In October 2023 HDI Global SE sent the Prudential Regulation Authority (PRA) revised figures for its FSCS liabilities: what its policyholders could claim from the Financial Services Compensation Scheme, which pays out when a firm fails. For the end of 2022 it gave £20.39m. A month later, after taking legal advice, it put the figure at £195.85m. The difference was compulsory insurance that its revisions had left out.1PRA, Final Notice to HDI Global SE, 16 July 2026, paragraphs 2.2, 2.6 and 2.9, and Annex A 1.2, 2.1, 4.3.1, 5.9, 6.2 and 6.3.
It was not one slip. On “at least eleven occasions between February 2022 and June 2024”, the UK branch of an insurer headquartered in Hanover sent materially wrong figures to the PRA, or to the Financial Conduct Authority (FCA), which collects the data used to set each firm’s share of the scheme’s levy. Some were corrections of earlier errors. The PRA fined it £4,165,000, after a 30% discount for settling; the penalty would otherwise have been £5,950,000.2Final Notice, paragraphs 1.1, 1.2 and 2.1, and Annex A 4.2. The PRA announced the penalty on 20 July 2026.
The PRA did not blame bad arithmetic. It said the errors “arose and persisted undetected because the Firm did not have effective written processes”, and that there was “a lack of clear accountability” for getting the figures right.3Final Notice, paragraphs 2.8 and 2.11.2. That suggests a simple test for any regulatory return: who owns it, what written method it follows, and who checks that method against the rule.
The fix that needed fixing
The £20.39m was not HDI’s first attempt at a fix. In summer 2023, answering a PRA request, it decided it had over-reported. It believed the scheme covered only policyholders with turnover under £1m. So it took its employers’ liability business out of the figures and resubmitted.4Final Notice, Annex A 5.4 and 5.5.
That was wrong. Employers’ liability and third-party motor cover are compulsory insurance, and the FSCS protects compulsory insurance claims in full. The original figures had not been too high at all. The firm had looked at the FSCS website, but there is no evidence that anyone involved consulted the PRA Rulebook before October 2023. Nor did it take advice before resubmitting.5Final Notice, Annex A 5.6, 5.7 and 5.11, and Annex B 3.2.3. It found the mistake in October 2023, after reading the rulebook, and sent the £195.85m figure in November.
It then missed a chance to catch a second, older error. Premium income for the levy should have been reported gross, before deducting the share passed on to reinsurers. HDI had long reported it net. In January 2024 someone asked about exactly this. “Despite being unsure what should be included”, the firm took no advice and kept to net. An independent adviser, appointed at the PRA’s request, found the error in June 2024.6Final Notice, Annex A 4.4.2, 6.5 to 6.8, and Annex B 3.2.3. Its 2022-23 premium income, first reported as £2,817,575, stood at £96,439,167 after all the corrections.7Final Notice, Annex A 4.3.2.
Everyone’s figure, nobody’s job
Behind both errors was the gap the PRA named. HDI had no written policy or process for calculating these figures until August 2024, and its senior managers knew it.8Final Notice, Annex A 3.2. The only document the firm could find, a note on how to submit the return, cited no rule and said nothing about method. Finance did the sums and, until September 2023, signed them off alone. Compliance then added its sign-off, but this “remained an informal and undocumented process”. The notice is blunt: “Neither the Finance nor Compliance function had clear responsibility” for getting the figures right.9Final Notice, Annex A 3.3 to 3.5.
The adviser’s findings read like a checklist in reverse: “no traceability to rule requirements”, challenge that went unrecorded, and no clear doer, checker and reviewer control documented. Its first recommendations were to name an owner for understanding the rules, and an owner for the accuracy of the return.10Final Notice, Annex A 6.9.2 and 6.10. HDI has since acted on them, and the PRA credited it for what the notice calls “a fulsome and candid account”.11Final Notice, paragraphs 1.2 and 2.10, and Annex C 18.1.
Same gaps, different rulebook
HDI is an insurer supervised by the PRA. Two recent FCA cases show the same gaps at investment firms.
Sigma Broking, a futures and options broker, was fined £1,087,300 in July 2025. Transaction reports are the trade-by-trade records firms send the FCA to help it detect market abuse. Sigma’s were incomplete or inaccurate for 924,584 transactions over five years, “close to 100% of reportable transactions”. The cause was “the reporting system initially having been set up incorrectly and these set up errors not having been corrected”. 12FCA, Final Notice to Sigma Broking Limited, 29 July 2025, paragraphs 1.1, 2.1, 2.2, 2.4, 2.5 and 4.4.
