visual approval continues, but it no longer stands alone as proof that published meaning survived distribution.

Before the next audit committee meeting, the CFO and CIO should ask a live AI service two questions:

1. What were the company's latest reported revenue and operating profit, and how do those figures compare with the previous period?
2. Which figures in the latest results relate to continuing operations, and which authoritative source supports that answer?

Those answers will not determine regulatory liability, yet they will reveal whether the publication process is preserving financial meaning beyond the screen. If the service returns the equivalent of $315 million instead of $420 million, management has evidence that the approved report and the public answer have parted company. The control question then becomes immediate, governable and impossible to return to the production team without an owner.

Verified main article word count: 2,574 words, including the standfirst and section headings, excluding the headline, keywords, questions and answers, and appendix.

RESEARCH LEDGER

Appendix: Sources reviewed

The appendix records the creator, title, publication date, jurisdiction or evidence category, use status, direct link and relevance of each source opened during research. Links are shown as plain text references, not as clickable links.