On 1 October 2026 the Basel Committee on Banking Supervision (BCBS) reported on its meeting in Indonesia on 28 and 29 September. One sentence under the heading “Digitalisation” promises risk managers work. The Committee, it says, agreed “to review the sufficiency and adequacy of existing ‘event type’ loss categories set out in the operational risk framework, with a focus on cyber risk and AI developments”. The release gives no timetable, no format and no direction. Under review is a sorting grid: the seven event types and their sub-categories that banks around the world use to classify their operational losses. In its current form the grid goes back to a working paper published in September 2001.
The release does not establish a capital charge for AI risk. No capital formula, whether in the Basel standard or in European Union law, is calculated by event type. Any change to the categories would affect what gets counted, not how capital is calculated. What it would reach is a taxonomy that has applied in the EU only since 23 September 2026, five days before the meeting began. Delegated Regulation (EU) 2026/1167 writes the seven Basel types into law as the top level and mentions AI in exactly one place, as a possible error in a model. Under the Annex to the Regulation, a loss of that kind must always carry the attribute “misconduct”, which under the Capital Requirements Regulation (CRR) covers wilful or negligent misconduct, while fraud committed with AI tools or the outage of an AI service carries no AI marker at all. What follows is derived from the wording of the Basel standard and EU law. It is not a forecast of where the review will land, and the Committee has published nothing on that so far.
What: the Basel Committee intends to review the event types for operational losses, with a focus on cyber risk and AI developments (BCBS press release, 1 October 2026)
Who: banks that keep a loss data set, in the EU, institutions with a business indicator of EUR 750m or more
Where it bites: in the loss data set, in supervisory reporting and in internal steering, not in the Pillar 1 capital formula
Legal basis: Basel Framework OPE25.17, CRR Articles 312, 316 and 317, Delegated Regulation (EU) 2026/1167 Articles 23 to 33 and Annex
Catch: the EU has just fixed the seven types in law, including back-mapping of losses to 1 January 2016
The seven event types began as a data collection form in 2001
Today the table sits in the Basel Framework under OPE25.17 as “Table 2”: seven types at the top level, from internal and external fraud through clients, products and business practices to system failures and errors in execution and process management, with 20 sub-categories beneath them. The same wording already appears in Annex 2 of the working paper on the regulatory treatment of operational risk that the Committee published in September 2001, then in Basel II as Annex 7 in 2004 and as Annex 9 in 2006. A word-level comparison of the 2004, 2006 and 2026 versions shows differences in layout only. The grid was built for collecting data, and by the Committee’s own account it was already in use before 2002 and applied “again” for the 2002 exercise. According to the Committee’s March 2003 summary report, 89 banks from 19 countries submitted 47,269 individual losses above EUR 10,000 from 2001 for the 2002 loss data collection exercise, each assigned to one of the seven types and one of the 20 sub-categories.
Cyber is no stranger to the grid. Since 2001 the table has listed, under external fraud, the sub-category “Systems security” with the examples “Hacking damage” and “Theft of information (with monetary loss)”. In the 2002 exercise this sub-category accounted for 0.14% of reported events and 0.28% of loss value, according to the same report, and system failures as a whole for 1.1% of events. Cyber risk is therefore not a new subject for the Committee. What is new, as far as its publications show, is that it has put the list itself up for review. A review of its publication list and of the full texts on cyber resilience, operational resilience and digitalisation between 2018 and 2026 finds no earlier statement on whether the event types are adequate.
The types describe the damage, not the cause or the technology behind it. A ransomware attack that shuts down systems, leaks customer data and triggers compensation touches three or four types at once on a plain reading of the Basel text, from external fraud through system failure to duties owed to clients. The standard asks only that the bank documents its criteria: “The bank must document criteria for allocating losses to the specified event types.” It says nothing about which mark the incident should leave in the statistics.
No capital formula is calculated by event type
In the Basel standardised approach for operational risk, finalised in December 2017, internal losses feed through the internal loss multiplier (ILM). Its basis, the loss component, is defined in OPE25.8 as “15 times average annual operational risk losses incurred over the previous 10 years”, an average across all losses with no distinction by type. The event types play a different part under OPE25.17. A supervisor “may request a bank to map its historical internal loss data” to the seven types, “for risk management purposes, and to assist in supervisory validation and/or review”. The industry association ORX (Operational Riskdata eXchange Association) spelt out the consequence in its May 2016 response to the draft standardised approach, then called the Standardised Measurement Approach (SMA): “Event type classification has no role in the SMA which undermines the requirement to invest in data collection by event types.”
