Gerd-Jan van Wiggen: The clause that only comes into play when things go wrong
This column was originally written in Dutch. This is an English translation.
By Gerd-Jan van Wiggen, Partner at Probability & Partners
As a consultant and investor, I have studied thousands of external disclosures from banks over the past 25 years. During that period, we have seen a major financial crisis, as well as significant changes in legislation and regulation. As a result, I have developed an instinct for the way in which matters are phrased.
The extent to which and the manner in which matters are described sometimes provide a better indication of how well or how poorly things are actually going underneath the surface than the impressive figures shown in the table above. With the advent of large language models (LLMs), the question therefore quickly arose in my mind as to how such tools might be helpful in analysing these disclosures. Saving time when sifting through lengthy texts is a good use case that encouraged me to experiment with this.
My initial attempts a few years ago were disappointing, as the LLMs missed important details. But the quality has gradually improved, and these days I am sometimes surprised by useful additional insights that prompt me to dig a little deeper into certain topics.
A BIS Bulletin was recently published (No. 134, ‘Supervisory screening with large language models: finding divergences’, Fernando Perez-Cruz, Kumar Rishabh and Ayush Uchil, 26 August 2026), in which researchers present the results of a study into how LLMs can assist in assessing prospectuses for the issuance of Additional Tier 1 (AT1) capital by banks.
AT1 instruments are perpetual instruments that function as subordinated debt in normal times but are required to absorb losses under stress. They must meet certain requirements in order to be counted towards banks’ regulatory capital base. These requirements were developed following the financial crisis, as a number of banks had issued AT1 instruments at the time which, in practice, proved unable to function as a capital buffer in a stress scenario. To prevent such a situation in the future, requirements have been formulated which are currently set out in the EU in CRR3/BRRD2, in the UK in the PRA Rulebook and in Switzerland in the Capital Adequacy Ordinance.
Regulators must determine whether proposed AT1 instruments actually meet these requirements. This is detailed work, in which an LLM could potentially add value. The researchers investigated this by carrying out two tests. In the first test, they examined a number of prospectuses from 10 major European systemically important banks to determine the extent to which the prospectuses had incorporated the formal requirements and the extent to which the characteristics of the instruments imposed restrictions on loss absorption or the allocation of losses. Any candidate non-compliance issues identified must then be assessed by an expert to determine whether or not the regulator should take action. This part of the study shows that most of the candidate non-compliance issues identified relate to mechanical triggers and the point of non-viability. It also provides an overview of the types of candidate non-compliance issues identified per jurisdiction.
The second test focused on two historical cases involving known deviations relating to AT1: Credit Suisse and Yes Bank. Deviations were identified in both cases. In the case of Credit Suisse, this concerned a narrower definition of the public support condition at the point of non-viability. In the case of Yes Bank, there had been three AT1 issuances in 2013, 2016 and 2017. In India, up to and including 2015, there was a requirement that shareholders had to bear the first loss. This requirement was subsequently abolished. The instruments issued complied with the rules in all cases at the time of issue.
The 2013 instrument therefore offered greater protection for investors than the instruments issued in 2016 and 2017. The model flagged the 2013 wording as a contradiction with the framework in force in 2020, whilst it did not flag anything regarding the 2016 and 2017 bonds. Until 2019, these instruments were traded at a similar yield, despite the difference in contractual protection. It was only when stress arose that the markets began trading these instruments at different yields. When the RBI intervened, the 2016 and 2017 instruments were written off, whilst the 2013 instrument was not.
This study shows that LLMs can provide signals that require further investigation. For the time being, it is primarily a tool for managing scarce supervisory capacity more efficiently. However, the Yes Bank case also illustrates how this affects investors: the market only priced in the clause once its application became likely.