> ## Content Index
> Fetch the complete content index at: https://fintekcafe.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI in Financial Reporting: What Controllers Can Automate in Filings and Disclosures, and What Auditors Will Ask
- URL: https://fintekcafe.com/ai-financial-reporting-disclosure-explained/
- Published: 2026-11-21T09:00:00.000Z
- Updated: 2026-11-21T09:00:00.000Z
- Description: AI in financial reporting for controllers: what to automate in filings and disclosures, what needs review, and the questions auditors will ask about AI.
- Author: FinTekCafe
- Tags: finance, reporting, ai, controls, seo

**AI in financial reporting means using models to check, tag, compare and draft the documents a company files with regulators and investors: the annual report, quarterly filings, footnotes, sustainability statements and responses to regulator comments. Most companies have started. KPMG's 2024 global survey of 1,800 companies found 72 percent piloting or using AI in financial reporting, but only about 10 percent had adopted it widely.** The position taken here: use AI to check the filing before using it to write the filing. Tie-outs, consistency checks and tagging are high-volume, verifiable work where a model's errors are caught by comparison to a source. Drafting narrative disclosure is where a fluent error reaches investors with the CEO's and CFO's certifications on it. Controllers who automate checking first build the evidence trail that auditors will ask for before drafting tools arrive at scale.

This is a business decision guide for finance leaders, not accounting, legal or audit advice.

![Adoption governance](https://fintekcafe.com/content/images/2026/11/inline-01-adoption-governance.png)

## Key Takeaways

- External reporting is the most scrutinized output a company produces, so AI enters it under the heaviest controls. The volume work (tie-out, tagging, benchmarking) is ready; narrative drafting needs a named preparer and reviewer.
- The vendors have shipped. Workiva launched a tie-out agent in July 2026, and all four large audit networks have put AI agents into their audit platforms between April 2025 and June 2026.
- Regulators have issued guidance and enforcement, not AI-specific rules. The PCAOB's 2024 generative AI publication is staff views, and COSO's February 2026 guidance applies the existing 2013 framework.
- The SEC's AI-washing cases show the disclosure risk runs both ways: overstating AI in the business, and failing to have disclosure controls over what is said about it.
- When AI sits inside a key control over financial reporting, the auditor will test it like any other control: design, evidence, change management and who reviews the output.

![Guidance not rules](https://fintekcafe.com/content/images/2026/11/inline-02-guidance-not-rules.png)

## Why Year-End Puts AI Under the Most Scrutiny

For calendar-year companies, the annual report is weeks away. It has the most reviewers and the least tolerance for error: the CEO and CFO certify it, the auditor opines on it and the audit committee approves it.

That is also why reporting teams want help. The work is repetitive in structure and unforgiving in detail: every number in the narrative must agree to the financial statements, every financial statement line must agree to the ledger, and every tagged value must agree to the face of the filing. Adoption in finance is broad but flat. Gartner's survey of 183 finance leaders, published 18 November 2025, found 59 percent using AI in the finance function, against 58 percent in 2024\. Error and anomaly detection was the third most common use case, at 34 percent.

The governance has not caught up with the tools. In the Center for Audit Quality's Spring 2026 Audit Partner Pulse Survey, released on 15 July 2026, nearly 90 percent of audit partners described AI governance at their largest client as developing or early stage, and CAQ's own analysis of the results put the share reporting advanced governance frameworks at only 6 percent.

![Task decision](https://fintekcafe.com/content/images/2026/11/inline-03-task-decision.png)

## What AI Does in External Reporting

### Tie-out and consistency checks

**A tie-out confirms that every number in a filing agrees to its source and to every other place it appears.** In a 10-K or annual report, the same revenue figure may appear a dozen times across MD&A, the financial statements, the footnotes and the segment tables. AI reads the whole document, finds each occurrence, compares it to the source schedule and lists variances.