Infinox Capital’s case was smaller. It was fined £99,200 in January 2025 for 46,053 transaction reports missing over six months from its new business in contracts for difference, which pay the change in an asset’s price without the buyer owning it. “It was the responsibility of a single individual at Infinox to manually identify which financial instruments were reportable when new business was commenced”. Nobody checked that work until a third-party review.13FCA, Final Notice to Infinox Capital Limited, 27 January 2025, paragraphs 1.1, 2.1, 3.1 and 4.12. The breach found was of Article 26(1) of MiFIR, the transaction reporting rule, alone.
The rules the regulators applied are worded alike. The FCA found Sigma in breach of Principle 3: “A firm must take reasonable care to organise and control its affairs responsibly and effectively, with adequate risk management systems.”14FCA Handbook, PRIN, 2.1.1, as last updated in June 2026; Sigma Final Notice, 5.2. HDI broke two of the PRA’s Fundamental Rules, its high-level rules for every firm it regulates. One, Rule 6, reads: “A firm must organise and control its affairs responsibly and effectively.”15PRA Rulebook, Fundamental Rules, 2.2 and 2.6; Final Notice, paragraph 2.11. The FCA and PRA texts are close in aim but not identical; the FCA’s adds reasonable care and risk management systems.
Principle 3 does not demand perfection. The FCA’s guidance says a firm is not in breach “simply because it failed to control or prevent unforeseeable risks”.16PRIN, 1.1.7. We found no equivalent in the PRA’s Fundamental Rules, but in our reading these errors were foreseeable: the PRA called the full protection of compulsory insurance claims “clear and obvious”.17Final Notice, Annex B 3.2.2.
Foreseeable errors, then, with a common pattern behind them, in our reading rather than the regulators’: one person or one unwritten method, a weak second check or none, and errors that ran for months or years, found late, after a regulator’s question or an outside review.
Rules as code, owners as people
Could software close these gaps? None of the three notices mentions AI or automation as a remedy, so what follows is our reading, not the regulators’. In short: code for the calculation, a model at most for a first sort of the inputs, and a named person who decides.
The obvious idea, having a model draft the return, helps little. It could produce a confident figure, but not an owner, a written method or a reviewer who knows the rule. HDI’s calculations were “reviewed, signed off and submitted without any effective challenge of the methodology”, including a figure about a tenth of the right one.18Final Notice, Annex A 4.3.1 and 5.11.
The better use of software, in our view, is to write the method itself as code. Rules like these change less often than the processes around them, and much of a calculation like HDI’s comes down to definitions and arithmetic. Whether premium income is counted gross or net of reinsurance, for instance, is fixed by a definition in the rules.19Final Notice, Annex A 6.8. Code can follow such definitions step by step, name the rule paragraph behind each, and be tested against worked examples. HDI’s net-versus-gross choice could then have sat in one visible line, open to anyone comparing it with the rule.
Recent research points the same way, towards keeping the model out of the calculation itself, with limits. In a July 2026 preprint, a paper not yet peer-reviewed, researchers at Aethis, a firm that sells such a system, tested the most capable current models on rules with exceptions nested three deep: “A is required UNLESS B applies, UNLESS C overrides B”. Several made confident errors. By April 2026 most of those errors had gone, but under the same model names and with no announced change. The authors’ point is that accuracy can move without notice, not that models stay wrong, and they warn against reading their findings beyond nested exceptions. Their design has a model draft the rules and a fixed engine, one that gives the same answer every time, apply them. They are candid about the limit: “The system guarantees correctness of execution given a correct specification; it does not guarantee correctness of the specification itself.”20Paul Simpson, John Kozak and Lisa Doake, Confidently Wrong: Exception Chain Collapse in Frontier LLM Rule Evaluation, arXiv preprint, July 2026, abstract and sections 4.2, 5.5, 6.3 and 10.1. The authors work for Aethis, which develops the system tested, and the benchmark used structured inputs rather than documents.
The harder part is sorting the inputs. Is this policy compulsory insurance? Is this instrument reportable? HDI’s compulsory-insurance error and the control Infinox lacked were both problems of this kind, and this is where a model is least to be trusted on its own. A model held to a procedure could still help with a first pass: propose a classification, quote the rule paragraph it relied on, and leave the decision to a named person. A model can, in our view, quote a plausible but wrong paragraph, so that person must check the quotation, not just the answer.