In the EU the multiplier does not even come into it. Under Article 312 of the CRR as amended by Regulation (EU) 2024/1623 (CRR3), the own funds requirement for operational risk is “the business indicator component”, and the text contains no loss component at all. As our 9 September analysis of the European Banking Authority (EBA) draft on operational risk management showed, the loss data set does not feed the capital requirement in the Union. The United Kingdom sets the multiplier to one under Policy Statement PS17/23 of the Prudential Regulation Authority (PRA), with the Basel 3.1 rules applying there from 1 January 2027. The March 2026 US re-proposal drops the multiplier according to its footnote 216, but has not yet been adopted.
There are, admittedly, jurisdictions where internal losses do enter capital, Canada, Switzerland and Japan among them according to an overview by ORX. Even there, though, it is the sum of losses that counts, not how they are spread across the types. A revision of the categories would only move the Pillar 1 requirement if it shifted the boundary of what counts as an operational loss in the first place, and the 1 October release says nothing of the kind. The work a new category creates arrives everywhere regardless: in the loss data set, in reporting and in the models banks use to size economic capital under the Internal Capital Adequacy Assessment Process (ICAAP). Whether supervisors weight the categories under Pillar 2 is not established. The capital stays where it is, while the database moves house.
The EU fixed the seven types in law on 23 September
For European banks the real lever lies in an act the Commission adopted on 28 May 2026. Delegated Regulation (EU) 2026/1167 was published in the Official Journal on 3 September and, under its Article 40, entered into force twenty days later, on 23 September. It is the same Regulation whose business indicator rules the EBA had publicly criticised in April, as we set out on 19 June. Its second part, the risk taxonomy under Article 317(9) CRR, stayed in the background at the time. Article 317(9) requires a taxonomy that “complies with international standards”, and the Regulation follows it to the letter: recital 24 says it should be “based on the same event types” as the Basel standard. Article 24 carries over the seven types unchanged as the top level, known as Level 1. Beneath them sit 26 sub-categories at Level 2 and 15 further characteristics, which the Regulation calls attributes.
The EBA chose this continuity deliberately. In its 4 August 2025 final report it wrote that the Committee had delivered a new capital methodology, but “it has not updated its risk taxonomy previously used in the Basel 2 framework”. When some respondents to the consultation asked for the top level to be changed, the EBA declined, citing the ability to map back to Basel. Two consequences of that choice are written into the Regulation. Article 33 requires losses from 1 January 2016 onwards to be assigned to the seven types retroactively, while sub-categories and attributes may be assigned voluntarily from 2025. And recital 29 justifies the back-mapping on the grounds that the Level 1 categories are “unchanged compared to the existing framework”. Re-keying ten years of history is reasonable only because nothing changes at the top.
Five days after the Regulation entered into force, the Committee put its event type categories on its review list. Whether it means only the seven types or the sub-categories as well, the release does not say. The two dates are unconnected, since the European work has been running since the June 2024 EBA consultation. On cyber, Europe found an answer of its own along the way. The consultation draft provided for a separate sub-category for cyber attacks, and in its final report the EBA dropped it “because cyber-related losses could cut across different Level 1 event types and Level 2 categories”. Two attributes took its place: “ICT risk – related to Cyber”, which refers to the definition of a cyber attack in the Digital Operational Resilience Act (DORA), and “ICT risk – not related to Cyber”. Alongside them, sub-category 2.3 “Data theft and manipulation” remains under external fraud, explicitly “by any means, including cyber-attacks”.
In doing so Europe puts into practice what the Basel Committee itself has provided for three times. Its March 2021 revised principles for the sound management of operational risk state that a taxonomy can mark risks arising from legal, conduct, model and information and communication technology (ICT) exposures “(including cyber)”, using the word “flag”. In its December 2022 FAQs on climate-related financial risks, the Committee likewise suggests marking climate losses within the existing categories, “for example, by using a flag”. Basel II already provided in 2006 for operational losses related to credit risk to be flagged separately in the loss database. The European taxonomy therefore follows the Basel line, only in more binding form than Basel has ever made it.