This is the strongest use case because the check is verifiable: every flag points to two numbers that a person can compare. It is also where an AI error is least dangerous. A false flag wastes a reviewer's minute; a missed variance is no worse than the manual process it supplements. The upstream discipline matters as much as the tool, which is why tie-out automation works best in companies that have already moved toward [a continuous close supported by AI](https://fintekcafe.com/continuous-close-ai-finance-function/), where the ledger is reconciled before the filing draft starts.

### First drafts of MD&A and footnotes

Models can draft the variance explanations in management's discussion and analysis from the flux analysis and commentary that business units provide, and roll forward footnotes with updated numbers. The risk is fluent prose carrying an explanation nobody gave, or a number nobody checked.

### XBRL tagging

**Inline XBRL embeds machine-readable tags in the filing itself, so each reported value carries a label from the accounting taxonomy.** The SEC adopted mandatory Inline XBRL for operating companies in June 2018, phased in from fiscal periods ending on or after 15 June 2019 for the largest filers, and it covers the financial statements and cover page. Suggesting tags, checking consistency with prior periods and flagging extension elements is a pattern-matching job that AI does well, with a person approving changes.

### Peer disclosure benchmarking

Comparing a draft disclosure with how peers described the same issue, such as tariff exposure or a segment change, used to take a day of reading. DFIN launched its Active Intelligence suite on 19 November 2025, starting with comparisons of draft SEC filings against a filer's own prior filings and chosen peers. Workiva's Benchmarking Agent, launched in July 2026, does similar work.

### Sustainability report drafting

Sustainability reporting suits AI drafting, but the rules are moving. In the EU, the Omnibus I amendments, adopted by the European Parliament on 16 December 2025 and the Council on 24 February 2026, narrowed the Corporate Sustainability Reporting Directive to companies with more than 1,000 employees and more than 450 million euros in net turnover, according to summaries by De Brauw and Clifford Chance. The Commission adopted revised, simplified reporting standards on 3 July 2026, cutting mandatory datapoints by more than 60 percent according to Cooley and Herbert Smith Freehills summaries; the act was still under scrutiny by Parliament and Council at that point. In California, the first Scope 1 and 2 emissions reports under SB 253 were due 10 November 2026\. On 9 November 2026, the Ninth Circuit let that requirement proceed but found the Scope 3 and climate-risk (SB 261) requirements likely unconstitutional, in a preliminary ruling rather than a final judgment, as Crowell and Moring reported. Outside the EU and US, S&P Global counted 35 jurisdictions that had adopted the ISSB standards as of 22 July 2026\. AI drafts must follow this year's rule set, not the one in the model's training data.

### Comment-letter research and disclosure checklists

Researching SEC staff comments on a topic, or completing the disclosure checklist, is retrieval work over a large public corpus. KPMG said in April 2025 that it had released a Financial Report Analyzer to help its auditors complete disclosure checklists. Companies can use the same approach internally, provided the output is checked against the source letter or standard.

## What Vendors and the Audit Firms Now Offer

| Provider | Offering                                                                 | Date and source                                                                      | What it does                                                               |
| -------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------ | -------------------------------------------------------------------------- |
| Workiva  | Tie-Out, Benchmarking and Sustainability Disclosure agents; Agent Studio | 29 July 2026 (Business Wire); Agent Studio at Amplify, September 2026 (SiliconANGLE) | Flags variances, benchmarks disclosures, custom agents with human sign-off |
| DFIN     | Active Intelligence in ActiveDisclosure                                  | 19 November 2025 (DFIN press release)                                                | Compares draft filings with prior filings and peers                        |
| Deloitte | Agentic intelligence network in Omnia                                    | 24 June 2026 (Deloitte press release)                                                | Agents flag risks and analyze evidence for auditor review                  |
| EY       | Multi-agent framework in EY Canvas                                       | 7 April 2026 (EY press release)                                                      | Covers 130,000 assurance professionals and 160,000 engagements             |
| KPMG     | AI agents in KPMG Clara                                                  | 23 April 2025 (KPMG press release)                                                   | Agents for expense vouching and searching for unrecorded liabilities       |
| PwC      | Next-generation audit tools                                              | 27 October 2025 briefing (Accounting Today)                                          | AI-integrated end-to-end audit planned for calendar 2026 audits            |

These are vendor descriptions, not independent evaluations. The practical point for a controller: the auditor will use AI on the company's filing whether or not the company does, so inconsistencies that once survived a sampled review are more likely to surface.

