Three structural pressures are converging on North American banks simultaneously, and the combination is what makes this moment different from prior years of incremental regulatory and competitive pressure. A compliance cost crisis that has reached $61 billion annually with demonstrably poor outcomes. An enforcement environment that has escalated from periodic to structural, with the largest AML fine in US Treasury history paid in 2024. And a fair lending exposure that most institutions have not yet quantified but that regulators have already moved to enforce. Each of these forces, acting alone, would justify a fundamental redesign of how banks make decisions. Acting together, they compress the timeline in ways that are now measurable in months, not years.
The compliance machine is producing volume, not intelligence
The most revealing number in the AML compliance system is not the $61 billion annual cost, striking as that figure is. It is what that investment actually produces at the point where it is supposed to matter. Between 2014 and 2023, US financial institutions filed more than 167 million Currency Transaction Reports with FinCEN. The GAO found that across the portal and agencies’ internal systems, law enforcement accessed less than 3% of those reports in 2023, the most recent full year for which data is available. The system is generating massive volume with limited actionable intelligence, and US financial institutions are paying $4.41 for every $1 lost to fraud when investigation, labour, recovery, and legal fees are included — a figure from the 2023 LexisNexis True Cost of Fraud Study that the 2025 edition has since revised upward to more than $5. That ratio reflects a system optimised for filing, not for outcomes.
This does not mean the compliance apparatus is without value — it means the current architecture produces poor signal-to-noise at an extraordinary cost. And the cost is not stable. Global AML spend is forecast to reach $51.7 billion by 2028 for the US and Canada alone, which means the trajectory of the current model is more spending for structurally similar returns, absent a redesign. The question boards need to force into the room is not whether the current system can be maintained. It is whether continuing to maintain it, while better approaches exist, is a defensible use of capital.
Enforcement has shifted from periodic to structural
The compliance cost story would be containable if enforcement remained predictable. It has not. In October 2024, TD Bank faced total penalties exceeding $3 billion across the Department of Justice, FinCEN, the OCC, and the Federal Reserve for AML program failures including inadequate transaction monitoring. FinCEN’s portion alone — $1.3 billion — was the largest civil penalty ever assessed by FinCEN against a depository institution in US Treasury history. The bank was simultaneously required to submit to a four-year independent monitorship. These are not isolated facts. They are the most visible data points in a pattern of enforcement escalation that is structural rather than cyclical, and that is accelerating across multiple vectors at once.
Fraud has been formally designated an AML/CFT National Priority by FinCEN. According to FinCEN SAR data, check fraud filings in 2024 exceeded 521,000 — nearly double the prior year. Banks with inadequate fraud AI now face direct enforcement risk within their AML programme, not as a separate line of regulatory concern but as a failure of their core compliance infrastructure. At the same time, the FTC’s 2024 Consumer Sentinel Network Data Book, published in March 2025, reported $12.5 billion in consumer fraud losses for the year — a 25% increase over 2023 — and financial institutions bear substantial liability for losses their systems demonstrably failed to prevent.
The structural incompatibility introduced by FedNow makes the picture more acute. The Federal Reserve’s real-time payment network, launched in July 2023, operates on irrevocable transfers. Banks that accept real-time payments while scoring fraud decisions in batch are making prevention decisions retrospectively — after the money has already moved. The liability exposure this creates is not incremental. It is architectural. A bank whose fraud scoring system was designed for a world of batch processing is operating with a structural gap in its risk infrastructure, and that gap widens with every quarter that real-time payment volumes grow.
Fair lending exposure has not been quantified — but the CFPB has
The third pressure is the least visible today and potentially the most consequential in the medium term. The CFPB has explicitly stated that algorithmic credit models are subject to the Equal Credit Opportunity Act. Models that cannot produce a specific, auditable explanation for an adverse action are in violation of ECOA adverse action notice requirements — not in theory, but as a matter of agency guidance that is already in force. The practical question institutions should be putting to their technology teams is direct: when the CFPB examines your AI credit model, can you produce a clear and specific explanation for every declined application, in a format that satisfies ECOA requirements, without human intervention? For most banks, the honest answer is no.
