Part of the AI in finance series.
The Financial Reporting Council commissioned academics at Lancaster and Loughborough universities to study how listed companies are using AI in the annual reporting process. The resulting research, published 9 September 2026 and built on 39 interviews with senior members of reporting and investor relations teams, is the most grounded account available of what actually works - and where the boundaries sit. This article draws solely on that research. The FRC has issued no guidance on AI in annual reporting.
What did the FRC research actually find?
Every use case the research identifies as working falls into one of three activities: reading long documents faster, comparing disclosures across peers or periods, and checking narrative for consistency and tone. The research presents three case studies: LLM tools used to navigate long documents, research technical accounting issues and analyse peer disclosures; AI applied to narrative consistency tracking and sentiment analysis across disclosures; and anomaly detection with diagnostic suggestions introduced earlier in the reporting cycle. In no case does the model draft numbers. The pattern is consistent - AI is useful when it is reading, comparing or flagging; it is not in the drafting seat.
What are the governance lines the research sets out?
The research is explicit on accountability. It states that 'ultimately directors are responsible for the output of the annual reporting process' and that 'AI should support, not replace, human judgement.' On checking, it describes 'proportionate checking based on materiality and risk' as the operating principle - meaning the intensity of human review should scale with what is at stake in each section, not be applied uniformly or omitted because a model has already run over the text. These three lines are the governance architecture the research offers, and they are enough to frame a practical operating model for a group finance team.
Where does this fit with the FRC's wider expectations in September 2026?
The FRC's annual review published the same month asks companies to ensure robust review processes are in place to identify common technical compliance issues in their reports. AI used as a review layer - reading, flagging anomalies, surfacing inconsistencies - fits that expectation in shape. It is a fit of principle, not a claim that AI catches specific errors; human reviewers remain responsible for every conclusion. The point for a group CFO is that investing in structured AI-assisted review ahead of final sign-off is consistent with the direction the FRC is signalling, even in the absence of formal guidance on AI specifically.
What is the model actually given - and who checks the output?
The research describes three distinct input contexts. For document navigation and technical research, the model is given the draft annual report, prior-year comparatives and selected peer filings. For narrative consistency and sentiment work, it is given successive drafts of the same report, or the narrative sections across multiple filings in a peer set. For anomaly detection, it is given financial data and disclosure text earlier in the cycle, before the report reaches its final form. In each case the output is a flag, a question or a comparison - not a conclusion. A human reviewer, typically a senior member of the reporting or investor relations team, reads the output and decides what to do with it.
Who checks matters as much as what is checked. The research draws on senior practitioners, and the implication is clear: AI output in the annual reporting context is not a task for junior staff to clear unreviewed. The model surfaces candidate issues; experienced judgement resolves them.
How should checking be scaled to materiality and risk?
The research's phrase 'proportionate checking based on materiality and risk' is the practical answer to the question of how much human review is enough. For a group with complex consolidation, significant estimates or new disclosure requirements - CSRD being the obvious current example - the materiality threshold for intensive review is lower, meaning more of the AI output warrants close attention. For boilerplate sections with low estimation uncertainty, lighter review is defensible. The CFO's role is to set that calibration explicitly, not leave it to individual reviewers to decide ad hoc. A written protocol that maps disclosure sections to review intensity, informed by the group's own materiality framework, is the practical output of applying this principle.
What does this mean for a group CFO thinking about year-end preparation?
The research supports a specific sequencing: AI belongs in the review layer, ahead of final drafting sign-off, where it can read the full document as an integrated whole and surface things that section-by-section human review misses - tonal shifts, factual inconsistencies between the strategic report and the financial statements, or disclosures that read differently against peer practice. That is a structural role in the year-end process, not an add-on. It also implies that the data environment needs to be ready before the tool is useful: the model needs access to a clean, complete draft, prior-year comparatives and, where peer analysis is the goal, a curated set of comparable filings.
The three use cases in the research are self-contained enough that a group can adopt one without the others. Anomaly detection earlier in the cycle is the lowest-friction starting point for most groups because it does not require the full draft to be complete. Narrative consistency checking is highest value for groups with long, multi-author reports where drift between sections is a known risk. Peer disclosure analysis is most useful where the group faces novel or contested disclosures and wants to understand market practice quickly.
This article covers the external annual report only. For the management board pack, see our piece on AI and variance commentary in the board pack. For audit trail considerations, see AI governance and the finance audit trail. For close readiness before annual reporting begins, see why AI close pilots stall.
What should a group CFO do before deploying any of these tools?
Three practical steps follow from the research. First, define the input perimeter: which documents, which vintage, which peer set. Second, write down who reviews the output and at what seniority, and map that to the group's materiality framework so that review intensity is not left to individual discretion. Third, document the process - what the model was given, what it flagged, and what the human reviewer concluded. The research's governance principle is that directors remain responsible for the output; a written record is how that responsibility is discharged in practice, not merely asserted.
Common questions
What does the FRC's 2026 commissioned research say about AI in annual reporting?
Research commissioned by the FRC from Lancaster and Loughborough universities, published 9 September 2026 and drawing on 39 interviews with senior reporting and investor relations professionals, identifies three working use cases: LLM tools for navigating long documents and analysing peer disclosures; AI for narrative consistency tracking and sentiment analysis; and anomaly detection with diagnostic suggestions earlier in the reporting cycle. Every use case involves reading, comparing or checking - none involves the model drafting numbers.
Has the FRC issued guidance on using AI in annual reports?
The FRC has issued no guidance on AI in annual reporting as of September 2026. The research published that month is commissioned academic work, not FRC guidance, policy or a statement of regulatory expectation. Group CFOs should treat it as evidence of what works in practice, assessed against their own governance frameworks.
Who is responsible for the annual report when AI tools are used in its preparation?
The FRC's commissioned research states plainly that 'ultimately directors are responsible for the output of the annual reporting process.' Using AI in the review layer does not transfer or dilute that responsibility. The research also states that 'AI should support, not replace, human judgement,' which means every output the model produces requires human review before any conclusion is drawn.
How should human checking of AI output be scaled across the annual report?
The FRC research describes 'proportionate checking based on materiality and risk' as the governing principle. In practice this means sections with higher estimation uncertainty, greater investor focus or new disclosure requirements warrant more intensive human review of AI-flagged items. A written protocol mapping sections to review intensity, aligned to the group's own materiality framework, is the recommended way to apply this principle consistently.
Where in the year-end process should AI tools sit for annual reporting?
The evidence base from the FRC's commissioned research points to AI being most useful in the review layer - after drafts are substantially complete but before final sign-off - where it can read the document as an integrated whole and surface inconsistencies, tonal shifts or anomalies that section-by-section human review can miss. Anomaly detection can also be introduced earlier in the cycle, before the full draft is complete, to surface issues while there is still time to address them.