## What Regulators and Standard Setters Have Said

The short version: there is guidance, enforcement and an inspection focus, but no AI-specific financial reporting rule.

### The PCAOB

On 22 July 2024, PCAOB staff published a Spotlight on generative AI in audits and financial reporting, based on outreach to firms and companies. It found that use in financial reporting appeared to be in early stages, and that firms expected AI to augment rather than replace people. The document says plainly that it represents staff views and is not a rule, policy or statement of the Board.

The binding change came through existing standards. The PCAOB adopted amendments to AS 1105 (Audit Evidence) and AS 2301 on 12 June 2024 addressing technology-assisted analysis; the SEC approved them on 20 August 2024, and they apply to audits of fiscal years beginning on or after 15 December 2025\. For calendar-year companies, the coming year-end audit is the first under them. The PCAOB has not issued an AI-specific auditing standard.

### The SEC

The SEC's AI activity is enforcement and advisory input, not rulemaking:

- **AI washing.** On 18 March 2024 the SEC settled its first AI-washing cases against two investment advisers, Delphia (225,000 dollars) and Global Predictions (175,000 dollars), for claims about AI they could not substantiate. On 14 January 2025 it settled with Presto Automation, described by law firms including DLA Piper as the first AI-washing case against a public company, over statements about an AI ordering product; the order also found Presto failed to maintain adequate disclosure controls, according to DLA Piper's summary. No civil penalty was imposed.
- **Priority.** In February 2025 the SEC created its Cyber and Emerging Technologies Unit, with AI-related fraud among its stated priorities. Gibson Dunn noted in a late-2025 review that the SEC under Chairman Atkins had yet to bring an AI-washing action against an investment adviser.
- **Disclosure review.** Debevoise reported in December 2025 that recent SEC comment letters signal the staff's view that companies should not suggest their AI is more autonomous, scalable or commercially mature than it is.
- **Advice, not rules.** On 4 December 2025 the SEC's Investor Advisory Committee recommended that the Commission issue AI disclosure guidance. It is advisory, and commentators, including the D&O Diary, noted that Chairman Atkins expressed skepticism.
- **Reporting frequency.** On 5 May 2026 the SEC proposed letting companies file semiannual reports instead of quarterly 10-Qs. Comments closed 6 July 2026\. It remains a proposal.

### COSO

On 23 February 2026, COSO published guidance on achieving effective internal control over generative AI. It does not create a new framework; it applies the five components of the 2013 Internal Control Integrated Framework to AI, sorts use cases into eight capability types including posting, judgment and monitoring, and recommends that management agree with internal and external auditors what counts as sufficient evidence for AI-enabled controls, as summarized by the Journal of Accountancy and Deloitte's DART.

## Where AI in Reporting Fails

### Invented numbers and citations

Models generate plausible text, and in a filing plausible is not enough. When Patronus AI released its FinanceBench benchmark in November 2023, GPT-4 Turbo with a retrieval system answered incorrectly or refused 81 percent of a 150-question sample about public-company filings. Models have improved, but the failure mode remains. In October 2025, Deloitte Australia agreed to partially refund a 440,000 Australian dollar government report that contained fabricated references; the revised report disclosed the use of a GPT-4o model, as the Associated Press reported. The general controls are covered in [the guide to AI hallucination controls](https://fintekcafe.com/ai-hallucination-controls-explained/); in reporting, the rule is simpler: no number enters a draft except by reference to a source schedule.

