Industries / Finance and business services / Finance and banking
Fraud and financial crime
This work keeps criminals from stealing through the financial system or using it to hide money, from card fraud and scams to money laundering, which disguises where criminal money came from.
7
Milestones achieved, from 1970 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1970
Banks must report cash
The US Bank Secrecy Act set rules for recordkeeping and reporting by banks, and it required them to report cash transactions over $10,000 on a Currency Transaction Report, leaving a paper trail. The Money Laundering Control Act of 1986 made money laundering a federal crime, and a 1992 law required Suspicious Activity Reports (SARs), in which banks tell the authorities about activity that may be criminal.
Credit
US Congress; the record is kept by FinCEN, the Financial Crimes Enforcement Network
1989
A task force on laundering
At their Paris summit of 14 to 16 July 1989, the Group of Seven leaders agreed to "convene a financial action task force" to assess cooperation against the use of banks for money laundering and to consider further measures, including changes to laws and regulations. The Financial Action Task Force (FATF) now designates the categories of crime that money-laundering laws should cover, and keeps a "grey list" of countries, which the IMF helps countries leave.
Credit
G7 leaders at the 1989 Paris summit
2006
Chip and PIN replaces signatures
In February 2006 chip and PIN fully replaced signatures for in-store card payments in the UK: the card carries a chip, and the shopper types a personal identification number (PIN) instead of signing. UK Finance, the banking industry body, says losses to counterfeit cards fell 95% over the following 20 years.
LessonFraud moved to payments the chip does not protect; remote purchase fraud, on online and phone orders, took £215 million in the first half of 2025, by UK Finance's figures.
Credit
the UK card industry, by UK Finance's account
2016
Bangladesh Bank theft
In February 2016 attackers sent fraudulent payment orders over the Swift network and moved $81 million from Bangladesh Bank's account at the Federal Reserve Bank of New York to accounts at Rizal Commercial Banking Corporation in Manila. Much of it went into the Philippine casino industry, and about $15 million had been recovered by February 2018, Al Jazeera reported. In May 2018 the payments committee of the Bank for International Settlements (BIS) set out a strategy for central banks, payment system operators and banks, warning that "weaknesses in security at one endpoint in the ecosystem can be exploited to commit payments fraud".
LessonA payment network is only as safe as the least protected computer allowed to send instructions into it.
Credit
the BIS Committee on Payments and Market Infrastructures, for the response
2020
Wirecard's missing cash
Wirecard, a German payments company, filed for bankruptcy protection in 2020 after €1.9 billion of cash on its books "could not be verified". On 3 November 2020 the European Securities and Markets Authority (ESMA) found deficiencies in the work of Germany's two supervisors of company accounts, BaFin and the Financial Reporting Enforcement Panel, including the "non-selection (or non-timely selection)" of Wirecard's reports for examination between 2016 and 2018 and examinations that did not address the "media and whistle-blowing allegations" against the company. In April 2023 Germany's audit oversight body, APAS, fined EY's German arm €500,000 over its 2016 to 2018 audits of Wirecard and barred it for two years from taking on new public-interest clients.
LessonWarnings already raised in public are worth nothing until a supervisor examines them.
Credit
ESMA; APAS
2024
Scam victims repaid by rule
In an authorized push payment (APP) scam, a criminal tricks a person into sending money themselves. From 7 October 2024 UK payment firms have had to reimburse most APP scam victims, under rules set by the Payment Systems Regulator (PSR). In the first nine months, the PSR found, 88% of the money lost and claimed back was returned, £112 million in all, against 66% in the same months a year earlier, and 84% of claims were resolved within five business days. Claims fell about 15% from a year earlier.
Credit
Payment Systems Regulator
2024
TD Bank pleads guilty
On 10 October 2024 TD Bank pleaded guilty to Bank Secrecy Act and money laundering conspiracy charges in the United States. Prosecutors said 92% of its transaction volume, about $18.3 trillion, went unmonitored from 1 January 2018 to 12 April 2024, because it left domestic automated clearing house (ACH) transfers, the US system for bulk bank-to-bank payments, most check activity and other types out of its automated monitoring. Three laundering networks moved more than $670 million through its accounts between 2019 and 2023. Attorney General Merrick Garland called it "the largest bank in U.S. history to plead guilty to Bank Secrecy Act program failures", and penalties across US agencies came to about $3.09 billion, by the Federal Reserve's count.
LessonA monitoring system cannot catch crime in payments it was never set up to see.
