The compliance industry’s enthusiasm for artificial intelligence in anti-money laundering (AML) work is understandable. AI can process transaction volumes no human team could match, flag patterns manual review would miss, and cut the time cost of due diligence.
Leo RegTech, which integrates tools such as Vartion Pascal for KYC data and Yoti ID for digital identity verification, has been building in this direction for years. But the industry, one veteran compliance lawyer argues, is not being honest about how often AI still gets things wrong, and in AML the consequences of that could be severe.
Leo RegTech recently discussed the point that whilst AI is transforming AML compliance, it should not be making the decisions.
A Microsoft study, DELEGATE-52, tested 19 AI models across 52 professional domains and found frontier models, including versions of GPT, Gemini, and Claude, lost an average of 25% of document accuracy across 20 delegated interactions, with degradation reaching 50% overall.
A separate Stanford study found AI legal research tools from LexisNexis and Thomson Reuters hallucinated between 17% and 33% of the time. Applied to AML files, where PEP status, beneficial ownership, sanctions exposure and source of wealth all require multi-step, judgement-heavy analysis, that error rate becomes a serious regulatory liability.
Effective AML compliance rests on a firm’s Business-Wide Risk Assessment and client risk appetite framework, built under regimes including the UK Money Laundering Regulations, FinCEN’s Customer Due Diligence rule, and the EU’s AML Directives. Any AI tool used in this process must ingest that framework rather than operate on generic assumptions.
The AML technology market, projected to reach $9.4bn by 2030, includes vendors such as Quantexa, valued at $2.6bn and eyeing a 2026 IPO, Feedzai, valued at around $2bn, Napier AI, and NICE Actimize. None currently offers a standalone compliance decision on client KYC.
Law has already learned this lesson. In Mata v. Avianca, two US attorneys were sanctioned over fabricated ChatGPT-generated case law. In April 2026, Sullivan & Cromwell apologised to a US Bankruptcy Court judge after an AI-generated motion contained inaccurate citations. Over 1,300 AI hallucination incidents have now been logged in legal filings globally.
Under the EU AI Act, high-risk provisions covering AI-driven creditworthiness evaluation have been pushed back from August 2026 to December 2027 following May’s Digital Omnibus agreement, though standalone AML monitoring sits outside that high-risk category. Penalties for non-compliance where it does apply reach €35m or 7% of global turnover.
The conclusion is not to avoid AI, but to keep a qualified human making the final call on risk classification and SAR decisions, with AI acting as accelerator rather than decision-maker.
Read the full Leo RegTech post here.
Copyright © 2026 FinTech Global