An AI error in a model counts as misconduct in the EU data set
The Regulation provides no attribute for AI. The term appears exactly once in the whole text, in the description of sub-category 4.8 “Model methodology”, which sits under clients, products and business practices: “Where Artificial Intelligence (AI) systems are components of the model, an error due to that technology could fall under the scope of model risk.” Which attributes belong to which categories is set out in the Annex, and the Annex is mandatory. Losses in 4.8 “shall always be assigned the attribute ‘Legal Risk – Misconduct’”, together with the model risk attribute. Model risk, in turn, may not be assigned outside 4.8, 7.4 data management and 7.5 model implementation and use. A loss caused by a faulty AI model therefore appears in the European data set as misconduct, in the same top-level category as mis-selling, market manipulation and breaches of sanctions rules.
The other AI cases leave no AI trace in the data set. Fraud against the bank using a cloned voice or a faked video falls, on a plain reading, under external fraud, for instance under 2.2 for fraud by people who are not customers of the bank. The outage of an AI service bought from a vendor belongs to 6.1, infrastructure and system failure. The Annex prescribes no attribute for either. Depending on the case, attributes such as ICT or third-party risk may apply, but there is no AI attribute under which the cases could be gathered. Whether the cyber attribute applies depends on the DORA definition of a cyber attack, and a deepfake in a video call need not attack a single system. The engineering firm Arup, which lost around USD 25 million in this way in January 2024, as it told CNN, stressed that none of its internal systems had been compromised.
A study by the Federal Reserve Bank of Richmond suggests that losses rising with AI use show exactly this spread. For large US bank holding companies, Atanas Mihov and Ping McLemore estimate that “if a bank increased its AI investments by 10%, its quarterly operational losses would go up by 4%”, driven mainly by external fraud, problems with clients and system failures. Those are three different event types. Annual reports are of little help on causes either. Commerzbank reports operational losses of EUR 576m for 2025, of which EUR 522m sit in the clients, products and business practices category. Its table has no line for cyber. Deutsche Bank stopped showing the granular breakdown by event type in its 2025 annual report, because the standardised approach “does not distinguish operational risk loss event type categories”. Anyone who wants to know which losses AI causes will not find the answer in any of these grids, including the new European one.
Conduct risk got its attribute fourteen years after the LIBOR scandal
The last wave of losses that strained the grid came after 2008. Redress for payment protection insurance (PPI) in the United Kingdom and for interest rate hedging products ended up in clients, products and business practices. In its 2013 Pillar 3 report Barclays showed 90.0% of its net losses in that category, “heavily impacted by provisions for PPI and interest rate hedging product redress”. Penalties for manipulated benchmark rates such as the London Interbank Offered Rate (LIBOR) and for collusion in foreign exchange markets were probably booked there too, not least because Basel lists “Market manipulation” as an example of that category, although no Pillar 3 report confirms it. Anyone who lived through the fixing years on a trading floor knows how long a loss pattern takes to reach the rulebook. No new Basel category came out of it, and the word “misconduct” still does not appear in the Basel standard.
The answer came by the European route, and late. In the 2024 CRR3 amendments, Article 317(1) names “misconduct events” explicitly as part of the loss data set, and only the 2026 Regulation introduces the attribute “Legal Risk – Misconduct” for them, fourteen years after the 2012 LIBOR scandal. The parallel with cyber and AI holds, though not entirely. Conduct losses clustered in one category and could be found there even without an attribute. Cyber and AI spread across at least three types, and that spread is precisely what makes them invisible in statistics organised by type. This argues for an attribute rather than a new top-level category, because a top-level category requires a single answer. Under Article 23(1) of the Regulation each event goes into exactly one Level 1 type, and where several apply, into “the most relevant”. A ransomware case with business interruption and customer compensation would then have to choose between cyber and system failure, and a year later the choice would be hard to reconstruct.