Credit
US Department of Justice; Federal Reserve and other US regulators
Watch for
Banks pool fraud models
Swift, which runs a messaging network that banks use to send each other payment orders, says a model trained with 13 banks, including ANZ, BNY and Intesa Sanpaolo, was "twice as effective in identifying instances of known frauds" as a model trained on one bank's data. It used federated learning, in which the model trains inside each bank so customer data is not shared, and the trial ran on ten million artificial transactions. Swift plans a second phase on real transaction data.
Achieved whenThe model runs on real payments and a party outside Swift, such as a participating bank's regulator or an independent reviewer, reports its detection and false-alarm rates.
Claimed by
Swift, September 2025
An EU anti-laundering supervisor
The EU Anti-Money Laundering Authority (AMLA), based in Frankfurt, began operating in summer 2025. It plans to select about 40 cross-border financial firms with a high risk of money laundering during 2027 and to supervise them directly from 2028.
Achieved whenAMLA names the firms it supervises and publishes the results of its first direct supervision.
Claimed by
AMLA, on its website, as of November 2026
Open
01
Test drafted suspicious activity reports
A public benchmark for AI drafts of the narrative in a suspicious activity report, the written account in which a bank investigator explains to the authorities who did what, when, and why it looks criminal. Build at least 200 cases from published synthetic transaction data, each with a list of reference facts. Have models draft the narratives, and have two reviewers with anti-money-laundering experience count invented facts and missing reference facts in each draft. No real customer data is used.
The briefCredit · unclaimed02
Match names across scripts
A test of sanctions name matching across writing systems, with results reported per script. Banks screen customers and payments against public sanctions lists, and an analyst clears each possible match by hand. Take entities in the public lists that already carry names in more than one script, such as Latin with Cyrillic, Arabic or Chinese, build checked pairs of matching and non-matching names, and report how often each matcher finds the true matches and rejects the false ones in each script.
The briefCredit · unclaimed
Open 01 · the brief
Test drafted suspicious activity reports
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark for AI drafts of the narrative in a suspicious activity report, the written account in which a bank investigator explains to the authorities who did what, when, and why it looks criminal. Build at least 200 cases from published synthetic transaction data, each with a list of reference facts. Have models draft the narratives, and have two reviewers with anti-money-laundering experience count invented facts and missing reference facts in each draft. No real customer data is used.
Why it is still open
The Co-Investigator AI paper (Prathamesh Vasudeo Naik and colleagues, September 2025) describes an agent system for drafting these narratives and does not release a dataset or benchmark. The software firm Hummingbird reported that early ChatGPT drafts invented a law-enforcement contact and a branch visit that were not in the data. Hummingbird calls narrative writing the "most time-intensive portion" of an investigation, and says compliance leaders rated it "one of the least impactful tasks to potentially automate".
Built with what exists
The synthetic anti-money-laundering transaction sets published by Erik Altman and colleagues (2023); models that can run inside a bank's own computers, such as Mistral Large 4 (see the Note "A trillion open parameters") with the ggml library (see the Note "A first stable number for local AI"); a hosted model such as Claude Fable 5.2 for comparison (see the Note "Three working devices"); a public scoring sheet.
First step
Pick 20 laundering patterns from the synthetic data, write their fact lists, have one model draft narratives, and count the invented facts.
Done looks like
200 cases with reference facts published, and for at least three models a table of invented facts per narrative and the share of reference facts included, scored by two reviewers whose agreement is reported.
Open 02 · the brief
Match names across scripts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of sanctions name matching across writing systems, with results reported per script. Banks screen customers and payments against public sanctions lists, and an analyst clears each possible match by hand. Take entities in the public lists that already carry names in more than one script, such as Latin with Cyrillic, Arabic or Chinese, build checked pairs of matching and non-matching names, and report how often each matcher finds the true matches and rejects the false ones in each script.
Why it is still open
OpenSanctions Pairs (Chandler Smith, Friedrich Lindenberg and colleagues, February 2026) has 755,540 labeled pairs, but only 3.83% of them cross scripts, it reports no results by script, and it finds that large language models "struggle with cross-script transliteration". Resistance to deliberate evasion is outside its scope.
Built with what exists
The OpenSanctions Pairs data on Hugging Face and its code on GitHub (non-commercial license, CC BY-NC 4.0); the open nomenklatura matcher used as its baseline; open-weight models; a model that returns probabilities for fixed-answer questions such as "same entity or not" (see the Note "A thousand questions, one pass"); readers of each script to check the pairs.
First step
Download OpenSanctions Pairs, count its cross-script pairs by script, and have a reader of one script check 50 of them.
Done looks like
At least 500 checked pairs for each of at least three script groups, released under the same license, and a table giving precision and recall per script for at least three matchers.