The nearest evidence is a 2026 preprint by researchers at Anhalt University on the Bundesbank’s checks of securities offered as collateral. General-purpose models pulled the relevant facts from long prospectuses, and fixed rules in code then tested them against six eligibility criteria. About 90% of the securities the system called eligible were, but it wrongly rejected about one in five that were eligible. Its authors read the design as cautious: any doubtful criterion makes a document ineligible, which they describe as flagging it for human review. It is one study of a related task, and they say hard data on their automatic evaluation is still missing.21Serhii Hamotskyi, Akash Kumar Gautam and Christian Hänig, LLM-Based Examination of Eligibility Criteria from Securities Prospectuses at the German Central Bank, arXiv preprint, 2026, Table 2 and sections 5, 7.2 and 8. Precision was 0.90 to 0.91 and recall 0.80 to 0.82, depending on the model.
Neither code nor a model removes the owner. In 2019 the PRA noted that firms had interpreted its rules over time, and “in many cases these have been hard coded into firms’ systems”. Firms might need to find and validate them.22PRA, Reliability of regulatory returns, letter to chief executives, 31 October 2019. In 2021 it added that where such interpretations may no longer be suitable, firms should identify, validate and correct them where appropriate.23PRA, Thematic findings on the reliability of regulatory reporting, letter to chief executives, 10 September 2021, section 1. Both letters were written for banks, and the first is seven years old. The point travels: code applies someone’s reading of a rule to every record, right or wrong. And rules do change, which is why HDI’s adviser recommended regular reviews of the regulation.24Final Notice, Annex A 6.10.5.
Accountability does not move either. In all three cases the penalty fell on the firm.25HDI Final Notice, 1.1; Sigma Final Notice, 1.1; Infinox Final Notice, 1.1. That is the argument of The expensive coin flip: the checking falls on the firm. In our reading, code and models can make that checking cheaper and better recorded, but not optional.
Why not just read the rulebook?
The first objection is that HDI is an outlier. Because the firm could not show it had checked the rulebook for its levy data before summer 2023, the PRA said its failures “reflect a complete lack of engagement rather than a misunderstanding resulting from complexity”.26Final Notice, Annex B 4.3. Sigma and Infinox are the answer. Both had been subject to their reporting rules since January 2018, and both knew the obligation.27Sigma Final Notice, 4.5; Infinox Final Notice, 4.4. Their errors came from a set-up nobody revisited and a judgement nobody checked.
The second is that none of this needs software. A written procedure, a named owner and a reviewer who read the rulebook would have caught HDI’s errors. That is fair, and for a small return it may be the whole answer.
But in our view a procedure that lives in people is harder to review than one that lives in code. Code can be reviewed only if someone does it, as Sigma shows, but every step is on the page and every change leaves a trace. An organisation’s way of working often is not written anywhere. A handover set up years ago outlives the reason for it, and nobody can say which step matters until it is removed. HDI’s own Compliance sign-off, added in 2023, stayed informal and undocumented, so there was nothing to review. In 2021 the PRA described something similar at banks: responsibilities dispersed across teams, in processes so fragmented that there was “a poor understanding and documentation of the entire process”.28HDI Final Notice, Annex A 3.4; PRA, Thematic findings on the reliability of regulatory reporting, section 1. Changing such a process means first finding out what it is.
Where software earns its place, in our reading, is consistency and memory: the same checked method every period, with a record of what it did. The case for a model is narrower, helping a reviewer spot misclassified records at volume. Neither is worth building before the owner and the method exist. Otherwise the software does what Sigma’s reporting system did: it applies set-up errors nobody corrected to nearly every record, for years.29Sigma Final Notice, 2.4 and 2.5.
Start with a name
This week, pick one regulatory return your firm submits. On one page, write down who owns its accuracy; the method, with the rule paragraphs it rests on; and who checks the method, not just the total. If a line is blank, or names a department rather than a person, you have found what the PRA found at HDI. Only then ask what to automate.
The rules were published all along. What was missing, at all three firms, was someone whose job it was to check the work against them.
This is commentary on published sources, as they stood on 13 November 2026, not legal or regulatory advice.
Drafted with AI assistance and reviewed by Ilja Orlovs before publication.