Use AI to Focus Supervision, Not to Replace It
The central subject of this series is the BIS procedure that compares capital rules with AT1 prospectuses and ranks possible divergences for expert attention. The evidence supports a useful but bounded conclusion: a language model can help supervisors find and order passages whose contractual wording may change a rule’s operation, while supervisory and legal experts must still decide whether the difference is permitted, material, actionable, or legally consequential. This final synthesis turns that distinction into an adoption standard. It asks what the procedure has demonstrated, what controls make its use defensible, and where the evidence still requires restraint.
What the BIS procedure can legitimately do
The procedure addresses a real supervisory bottleneck. A regulatory package states the conditions an instrument must satisfy; a prospectus states the contractual terms on which it is issued. Those texts are expected to differ because a contract adds detail and expresses regulatory requirements in a different form. The hard task is to notice when wording changes the scope, condition, sequence, or consequence of a requirement rather than merely restating it.
AT1 instruments make that task more than a document-management exercise. They are perpetual securities that can pay discretionary, non-cumulative coupons in ordinary conditions and absorb losses under stress. Mechanical triggers determine when a capital threshold is reached. Point-of-non-viability provisions describe when an authority can require write-down or conversion. Coupon terms govern cancellation, and hierarchy determines whether common equity or AT1 bears losses first. A small qualification in one of these provisions can remain unimportant during normal trading and become central when a bank approaches failure or an authority changes the allocation of losses.
The BIS procedure is appropriately narrow in response. It supplies the LLM with the relevant regulatory documents, the prospectus, and a structured prompt. The model identifies provisions addressing the same requirement, quotes the relevant passages, flags wording that may alter scope, conditions, or consequences, and classifies the possible relationship as narrowing, expansion, omission, ambiguity, or contradiction. The result is not a compliance score detached from its sources. It is a candidate record that an expert can inspect.
That record has three layers, and each should retain its own status. The paired passages are evidence about what the texts say. The divergence label is an analytical description of how the passages may differ. Recurrence across independent model runs is a signal about how consistently the procedure returned an equivalent candidate. None of those layers is the final legal or supervisory conclusion. A reviewer still has to determine whether the passages address the same requirement, whether the apparent difference is permitted elsewhere in the framework or contract, whether it is material to the supervisory mandate, and what action is authorised.
The ranking function is nevertheless significant. Supervisory expertise is scarce, and a reviewer who starts with an undifferentiated mass of rules and prospectuses must spend time both locating possible issues and interpreting them. A queue changes the starting point. It places recurrent clause pairs and their source passages earlier in the reviewer’s field of attention. That is a meaningful operational benefit even though it is not adjudication.
The benefit also creates a responsibility. What is placed near the top is more likely to be inspected, and what is placed low may wait. Ranking is therefore agenda-setting power. It must be evaluated not only by whether its candidates sound plausible but by whether it improves the allocation of expert time without hiding material differences. A repeated candidate may turn out to be permitted or immaterial. A candidate returned once, or not returned at all, may still matter in an unusual fact pattern. A shorter queue is not automatically a safer queue.
The BIS results show the shape of that promise. In the exercise covering AT1 instruments from 10 European G-SIBs, the procedure produced 576 individual divergences. After equivalent findings were merged, 193 distinct candidates remained, and 79 appeared in at least four of the five independent runs. Mechanical triggers and PONV provisions accounted for roughly two thirds of the retained high-recurrence candidates. The mix of divergence types varied by jurisdiction, with narrowings most common in Switzerland, expansions in the UK, and omissions in the EU.
Those figures describe the queue produced by that exercise. They do not measure the prevalence of divergence across all banks or legal systems, and they do not show that 79 instruments breached a rule. The jurisdictional sample sizes differ, and the candidates were not presented as independently adjudicated legal findings. The numbers are useful because they show how a screening process can concentrate attention on provisions governing loss absorption. They are not evidence for a compliance league table or a universal recurrence threshold.
The historical tests make the same point through cases. The procedure recovered a Credit Suisse narrowing and a Yes Bank contradiction in anonymised documents, and the findings disappeared when the relevant regulatory and prospectus wording was separately modified to agree. That supports the interpretation that the procedure responded to the supplied textual relationship rather than merely reproducing a known case. It does not establish a general recall rate, a production false-negative rate, or legal correctness across other documents.