The time series argues for a new top-level category
Level 1 assigns each event to exactly one type and is mandatory back to 2016. Attributes are also mandatory for current losses, but they overlap, and for historical data banks may assign them only voluntarily, from 2025, under Article 33. A cyber time series built from attributes therefore has reliable data only from 2025, while one built on Level 1 would have ten years of history. Only the top level is internationally comparable, moreover, because the attributes are European law and other jurisdictions build their own solutions. In its 2019 reference taxonomy ORX counted, across 60 taxonomies of banks and insurers, an average of 14 risks at the top level against Basel’s seven, with information security, cyber, model and third-party risk among those that had moved up. Practice extended the list long ago, and the Committee’s review may simply catch up with what banks have been running internally for years.
That argument has another side, however. If Basel changes the top level, the EU faces a second conversion within a few years, with an amended Regulation and a fresh back-mapping that recital 29 has so far justified only for unchanged types. Some respondents to the consultation had themselves advised the EBA to treat cyber as an attribute: “A cyber risk attribute instead of the L2 category might be easier to implement and more helpful.” Three outcomes are conceivable. The Committee can confirm the grid as adequate. It can stay with its approach in 2006, 2021 and 2022 and add markers. The European solution would be confirmed, and the AI gap would stay open until someone writes an attribute for it. Or it introduces new types, followed by an amendment to the Delegated Regulation and a re-keying of the history. None of these outcomes costs Pillar 1 capital, and the last two cost data work.
Recommendations for operational risk, non-financial risk and reporting teams
Now: Article 317(4)(b) CRR requires “descriptive information about the drivers or causes of the loss events”, with a level of detail that depends on the size of the loss. The obligation applies to institutions with a business indicator of EUR 750m or more, and smaller banks benefit from the same recording for steering. An internal field that records for each event whether and how AI was involved, as an in-house model, as a purchased model or as an attacker’s tool, fits exactly there. The Regulation provides for no sixteenth attribute in the reporting data set, and it does not prohibit an internal field.
Now: Whether a model loss falls under model methodology (4.8) or model implementation and use (7.5) decides whether it appears as misconduct in the data set. That boundary deserves a documented rule with examples from the bank’s own practice, before the first AI losses are assigned and a practice forms on its own.
Ongoing: In its final report the EBA expects losses carrying the attribute “ICT risk – related to Cyber” to be consistent with the major ICT-related incidents reported as “cybersecurity-related” under Implementing Regulation (EU) 2025/302. That is the authority’s expectation, not an obligation in the text of the Regulation, but a gap between the two lists is what supervisors will notice first.
Preparatory: Store the description, cause, affected systems and parties involved for each event so that the history can be mapped to a changed grid. If Basel changes the top level, a bank holding this data re-keys its records instead of researching ten years of losses from scratch.
Watch: The Committee has named neither a timetable nor a format, and by its own account its decisions have “no legal force”, relying instead on its members’ commitments. Pillar 1 capital planning has nothing to set aside for a new category. Within the ICAAP, however, it is worth checking whether AI and cyber losses are visible separately in the economic capital model.
Glossary
Event type (Level 1): Top level of the classification of operational losses. Basel and the EU use seven types, and every loss is assigned to exactly one of them.
Level 2 category: Sub-category of an event type. The EU Regulation has 26, the Basel standard 20.
Attribute: Additional characteristic of a loss, such as cyber, model risk or misconduct. A loss can carry several attributes, and the Annex to the Regulation determines which belong to which categories.
Loss data set: Structured collection of realised operational losses under Article 317 CRR, mandatory for institutions with a business indicator of EUR 750m or more.
Business indicator component: Own funds requirement for operational risk under CRR3, calculated from the bank’s income figures, without a loss component.
Internal loss multiplier (ILM): Factor in the Basel standard that scales the requirement by the losses of the past ten years. Not adopted in the EU, set to one in the United Kingdom.
Christian Schablitzki – the agentic banker
Strategy & Management Consultant · Agentic AI expert for financial institutions
More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Currently a Partner at Infosys Consulting in Germany. Certified as Claude Partner – Claude Code (Anthropic), as well as in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.
LinkedIn profile → · agenticbanker.ai →Keep reading – every 14 days in your inbox.
Capital-markets insights, regulatory updates and AI trends. Concise, well-founded, free.
GDPR-compliant. Unsubscribe at any time.