Investigation report, 29 Sept 2026
What have recent advances in AI for payment fraud and financial-crime detection made reliable, what still fails, and what products could a startup or bank build now?
The most dependable result is machine-learning triage placed after a bank's existing alert rules. At two banks it removed most false alerts while keeping about 90% of true cases.1,2 Network features added to standard tree models also work, and fast enough for real time.3,4 Three things still fail. Large language models make poor fraud classifiers.5,6 Transaction foundation models did not work for money laundering.7 Most models break when crime patterns or labels shift.8,9
Recommendation: a UK bank or payment firm should first build or buy mule-account scoring for incoming payments, with an explanation for each flag. Under UK scam rules, the receiving firm now pays half of each reimbursement.10 A live bank pilot of this kind of scoring beat the bank's rules.11
Caveat: most results are offline, from a single institution, and reported by the authors themselves.
Works citing the seeds 47 · Related works 38 · Business and economics papers 24 · Adjacent-topic search 50 · Industry whitepapers (web search) 107
Every source was screened from its title and abstract before any was read; 30 shortlisted sources could not be fetched (no open copy or unreadable). 25 verified claims came from the sources read.
240 checks run against the source passages; 46 claims caught and corrected on the way.
Distilled into 6 findings, 4 opportunities, 3 combinations.
A model that re-scores rule alerts removed 80% of false positives while keeping over 90% of true positives at one bank. An independent study at a second bank cut false positives by 0.76 (as a share) at 0.90 recall. Both are offline tests on alert sets dominated by false positives.1,2+3
Adding subgraph features to boosted trees beat graph neural networks by up to 8% on synthetic AML data, and small-batch CPU latency averaged 30 ms. On real Capital One data a graph model beat production XGBoost; the authors credit the graph formulation, not model depth.3,4+3
WeChat Pay, Revolut and Visa report large relative fraud gains over their production models, one in a live A/B test. Revolut's model scored 47.1% below its AML baseline, blamed on missing cross-account network features. All figures are relative and unreproducible.7,18+2
Untuned frontier LLMs scored low F1 on real e-commerce fraud, and even retrieval-aided LLMs trail a tabular model. LLM-written mule-case narratives took seconds per alert on one GPU and a bank triage agent beat its legacy flow, but both need human review.5,6+5
A model trained on one synthetic laundering mix scored F1 near 0 on another until retrained. Temporal shift degraded fraud models, labels are sparse and noisy at real banks, and federated training across banks helped only slightly.8,9+4
LexisNexis/Forrester put global financial-crime compliance cost at $206B, with salaries the main driver in most regions. UK rules now make receiving banks fund 50% of scam reimbursements up to £85,000, and the regulator says no bank detects all such fraud.10,30+3
Each is rated on evidence, technical readiness and commercial access separately; there is no overall score. Hover a rating, or open “Why these ratings”, for the reasoning.
A scoring layer that sits after a bank's existing monitoring rules, ranks or auto-closes low-risk alerts, and drafts an explanation of each score for the analyst. Rules stay in place, which keeps the audit trail regulators expect.1,2+6
Main risk. Narratives that miss facts, and labels built from past investigator decisions, can hide missed laundering and draw regulatory criticism.
A mule-risk score for incoming Faster Payments accounts that combines account behaviour with network features and explains each flag, so the receiving firm can hold or freeze funds before they move on.4,10+5
Main risk. Fraud may pivot to international payments outside the scheme, shrinking the liability that motivates buyers.
A streaming service that computes fan-in, fan-out, cycle and other subgraph features per transaction and feeds them into the boosted-tree models banks already run, without replacing those models.3,4+5
Main risk. Gains measured on synthetic laundering patterns may shrink on real data where launderers blend in.
A validation kit that runs a bank's model against synthetic scenarios: rarer crime, new laundering mixes, time drift, scarce or noisy labels, and group fairness. It outputs a model-risk report.8,26+6
Main risk. Without proof that synthetic results predict real performance, buyers may treat the report as a checkbox.
Two concepts from different sources, tied into an idea neither states. These are hypotheses to test, not findings.
An LLM scambaiting system engaged real scammers and extracted financial details such as mule accounts in 31.74% of conversations where scammers replied.47,48,49
From 'Send to which account?' Evaluation of an LLM-based Scambaiting SystemA bank's mule classifier works in production, but its labels come from police filings and analyst confirmations that arrive late and are biased.11,29
From Detection, Attribution, Narration: An End-to-End Pipeline for Explainable Money Mule IdentificationFeed mule accounts harvested by scambaiting agents, shared through industry channels, into receiving banks' mule models as early positive labels and watchlist hits. This targets the label delay that weakens today's mule models, and catches accounts before any victim reports them.
The scambaiting work stops at intelligence sharing; the mule-detection work relies on slow police and analyst labels. Neither links the two.
A bank transaction foundation model failed on AML, blamed on missing cross-account network features; its kind of embeddings can be cached and fused live.7,50
From PRAGMA: Revolut Foundation ModelA graph-feature preprocessor computes subgraph patterns per transaction at real-time latency and lifts tree models sharply on laundering data.4,38
From Graph Feature Preprocessor: Real-time Subgraph-based Feature Extraction for Financial Crime DetectionBuild the AML model as a tree model that takes both per-customer foundation-model embeddings (behaviour over time) and streaming subgraph features (who pays whom). Each supplies what the other lacks: behavioural history versus network structure.
Foundation-model papers leave graph modelling as future work, and graph-feature work uses no learned behavioural embeddings.
Banks can find cross-institution laundering groups by sharing only hashed counterparty summaries, not raw transactions; tested on real Alipay data.51,52,53
From Towards Collaborative Anti-Money Laundering Among Financial InstitutionsUK rules make receiving firms pay half of each APP scam reimbursement, explicitly to give them an incentive to detect the mule accounts they host.10,33
From PS25/5 APP scams reimbursement requirement consolidated policy statementOffer a shared service where sending and receiving firms exchange hashed counterparty sketches at payment time, so both sides can spot mule clusters that span firms without breaching data-protection rules. Shared liability gives both sides a reason to join.
The cross-bank method was pitched for AML research, not tied to a liability rule that pays both parties to participate.
The detailed analysis behind the cards above, section by section.
Question. The prompt supplied this question. It asks what AI for payment fraud and financial-crime detection now does reliably, what still fails, and what a startup or bank could build now.
Evidence base. We read all the seed papers. They cover graph methods for anti-money-laundering (AML), synthetic data and benchmarks, transaction foundation models, and LLM agents. We screened a further set of candidate sources and read five of them:
Strength. The evidence on alert triage is strong, because two banks report similar results. Other evidence is moderate:
Not reached.
Every citation marker opens the passage it rests on in one of these. Sources the agent added show why, and web pages show where and when they were read.
Start with a few seed papers and a question. The free preview gives you the answer and a couple of findings; the full report is the same run, continued.