<- Take me back to my Company Secretaries (UK) overview
Executive Summary
Company Secretaries supporting UK board oversight of Consumer Duty compliance prepare the annual board report mandated by PRIN 2A.8.3R, the cross-cutting rules summary for board papers, and the regulatory horizon-scan for non-executive director briefings. Across nine findings in this cell, frontier AI models tested produced confidently wrong reconstructions of the FCA's text: a multi-condition safe harbour at PRIN 2A.2, a binding PRIN 2A.5.10R citation imposed where FG22/5 guidance lives, a reversed group insurance scope exclusion, an inverted fair-value methodology, an invented FS25/2 supervisory timeline reproduced across multiple questions, and an evasion response combined with a fabricated Clifford Chance citation.
Each of these is the kind of detail that would be included in a board pack 'rule summary' and become the firm's working board-level understanding until a director's challenge or an FCA interaction exposes the discrepancy.
How AI gets this regulation wrong
The findings in this cell are inference drift and rule-misstatement, not refusal. The models committed to specific board-paper-ready answers where the correct posture would have been to surface the actual FCA text or to flag that the model could not locate authoritative text. Instead, the models produced content with the structural features of board-paper regulatory analysis, where the underlying claims were either fabricated or directly contradicted by the FCA's published text.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Misstated Rule | 1 | Finding#1 |
| Inference Drift | 1 | Finding#2 |
| Inference Drift | 1 | Finding#3 |
| Inference Drift | 1 | Finding#4 |
| Inference Drift | 1 | Finding#5 |
| Inference Drift | 1 | Finding#6 |
| Misstated Rule | 1 | Finding#7 |
| Inference Drift | 1 | Finding#8 |
| Inference Drift | 1 | Finding#9 |
What that means for your practice
For Company Secretaries, the nine findings cluster on governance-grade exposure: board papers that import the AI's framing carry forward the wrong understanding of what the Duty requires, what its scope is, and what the FCA's recent supervisory record actually says. The annual board report under PRIN 2A.8.3R is the most acute exposure surface: a report that recites the model's multi-condition safe harbour, its rule-versus-guidance confusion, or its fabricated supervisory timeline would be challenged on review.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Regulatory enforcement / professional liability exposure | 5 | Finding#1 · Finding#2 · Finding#3 · Finding#4 · Finding#7 |
| Operational decisions based on a fabricated regulator record | 4 | Finding#5 · Finding#6 · Finding#8 · Finding#9 |
When this affects Company Secretaries
Company Secretaries supporting UK board oversight of the Consumer Duty encounter the framework continuously: preparing the annual Duty board report under PRIN 2A.8.3R, drafting board-pack rule summaries on the cross-cutting obligations and four outcomes, supporting board engagement with the FCA's recent supervisory record (Dear CEO letters, multi-firm reports, feedback statements), and maintaining the firm's Duty governance map across product lines, distribution chains, and customer categories.
Each of these workstreams puts the Company Secretary in the position of summarising specific FCA text for the board, and each is increasingly supported by AI-assisted research at the drafting stage. The findings in this cell map onto exactly the kinds of summaries that would appear in a board pack. The foreseeable-harm safe harbour (Finding#2) is the kind of rule precis a board would expect to see; the model's multi-condition reconstruction sets the wrong baseline. The PRIN 2A.5 versus FG22/5 confusion (Finding#3) is exactly the kind of detail a non-executive director with a regulatory background would catch on review.
The fair-value methodology inversion (Finding#4) would, if it reached a board pack, lead the board to approve a methodology that exceeds what FG22/5 requires.
The FS25/2 findings (Findings#5, 6, 8, 9) are the most reputationally damaging if imported: a board paper that records 'April and August 2025 Dear CEO letter withdrawals' as the FCA's record contradicts a publicly available feedback statement and undermines the paper's authority. The Clifford Chance fabricated citation in Finding#9 is the cleanest example of a structural failure mode that the Company Secretary's review process should be designed to catch.
