AI in corporate reporting: narrative is changing, financial statements are still human work
Published 14 August 2026
AI in corporate reporting is moving into the annual report, but not evenly. The narrative pages are changing first. The financial statements, where judgement, evidence and accountability are hardest to delegate, remain largely human work.
That is the central finding from the Financial Reporting Council’s July 2026 research into how UK companies are using artificial intelligence in corporate reporting. It is a useful correction to two competing assumptions: that AI is already producing whole annual reports, or that it has no meaningful role in external reporting at all.
The practical position sits between them. AI can help finance and reporting teams draft, compare and check narrative at speed. It cannot take responsibility for whether the final report is accurate, balanced and properly supported. That responsibility remains with management and the board.
What the FRC found about AI in corporate reporting
The FRC’s research found increasing but cautious use of AI. Adoption is uneven and generally concentrated in lower-risk, task-specific activities where the output can be reviewed by an experienced person.
Generative AI is proving most useful in narrative reporting. That can include producing a first draft from approved source material, comparing language across reporting periods, checking consistency between sections and helping teams manage the production process. These are real uses. They reduce time spent assembling and reviewing text without pretending that the model owns the conclusion.
Use in the financial statements themselves remains limited. That makes sense. Recognition, measurement, provisioning, impairment and going-concern judgements depend on evidence, accounting standards and context. The issue is not whether a model can produce plausible wording. It is whether the organisation can demonstrate how a material conclusion was reached and who challenged it.
The FRC also found that trust, data quality, governance controls and legal and reputational risk remain significant barriers. Investors continue to value authenticity, accuracy and accountability. A faster reporting process is useful. A report that feels generic, obscures management judgement or contains an unsupported statement is not.
Narrative drafting is not a low-control activity
It is tempting to classify narrative as the safe part of the annual report because it sits outside the primary statements. That is too relaxed.
The strategic report, risk disclosures, viability statement and performance commentary shape how investors understand the numbers. Language that overstates progress, weakens a risk disclosure or introduces inconsistency between the narrative and the accounts can be materially misleading even when every figure in the financial statements is correct.
AI adds a particular risk here. It produces fluent text quickly and confidently. That fluency can make weak analysis look finished. A paragraph can read well while subtly changing the meaning of the approved evidence behind it.
The control should therefore attach to the assertion, not the format. If the text contains a material claim about performance, risk, outlook or controls, the reviewer needs to trace that claim to an approved source. The fact that it appears in prose rather than a table does not reduce the evidential standard.
In practice, that means every AI-assisted reporting workflow needs three things: controlled source material, a named reviewer with the right subject knowledge and a record of the final human decision. Without those, the organisation has accelerated drafting but weakened control.
The disclosure-control questions a CFO should ask
The first question is not which model the reporting team uses. It is what the model is permitted to do.
Can it summarise approved board papers? Can it compare this year’s risk disclosures with last year’s? Can it propose wording for a section that will then go through the normal review process? Can it access unpublished results, personal data or commercially sensitive forecasts?
Those permissions should be explicit. A reporting team should not have to infer them from a general corporate AI policy written for every department.
The second question is provenance. For each material paragraph, can the reviewer identify the source information used to generate or support it? A model-generated statement that cannot be traced should not survive review simply because it sounds reasonable.
The third is consistency. Annual reports repeat themes across the strategic report, principal risks, sustainability disclosures, remuneration reporting and the financial review. AI can help identify contradictions, but it can also create them when separate teams use different prompts or source packs. A final cross-report consistency review remains necessary.
The fourth is change control. Models, retrieval systems and prompts change. A workflow tested during the half-year process may behave differently by year-end. The AI governance framework for finance functions should therefore cover reporting tools in the same way it covers transaction-processing tools: authorised access, review thresholds, error handling and validation after material changes.
Reviewer accountability cannot be delegated
The phrase “human in the loop” is too weak for external reporting. It describes presence, not responsibility.
A person can be in the loop and still provide little effective challenge. They can accept a draft because the deadline is close, because the language is polished or because they assume another reviewer will check the detail. That is not accountability. It is a hand-off with a human name attached.
Effective review needs a clear standard. The reviewer should know which assertions they own, what evidence they must inspect and what changes require escalation. Material narrative should have the same disciplined ownership as material numbers.
This is particularly important when AI is used to summarise large volumes of source material. The productivity gain comes from not reading every document in the same way as before. But someone still needs to test whether the summary omitted a contrary fact, lost an important qualification or gave too much weight to one source.
The right review design is risk-based. A low-risk drafting task can use sampling and standard checks. A statement about liquidity, viability, covenant headroom, a material control weakness or a principal risk needs direct review by the person accountable for that subject.
The board also needs a candid description of where AI was used. That does not mean labelling every sentence. It means the Audit Committee should understand which parts of the reporting process are AI-assisted, what controls apply and whether any significant issues arose during production.
Investor trust is the real constraint
The reporting question is often framed as efficiency: how many hours can AI remove from annual-report production? That is useful, but incomplete.
External reporting is a trust product. The annual report is valuable because investors believe the company has applied judgement, tested the evidence and accepted responsibility for what it says. If AI makes the document faster to produce but less specific, less candid or harder to evidence, the economics are poor.
Generic language is already a problem in corporate reporting. AI can make it worse by producing polished text that converges on the same safe phrases used by every other company. The result may pass a surface review while telling investors very little about the business.
The better use is to remove production friction so experienced people can spend more time on the parts that require judgement. Draft the first version faster. Compare disclosures more comprehensively. Identify inconsistencies earlier. Then use the time released to improve specificity, challenge and evidence.
That is how technology supports investor trust rather than diluting it.
A controlled reporting workflow for the next annual report
Start with an inventory. Identify every AI tool already being used in annual-report production, including general-purpose tools used informally by individual team members. Shadow use is still use.
Classify each use by risk. Formatting, document comparison and controlled summarisation are different from drafting viability language or interpreting a new accounting standard. Give each category an approved use, required reviewer and evidence standard.
Create a controlled source pack for each material section. Limit the model to approved documents where the tool allows it. Keep prompts and outputs for high-risk sections so the review trail can be reconstructed.
Add an AI-specific assertion check to the disclosure process. Ask what source supports the statement, whether important qualifications survived summarisation, whether it conflicts with another section and who approved the final wording.
Finally, report the outcome to the Audit Committee. The useful metrics are not number of prompts or pages drafted. They are hours removed from low-value production, exceptions found, material corrections made and any control failures identified.
AI will write more annual-report narrative. That is already happening. The standard that matters is whether the report remains specific, evidenced and accountable when it does.
For the wider control framework, read AI governance for finance functions and the ethical bypass problem in AI. If your reporting process is changing alongside the rest of the finance function, work with me to discuss the leadership and transformation required.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader.