Reg42

A letter from Avner Yoffe · November 2026

Don’t Panic.Regulation is Cool.

On why compliance is ready to be rebuilt, the piece AI was missing, and a game both sides of the counter can win

To whom it may concern,

Ask a bank whether it is compliant, and time how long the answer takes.

It is a fair question for a firm that holds other people’s money, and the honest reply is still a project in three parts: know what you owe, decide how you comply, prove what you do. Most of it is reading, mapping, checking and explaining, done by people. By our own estimate, built on public data, compliance costs financial firms worldwide nearly $1 trillion a year.

I have spent 10 years on the technology meant to make this easy: 3 at a regulatory reporting provider, then 7 at a global trading platform, where I led the creation of its regtech department from scratch. It was a privilege. Hypergrowth teaches you to scale without losing your grip on the rules, and I learned it next to some of the sharpest people in regtech. In all that time, I never saw a tool that could answer the question on the day it was asked.

Think of the letter every compliance team knows. A regulator asks, by a date, which policies cover a product, who approved them, which controls ran last quarter and what they found. The answers exist, scattered across a dozen systems and a few people’s inboxes, and a good team loses a week putting them together. Or think of the question no regulator ever dates: what is good enough? The rules ask for a risk-based, proportionate approach, then leave the firm to decide how many checks, which thresholds and how much evidence. Too little, and an examiner finds it. Too much, and honest customers are kept waiting and pay for the caution.

Too little is the dangerous side, and the harder one to see: has every obligation that applies been found, owned and met? Only 45% of practitioners in CUBE’s latest survey rated their approach to regulatory change effective, down from 59% a year earlier. When the answer is no, the bill is public. Penalties on financial institutions for anti-money-laundering, sanctions and due-diligence failures came to $3.8 billion in 2025, by Fenergo’s count, and enforcement shifted rather than stopped: down 61% in the US, up 767% across Europe, the Middle East and Africa. In March, the US Treasury’s FinCEN fined a broker-dealer a record $80 million because its anti-money-laundering program “was not proportional to the risks of its business model.” At the far end, TD Bank left 92% of its transaction volume unmonitored for more than 6 years; it pleaded guilty, agreed to $1.9 billion in Justice Department penalties alone, and had its US growth capped.

So the industry does what it knows, and hires. The United States counted 417,070 compliance officers in May 2025, across all industries, more in every annual estimate since 2010. At the 20 banks in a Bank Policy Institute survey, compliance hours rose 61% from 2016 to 2023, while all hours rose 20%.

The cost is not the worst of it. A financial firm grows by adding: a market, a license, a product, a partner. Each one starts with the same question: what do we owe now? And each answer is a project staffed by hand. Growth can compound; a program that scales by hiring cannot. That makes compliance probably the biggest bottleneck between a financial firm and exponential growth.

Everyone in finance is talking about AI now. Far fewer have looked hard at what it does to that question. Some of us have seen it coming: the people who built the surveillance stacks and the reporting pipelines inside banks, brokers and platforms, who have read the same duty in the FCA Handbook, the FINRA rulebook and MiFID and know it is one duty wearing three outfits. What we saw coming was not another rule. It was a new kind of reader.

Seeing it takes no special insight, only a willingness to stop arguing with two charts. The first is ours: for 15 years, the answer to more compliance has been more people. The second belongs to the machines. In June 2024, the best AI agent on METR’s software tasks could finish, half the time, work that takes a human expert about 11 minutes. By April 2026, METR measured more than 16 hours, beyond what its tasks can reliably gauge. GDPval, OpenAI’s test of real professional work, includes a suspicious activity investigation at a fictional bank. Across its 44 occupations, the best model matched or beat the professionals 47.6% of the time in October 2025; by April 2026, OpenAI was reporting 84.9% for its own GPT-5.5. Read side by side, the two charts say something uncomfortable: the work that keeps adding compliance officers is the work AI is learning to do.

Not every result is that clean, and the exception is the point. GDPval hands each model the full context of a task, often with the files attached. SemiAnalysis’s new GDPx puts models to work as agents in realistic environments instead, and the best scored 12%. On its finance tasks, graded pass or fail, every model scored zero; the legal ones were too hard for all of them. One test hands the model the context. The other makes it work for it.

Inside financial firms, the gap shows up as caution. In the Bank of England and FCA survey published in October, 80% of responding firms had AI agents deployed, and 2% let them act with broad autonomy. Most firms have hired the graduate. Hardly any let it sign.

A frontier model is a brilliant hire on its first morning: it can read anything and knows nothing about the firm. Fitting it into real work is proving slower than building it, and in compliance the fitting is most of the job.

The intelligence has arrived. Nobody has onboarded it.

For once the word belongs to the machine rather than the customer. In compliance, onboarding means every obligation that applies, every policy and control written to meet it, who owns each one, what proves it ran, and the date on all of it. Too much of that sits in a spreadsheet tab called FINAL_v7, in committee minutes nobody can find and in the memory of whoever wrote the policy before leaving in 2022. Ask the new hire to explain the firm to an examiner, and you will see the problem.

Supervisors see the same charts. The Bank of England’s Deputy Governor, Sarah Breeden, said in June that “an already exponential increase in capability appears to be accelerating,” then added:

Our frameworks were not built to contemplate autonomous agents, and relying on a human in the loop for all agent actions is unlikely to be realistic.

Sarah Breeden, Deputy Governor, Bank of England, 30 June 2026

This month, America’s national bank supervisor stopped calling it artificial intelligence. The OCC’s new supervisory framework calls it super intelligence, and lets a bank classify the risk of each use by its potential consequences, how much agency it hands over and “the degree of human oversight rather than the technical sophistication” of the system. If people cannot check every step, the record has to, down to which obligation, which version, which control, whose signature and what proof.

That record is what I mean by situational awareness: the firm knowing its own situation, on any day, with dates attached. Every institution that holds other people’s money should have it.

Building it is the problem my co-founder Eitan Ezra and I chose, and it is unglamorous on purpose: we do the onboarding. The missing piece is an ontology, one shared way to write down what a firm owes, what it does and what proves it. Ours is CLHEAR, patent-pending: it writes a regulator’s clause and the policy that answers it in the same form, readable by people and AI alike, every line traced to its clause. On top sits a context graph that links all of it, owners and evidence included, and keeps each change with its date and context. Foundation Capital named the idea in December 2025; we apply it where a dated record is the job itself.

Solon, our compliance AI, reasons over that graph. A general AI model answers from its training data and the web; it can read a firm’s files too, but with no compliance standard or graph it can trust. Solon answers from the firm’s graph, with full provenance: the rules, obligations, policies, controls and evidence behind each answer, and the reasoning that links them. It finds what applies, maps what the firm has and flags the gaps. It says when it does not know, and shows each person only what they may see. People decide and sign. It runs in a dedicated, private AWS account that Reg42 manages, or in the client’s own cloud, and firms start with one domain.

With that piece in place, I believe today’s models can already run most of a compliance program automatically, and that within 2 to 3 years this will be the mainstream way financial firms run theirs. Judge it on your own tasks, against the same model without the graph. Not all of it is built yet, and I would rather you read that here than discover it in a demo: Part V keeps what runs today apart from what comes next.

The prize is not only ours. Rules are how a society writes down what it expects of the people it trusts with its money, and good ones protect customers without stopping anyone from building. Our vision is a financial game that both sides of the counter win: each firm’s rules set at the sweet spot, as much control as the risk calls for and no more, and run with cutting-edge AI. The firm spends less proving what it does and grows without hiring a department for every market. Its customers get decisions made about them rather than about their category.

Regulators are pointing the same way. FATF now asks countries to “allow and encourage” simplified measures where risk is lower, and warns that overly cautious safeguards can shut legitimate customers out or raise what they pay. The OCC says a bank’s control intensity could rise with the potential consequences of an agent’s failure or misuse. In the UK, a bank closing an account opened since 28 April 2026 owes the customer longer notice and a reason they can understand, financial-crime cases aside; with the evidence attached, it can stand behind that reason. The regulator’s letter becomes a query rather than a project, and the fraud team can spend its hours on the scam instead of the file about the scam.

Protection and growth can travel together. BCG’s new benchmark of 26 systemically important banks found those with more mature risk management 2.5 times more likely to be advanced AI adopters: governance and speed, in its words, “need not be a tradeoff.” They have traveled together before. After the UK made payment firms reimburse victims of push-payment scams, an independent evaluation published by the Payment Systems Regulator estimated the policy cut losses by £73 million a year, with no sign of firms leaving the market, though some scams may have moved to international payments.

Read this part twice: now, and again when the dates come due, with a red pen. Here are my dates, and how you will know I was wrong.

By the end of 2027, the first regulated firms will hold a complete map of what they owe and what they do, drawn by machines from the rulebooks and their own records, evidence attached, signed by the people responsible. I am wrong if, by the end of 2027, no regulated firm says publicly that it holds such a map for the whole firm, signed by named people. If by then no supervisor has published guidance on relying on one, or what appears demands a line-by-line human review, Part VI counts that against me too.

In 2028, those firms will open a new country by changing the description of themselves that the system reads (their products, licenses and customers), not by hiring a department, and compliance will stop being the reason a launch waits. I am wrong if, by the end of 2028, no regulated firm reports entering a new jurisdiction that way.

Meanwhile teams will be rebuilt around the system rather than the spreadsheet. People will keep the judgment and the signature and machines will take the legwork, so the number of US compliance officers will stop climbing at the pace it has kept since 2010, even as the work grows. The signpost is the Bureau of Labor Statistics count for May 2028, published in spring 2029. It counts every industry, not just finance. Above 460,000, the old trend held and I was wrong. Below 430,000, the line has bent; in between, it is too early to call.

I could be wrong about the dates; I have watched enough roadmaps to know which way they slip. Every figure here is public and sourced at the end, our estimate is marked as ours, and none comes from a client. The six articles below make the case in detail, each written to stand on its own, and Part VI argues against me. If you are a skeptic, start there. I hope you are.

  1. IFrom Tasks to the MapWhy years of regtech spending automated tasks while the map of what a firm owes stayed in people’s heads, why cheaper, better machine reading changes that, and a keyword test on FCA, FINRA and MiFID II texts that pairs British and European duties and misses every American one.
  2. IIThe WallFewer US final rules and smaller FCA fines, yet more regimes arriving at once, each written in its own words and run as its own program: a cost that rises with every regime a firm adds, and one no hiring rate can match.
  3. IIIOne Common LanguageThe same four duties in the regulators’ own words from Brussels, London and New York, what a word-matcher finds and misses across them, what eight industries learned from building shared data languages, and what a language for compliance has to hold.
  4. IVProof Before AutonomyWhat hallucination studies, legal AI benchmarks and regulators in London, Paris, Washington and Brussels say about who answers for a machine’s work, and four conditions, drawn from how aviation, medicine and driving earned autonomy, that automated compliance should meet before it acts alone.
  5. VWhat We Are BuildingRules on AI in finance are arriving in the EU, the US and the UK at once, as a statute, a contest and a posture, so one activity carries three separately worded duties; how Reg42’s system maps them to a firm’s own records, with what runs today kept apart from what comes next.
  6. VIWhat Could Prove Me WrongSeven ways the predictions could fail, from regtech’s track record and stalled enterprise AI to hallucinating legal tools, late standards and an official forecast of more compliance officers, each paired with a public number and a date that will settle it.