The findings at a glance
The table below lists each finding from the Consumer Duty testing in this cell, showing the question area, the AI's failure mode, and the citation identifier for the underlying finding record.
Aggregate impact
The nine findings in this cell describe a pattern of model behaviour that the Company Secretary's review process should specifically anticipate: confident, board-paper-ready reconstructions of FCA text that the regulator's actual publications contradict. The pattern is consistent across both Opus 4.7 and Sonnet 4.6 with web search, and consistent across the cross-cutting rules (Finding#2), the four-outcomes structure (Finding#3), the fair-value methodology (Finding#4), the scope exclusions (Finding#7), and the supervisory-letter record (Findings#5, 6, 8, 9).
For Company Secretaries, the implication is structural rather than tactical. AI-assisted summaries of FCA text are not safe to import into a board pack without source-text verification, because the failure mode is not a random slip, it is a consistent over-confident reconstruction that has the surface features of a careful regulatory precis. The audit trail of the firm's Duty governance becomes the durable record of how the board engaged with the framework, and importing fabrications into that record undermines the firm's ability to demonstrate good governance under PRIN 2A.
The forward-looking implication is also significant. The FS25/2 fabrications appear under multiple differently framed questions, which suggests the model has an internalised but wrong account of the supervisory record. This is the kind of internal representation that will reproduce across subsequent questions on the same topic, so a board pack that imports the AI's framing on one question is likely to import the same fabrication on adjacent questions in the next cycle.
What your team should do
Company Secretaries should treat AI tools as a research-orientation aid for Consumer Duty board work, not as a source of board-paper text. Any output that summarises a PRIN 2A provision, characterises the FCA's scope position, or recites figures from a feedback statement requires direct verification against the FCA Handbook or the published feedback statement before it can appear in a board pack.
For practical safeguards on the annual Duty board report under PRIN 2A.8.3R: (a) every PRIN 2A citation in the draft report should be cross-referenced against the FCA Handbook before the report goes to the board. (b) Every figure from an FCA publication (number of withdrawn letters, dates of supervisory actions, quantitative thresholds) should be confirmed against the underlying PDF. (c) Every scope statement (which product lines are in or out of the Duty's perimeter) should be confirmed against PRIN 2A.1.8R or the corresponding source provision.
Where AI tools are most safely used in this practice area: drafting the structure of the annual board report, identifying which Duty workstreams should be covered for a particular product mix, surfacing cross-references between Duty obligations and adjacent FCA expectations, and producing first-draft summaries for review against the source text. The risk concentrates in the next step, where the AI is asked to specify the actual rule text, the applicable scope, or the FCA's recent supervisory record. At that point the source document is the only reliable input.
How RLB Can Help
RegLeg's published Hallucination Research gives UK Company Secretaries a free pre-flight check before relying on AI tools for Consumer Duty board work. Before the annual Duty board report under PRIN 2A.8.3R is finalised, the research identifies which areas of the Duty, the cross-cutting rules, the four outcomes, the scope exclusions in PRIN 2A.1.8R, the fair-value methodology in FG22/5, and recent FCA feedback statements, have historically generated confident but incorrect AI output that would mislead the board on review.
Beyond the published research, RegLeg works with UK boards on bespoke deep-dives that map AI-supported governance workflows to their actual hallucination exposure. The deep-dive identifies which board-pack workstreams (annual Duty report, supervisory horizon-scan, scope mapping across product lines) warrant additional controls or independent verification steps, and supports the Company Secretary's role in ensuring board papers do not import undetected regulatory misstatements. RegLeg also offers training and CPD-aligned content tailored to the UK Company Secretary context.
For teams that want to build durable in-house capability, RegLeg develops material covering how to interpret AI-generated regulatory summaries critically, how to structure escalation where AI confidence is high but human verification is essential, and how to document AI-assisted board-paper drafting consistently with good governance standards under PRIN 2A.