The cost implications of that gap depend on when it is addressed. Banks that build ECOA-compliant explainability into their models from the start bear a certain cost. Banks that deploy models without it and then remediate under regulatory pressure typically bear a cost that is three to five times higher, because retrofitting explainability into a production model touches data pipelines, monitoring infrastructure, and regulatory documentation across the entire deployment. That arithmetic is not speculative — it is the observed pattern across institutions that have gone through remediation programmes. The point at which the CFPB begins examining AI credit models in earnest is not knowable in advance. What is knowable is that the cost of being caught unprepared is substantially higher than the cost of building it in now.
The addressable value is $16.8 to $29.3 billion across North American banking
Against that backdrop of cost and risk, the opportunity for institutions that act is material. The five categories where AI-led decision improvement generates the most direct value — AML false positive reduction, real-time fraud prevention, credit origination conversion, check fraud detection, and ECOA compliance remediation — together represent an estimated $16.8 to $29.3 billion in annual value across North American banking.
| Decision type | Estimated annual value | Basis |
|---|---|---|
| AML false positive cost reduction | $5.5–9.2B | 90–95% false positive rates at large institutions; $3B annual investigation cost (Unit21). A 50% precision improvement represents multi-billion savings industry-wide. |
| Real-time fraud prevention (FedNow) | $4.1–7B | Batch-scoring architectures cannot prevent losses on FedNow irrevocable transfers. The gap is structural, not incremental. (FTC Consumer Sentinel 2024) |
| Credit origination conversion | $3.2–5.8B | Digital-first lenders are capturing origination from incumbents on decision speed alone. (Industry estimates) |
| Check fraud detection | $2.8–4.5B | SAR filings exceeded 521,000 in 2024, nearly double the prior year. Check fraud projected at $24B globally. (FinCEN SAR data) |
| ECOA / CFPB compliance remediation | $1.2–2.8B | Remediation cost for adding explainability post-deployment is typically 3–5x the greenfield cost. (Industry estimates) |
| Total | $16.8–29.3B | Ranges are order-of-magnitude estimates informed by publicly available regulatory and market data. |
The ranges reflect genuine uncertainty in market sizing and should not be read as guaranteed outcomes. What they do not capture is the compounding dynamic, which is in some respects more important than the point estimates. Banks that deploy better decision systems accumulate labelled outcome data that improves model quality with every decision the model sees. The first-mover advantage in AI decisioning is not simply a matter of being first. It is a matter of building a continuously improving system while competitors are still debating whether to act. The gap between early movers and followers does not hold constant. It grows.
Part 1 of 3.
Sources
LexisNexis Risk Solutions / Forrester Consulting. True Cost of Financial Crime Compliance Study — US and Canada. February 2024. LexisNexis Risk Solutions / Forrester Consulting. True Cost of Fraud Study: Financial Services and Lending — US and Canada. 2023 and 2025 editions. US Government Accountability Office. Currency Transaction Reports: Improvements Could Reduce Filer Burden While Still Providing Useful Information to Law Enforcement. GAO-25-106500, December 2024. Financial Crimes Enforcement Network. FinCEN Assesses Record $1.3 Billion Penalty against TD Bank. Press release, October 2024. US Department of Justice. TD Bank Pleads Guilty to Bank Secrecy Act and Money Laundering Conspiracy Violations in $1.8B Resolution. October 2024. Federal Trade Commission. Consumer Sentinel Network Data Book 2024. March 2025. FinCEN. SAR Filing Trend Data, Fiscal Year 2024. fincen.gov/reports/sar-stats. Federal Reserve. FedNow Service. federalreserve.gov/paymentsystems/fednow.