If you run compliance at a bank, broker, insurer or payments firm, start with one domain and hold the result to the standard you hold your own team to. If you have spent a career inside a regulator, tell us where a machine-drawn program would fail your exam; none of this should rest on my word. Builders will find the hard parts still unbuilt. Funders know where to find me, and now so does everyone else: avner@reg42.ai.

If all of this sounds early, consider for a moment that it might be on time. I have tried to be concrete, and the red pen is yours.

A word on the title. Don’t panic, because this is good news: trust is what lets a stranger hand you their savings, and proving you deserve it is about to get cheaper. Regulation is cool, a line I have used for years, because good rules are what make that trust possible, and running them well at scale is the most interesting problem I know. As for the 42: in the book our name comes from, a computer thinks for 7.5 million years, answers 42, and nobody can say what the question was. Compliance has the opposite problem: the question is plain, and the answer is slow. Even the number checks out: at the banks in the Bank Policy Institute’s survey, compliance took 42% of the C-suite’s time in 2023. We would rather keep the 42 in our name and out of their calendars.

Don’t panic. The answer is 42.

Avner YoffeCo-founder and CEO, Reg42 · Tel Aviv, November 2026

Sources

  • Nearly $1 trillion a year: Reg42’s own estimate, built on public data, and an extrapolation, not a measurement. Method: TheCityUK and PwC’s finding that total compliance costs exceed 13% of operating costs in UK financial services (TheCityUK, 5 November 2025) is applied to the operating costs of banks (McKinsey, Global Banking Annual Review 2026, May 2026: about $6.4 trillion of revenues before risk costs in 2025), insurers (Swiss Re Institute, sigma 2/2026, World insurance in 2026, July 2026: $8.48 trillion of premiums in 2025) and asset managers (BCG, Global Asset Management Report 2026, 28 April 2026: $147 trillion of assets under management, margins above 30%), with Reg42’s own assumptions, which no source gives: banks’ costs at 50% to 55% of revenues, insurers’ operating expenses (excluding claims) at 15% to 25% of premiums, and asset managers’ costs at about 70% of an assumed $0.5 trillion to $0.65 trillion of revenue. That gives roughly $600 billion to $800 billion a year for banks, insurers and asset managers. Payments firms, brokers, lenders and other financial firms are not in these sources; the step from there to nearly $1 trillion is itself Reg42’s assumption. The 13% is a UK ratio (the part UK firms directly measure is 2.6%; the rest is spread across the firm), from interviews with 19 chief compliance officers and a survey, extrapolated here to the world; it may overstate or understate other markets.
  • CUBE, From adoption to assurance: the cost of compliance in an AI-enabled world, released 4 November 2026 and read from press coverage of the release (FF News, 4 November 2026; TechIntelPro, 5 November 2026). More than 2,000 professionals at manager level and above in 11 countries, across banking and finance, insurance and professional services, surveyed in June and July 2026: 45% rated their approach generally or highly effective, down from 59% a year earlier. The coverage reports the question in the context of how firms embed regulatory change. CUBE is a regtech provider.
  • Fenergo, Global financial regulatory penalties fall by 18% in 2025 as enforcement shifts from US to EMEA and APAC, 13 January 2026: penalties on financial institutions for failing to comply with anti-money-laundering, KYC, sanctions and customer due diligence regulations totalled $3.8 billion in 2025, against $4.6 billion in 2024 and $6.6 billion in 2023. US penalties fell 61% (the release gives 58% for North America), while EMEA and APAC penalties rose 767% and 44%. The largest single penalty was $985 million (€835 million), issued to a Swiss bank by French authorities. Fenergo is a client lifecycle management provider; the count is its own, and the release does not set out its method.
  • Financial Crimes Enforcement Network, FinCEN Assesses Historic $80 Million Penalty Against Canaccord Genuity LLC, 6 March 2026: “the largest penalty ever imposed against a broker-dealer for violating the BSA”. Canaccord’s “significantly under-resourced AML program was not proportional to the risks of its business model”, and it failed to file at least 160 suspicious activity reports.