### Stale prior-period language

Rolling forward last year's footnote is where many reporting errors start: a debt facility that was refinanced, a contingency that was settled. AI rollforward makes this more likely, because the draft reads well. Every rolled-forward paragraph needs an owner who confirms the facts still hold.

### Inconsistency with the ledger

A model drafting MD&A from business-unit commentary may produce explanations that do not match the actual drivers in the ledger. Revenue "grew due to volume" when the ledger shows price. The fix is structural: drafts start from the reconciled flux analysis, not from free text.

### Material non-public information in external tools

A draft annual report before release is material non-public information. Pasting it into a consumer AI tool may expose it to a third party. Orrick warned in December 2023 that use of public generative AI systems creates a risk of disclosing material non-public information, and that Regulation FD can be violated even by accident. There is no SEC rule specific to AI here; the existing rules apply. This is the reporting-specific case of [shadow AI and how to govern it](https://fintekcafe.com/shadow-ai-enterprise-governance/): approved, enterprise-contracted tools with no training on company data, and a ban on everything else for draft filings.

## ICFR When AI Sits Inside a Control

Internal control over financial reporting (ICFR) is what management certifies under Sarbanes-Oxley Section 404, and what the auditor of an accelerated or large accelerated filer opines on. When AI performs or supports a key control, such as an automated tie-out replacing a manual review, it becomes part of that system.

Four practical consequences follow:

1. **The control owner stays human.** The AI performs a step; a named person owns the control, reviews exceptions and signs off.
2. **The review must be precise.** A reviewer who accepts every AI-cleared item without evidence of what was checked is a rubber stamp. Document what was examined.
3. **A model update is a change to the control.** If the vendor changes the model, the control may behave differently. Change management, re-testing after updates and version logging belong in the IT general controls.
4. **Completeness of input matters as much as output.** If the AI only saw part of the filing, its clean result means nothing. Evidence of completeness is part of the control.

Internal audit is usually the first function to test this. The approach is set out in [how internal audit audits the AI the business runs](https://fintekcafe.com/ai-internal-audit-explained/). The tax footnote is a special case: the income tax provision feeds it, and [what AI changes in the tax function](https://fintekcafe.com/ai-tax-function-explained/) explains why the data behind that number matters more than the drafting.

## The Decision: Task by Task

| Reporting task                              | Recommended mode        | Why                                                    | Control that makes it safe                                     |
| ------------------------------------------- | ----------------------- | ------------------------------------------------------ | -------------------------------------------------------------- |
| Tie-out of numbers across the filing        | Automate                | Verifiable against source; misses no worse than manual | Reviewer clears every flagged variance; completeness check     |
| Cross-reference and consistency checks      | Automate                | Mechanical comparison                                  | Exception log retained as evidence                             |
| XBRL tag suggestions                        | Automate, with approval | Pattern matching against taxonomy                      | Person approves new tags and extensions                        |
| Peer disclosure benchmarking                | Automate                | Research input, not filed text                         | Analyst confirms quotes against source filings                 |
| Comment-letter and standards research       | Preparer review         | Citations can be wrong                                 | Every citation opened and checked                              |
| MD&A variance narrative first draft         | Preparer review         | Fluent text can carry unchecked numbers                | Drafts built from reconciled flux; numbers linked to schedules |
| Footnote rollforward                        | Preparer review         | Stale facts survive rollforward                        | Owner confirms each fact still holds                           |
| Sustainability narrative first draft        | Preparer review         | Rules changing by jurisdiction                         | Drafted against the current year's rule set                    |
| Critical accounting estimates and judgments | Controller keeps        | Requires management judgment                           | Controller and CFO own; AI only for research                   |
| Non-GAAP measures and reconciliations       | Controller keeps        | Direct SEC enforcement exposure                        | Disclosure committee approval                                  |
| AI-related disclosures about the business   | Controller keeps        | AI-washing risk                                        | Legal and disclosure committee review                          |
| Final sign-off and certifications           | Controller keeps        | Personal certification by CEO and CFO                  | No delegation                                                  |

The pattern: automate where the output is checked against a source, review where the output is new text, and keep anything that is a judgment or a certification.