  • US Department of Justice, TD Bank Pleads Guilty to Bank Secrecy Act and Money Laundering Conspiracy Violations in $1.8B Resolution, 10 October 2024: “92% of total transaction volume going unmonitored from Jan. 1, 2018, to April 12, 2024”, about $18.3 trillion of transaction activity; a total Justice Department financial penalty of $1,886,945,780.40; TD Bank “became the largest bank in U.S. history to plead guilty to Bank Secrecy Act program failures”. The Federal Reserve, the OCC and FinCEN took concurrent actions, with penalties of their own.
  • Office of the Comptroller of the Currency, News Release 2024-116, 10 October 2024: a $450 million civil money penalty against TD Bank, N.A. and TD Bank USA, N.A., and “a restriction on the growth of the bank”, which the Acting Comptroller called an asset cap.
  • US Bureau of Labor Statistics, Occupational Employment and Wages, May 2025, 15 May 2026, Table 1: 417,070 compliance officers (occupation 13-1041), in all industries, not only finance; 46,460 of them worked in finance and insurance (BLS OEWS industry series for NAICS 52). The archived annual releases from May 2010 (204,000) onward show a rise in every year; BLS cautions against strict year-to-year comparisons.
  • Bank Policy Institute, Survey Finds Compliance Is Growing Demand on Bank Resources, 29 October 2024. Twenty banks, about half of BPI’s membership, responded: global systemically important banks, regional banks and major foreign banks doing business in the United States. Between 2016 and 2023, employee hours dedicated to complying with financial regulations and examiner mandates rose 61%, while aggregate employee hours rose 20%. In 2023, 42% of C-suite time and 43% of board time went to regulatory or supervisory compliance, against 24% and 27% in 2016. A survey of BPI’s member banks, not of the whole industry.
  • METR, Task-Completion Time Horizons of Frontier AI Models and its raw data file, page last updated 8 September 2026 and no longer actively updated. The 50% time horizon on software tasks: about 11 minutes for Claude 3.5 Sonnet (June 2024) and more than 16 hours for an early version of Claude Mythos Preview (April 2026), above the 16 hours METR says its current task suite can measure reliably. The horizon is the length of task, in human expert time, that an agent completes half the time, not how long the AI runs.
  • OpenAI (Patwardhan et al.), GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks, arXiv, 5 October 2025, with the public gold-subset tasks. Claude Opus 4.1’s deliverables were graded as better than or as good as those of professionals averaging 14 years’ experience 47.6% of the time on the 220-task gold subset. Compliance Officers are among the 44 occupations, and one public task is a suspicious activity investigation at a fictional bank. The 47.6% is across all 44 occupations; GDPval publishes no compliance-specific score. The paper calls its tasks “precisely-specified and one-shot, not interactive”: each comes with “the full context of the task in the prompt”, often with reference files.
  • VentureBeat (Carl Franzen), OpenAI’s GPT-5.5 is here, and it’s no potato: narrowly beats Anthropic’s Claude Mythos Preview on Terminal-Bench 2.0, 23 April 2026, reproducing OpenAI’s launch table: GDPval (wins or ties) 84.9% for GPT-5.5. Reported by OpenAI about its own model and read from press, because OpenAI’s page was not reachable. OpenAI may have used a later version of GDPval, so the 47.6% and the 84.9% are not a like-for-like series.
  • SemiAnalysis (Max Kan, Andrew Megalaa and Dylan Patel), GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains, 2 November 2026. Models run as agents inside realistic work environments: “GDPx is a true frontier-level benchmark: even the best model scores just 12%.” On finance, “all the models score 0% if you use binary grading”, and legal tasks were “too difficult for all models when using the intended binary grading”. The finance tasks are private-equity deal work, not compliance. A paid post, read in its public portion; a new benchmark from an industry analyst firm.
  • Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services, 2026, 26 October 2026. 80% of respondents had AI agents deployed, mostly for decision support (agents recommend, humans decide); 2% reported agents with broad autonomy. UK respondents to the full survey only.
  • Bank of England, Sarah Breeden, Agents of change, ECB Forum on Central Banking, Sintra, 30 June 2026. “So an already exponential increase in capability appears to be accelerating” (her reading of METR’s data, cited in her footnotes), and the block quote on frameworks and the human in the loop.
  • Office of the Comptroller of the Currency, Bulletin 2026-54, Super Intelligence: Supervisory Framework, 4 November 2026, addressed to national banks, federal savings associations and federal branches and agencies. “Super Intelligence” refers to the technologies and systems “previously encompassed by the term ‘Artificial Intelligence’”. A bank’s risk classification of SI use cases may rest on “potential consequences, transfer of agency, and the degree of human oversight rather than the technical sophistication of the SI used”, and its “control intensity could increase with the potential consequences of an SI agent’s failure or misuse”. The OCC adds that “nothing in this guide is intended to establish new regulatory expectations”.
  • Foundation Capital (Jaya Gupta and Ashu Garg), AI’s trillion-dollar opportunity: Context graphs, 22 December 2025. The essay that named and defined context graphs.
  • Caribbean Financial Action Task Force, Revisions to Recommendation 1: Financial Inclusion, Part 1, 2 September 2025. FATF’s February 2025 revision changed “may decide to allow” simplified measures to “allow and encourage” them in lower-risk situations (FATF’s wording as the CFATF reproduces it); FATF’s 2025 guidance, as summarized by the CFATF, says several factors have contributed to overly cautious and non-proportionate safeguards that can exclude legitimate customers or raise their costs.
  • legislation.gov.uk, The Payment Services and Payment Accounts (Contract Termination) (Amendment) Regulations 2025, SI 2025/688, in force 28 April 2026. For contracts from that date, at least 90 days’ notice (up from two months) and reasons “sufficiently detailed and specific” for the customer to understand, with financial-crime exceptions.
  • BCG, Operationalizing AI Governance Gives Banks an Advantage in Scaling AI, 13 November 2026, on the report AI Safety and Governance in Banking 2026, a benchmark of 26 global and domestic systemically important banks: “banks with more mature risk management are 2.5 times more likely to be advanced AI adopters”, and “governance and speed need not be a tradeoff”.
  • Payment Systems Regulator, Payment fraud falls by £73m following PSR reimbursement scheme, 1 July 2026, reporting Frontier Economics’ independent evaluation: APP fraud losses down an estimated £73m a year because of the policy, a positive short-term net benefit after firms’ costs (a conservative assessment) and “no evidence of market exits”. The displacement point comes from Frontier Economics’ own evaluation deck (Final Report webinar, dated 7 July 2026), whose findings the PSR’s page summarizes: “Some of this reduction may be offset by the policy driving an increase in international APP scams”.
  • The signposts follow Part VI. 2027: Part VI’s first claim is a map that covers the whole firm, drawn by software and signed by named people; its end-2027 signpost is supervisory guidance from the FCA, ESMA, the SEC or a national authority that names machine-generated obligation registers, where silence or a line-by-line human review requirement counts against the letter. 2028: a regulated firm reporting a new jurisdiction entered by changing the description of itself, its products and its licenses, not by building a local compliance team. Headcount, identical to Part VI’s Risk seven: the US Bureau of Labor Statistics OEWS national employment estimate for occupation 13-1041 for May 2028, due in spring 2029; above 460,000 the trend held, below 430,000 the line has bent. The 2010 to 2025 pace was about 4.9% a year (204,000 to 417,070). For orientation, Part VI computes about 481,000 for May 2028 on that trend and about 422,000 at the rate implied by the BLS Employment Projections, Table 1.2, 2025 to 2035 (+3.8% over the decade, on a different employment base of 436,400 that includes the self-employed; the two counts are not mixed in one ratio).