## What Auditors Will Ask

Expect these at year-end planning, and document the answers before fieldwork:

1. **Where is AI used in the financial reporting process?** An inventory: tool, task, owner, whether it sits in a key control.
2. **Which tools touch draft filings, and are they approved?** Contracts, data handling, and the policy that keeps material non-public information out of consumer tools.
3. **For each AI-supported control, who reviews the output and what evidence shows the review?** Exception logs, reviewer sign-off, and what was examined.
4. **How do you know the AI saw the complete population?** Input completeness evidence.
5. **What happens when the model changes?** Change management, re-testing and version records.
6. **How are AI-drafted numbers tied to the ledger?** The link from every number in a draft to its source schedule.
7. **How did you decide what the filing says about AI itself?** Disclosure committee minutes and support for each AI claim.

Companies that cannot answer these will find the auditor tests the affected controls manually, which removes the time saving.

## The Position, and Its Tradeoff

The position: check first, draft second. Fund AI for tie-out, consistency, tagging and benchmarking this year, document it as part of ICFR, and introduce drafting tools for MD&A and footnotes only once every drafted number links to a reconciled source.

The tradeoff is real. Checking tools are less visible. They do not give the CFO a first draft of MD&A on day three of close, and the reporting team may feel the slower path wastes the technology. Drafting tools would cut more hours sooner. But a drafting error in a filing is a disclosure problem with certifications attached, while a checking error is a missed flag in a process that already had manual review behind it. Starting with checks also builds the evidence and inventory auditors will request either way. The exception is a company with a mature, reconciled close and strong disclosure controls; it can pilot drafting in parallel, on lower-risk sections, with a preparer named on every paragraph.

## Frequently Asked Questions

### Can AI write an SEC filing?

AI can produce first drafts of sections such as MD&A variance explanations and footnote rollforwards, but it should not write a filing unsupervised. The CEO and CFO certify the report personally, and fluent AI text can carry numbers or explanations that do not match the ledger. Use AI for checking and drafting support, with a named preparer and reviewer for every section.

### Does the SEC or PCAOB have rules on AI in financial reporting?

Not AI-specific rules. The PCAOB's July 2024 generative AI Spotlight is staff observation, not a rule, and its binding change is the technology-assisted analysis amendment to AS 1105 and AS 2301, effective for fiscal years beginning on or after 15 December 2025\. The SEC has brought AI-washing enforcement cases and its Investor Advisory Committee recommended disclosure guidance in December 2025, but no AI disclosure rule exists.

### What happens to SOX controls when AI is used?

If AI performs or supports a key control over financial reporting, that control is in scope like any other. Auditors will look for a human control owner, evidence of review, completeness of the data the AI used, and change management when the model is updated. COSO's February 2026 guidance applies the existing 2013 framework to these cases.

### What is AI washing in financial disclosure?

AI washing is overstating how much a company uses AI or what its AI can do. The SEC's first cases, in March 2024, involved two investment advisers; a January 2025 case against Presto Automation, which law firms describe as the first against a public company, also found weak disclosure controls. Treat every statement about the company's AI as a disclosure that needs support.

## The Bottom Line

Financial reporting is where AI's value and its risk sit closest together. The volume work, tying out numbers, tagging and comparing, is ready for automation now and is verifiable by design. The narrative work is ready for drafting support, but only when every number traces to the ledger and every paragraph has an owner. Regulators have applied existing rules, not new AI rules, and the auditor will do the same. Controllers who can show an inventory, a reviewer and an evidence trail for every AI-supported step will get the time savings. Those who cannot will get more audit testing instead. This guide is a framework for that business decision, not accounting, legal or audit advice.