top of page

The full briefing: AI in KYC and Remediation

What the authoritative sources actually say — and the guide nobody has written yet

A short, sourced reading of where AI sits in KYC and customer-file remediation: what the standard-setters have legitimised, what the independent evidence base actually shows, and how the regulatory position reads across Australia, Singapore and the UK. Primary sources only; everything is listed at the back, and anything we could not confirm is named as unconfirmed.

1 · The claim

There is no single authoritative guide to “AI in KYC and remediation” — and remediation specifically is almost entirely vendor and consultancy territory.

That sounds like a problem. It is actually the most useful thing to know before you buy anything in this market, because it tells you what to trust. No standard-setter has published a remediation playbook, so anyone who hands you one is giving you their opinion dressed as a standard. What does exist is a credible three-layer stack you can stand on: standard-setter principles that legitimise AI in financial-crime compliance, an independent evidence base with quotable numbers, and jurisdiction-specific regulatory positions. Anchored to those, an AI-assisted remediation reads as governed rather than clever — which is the only version a senior compliance reader will accept.

The single most important line in the whole evidence base is quiet: data quality and legacy systems are the documented barriers to scaling AI. Not the model, not the algorithm — the data. That is the remediation case, stated independently by people with no product to sell.

This issue is the map of that three-layer stack, the market claims around it that don’t survive checking, and what a compliance function can do about it now without buying anything.

2 · The evidence

2.1 Layer 1 — the standard-setters legitimise it, on conditions

  • Wolfsberg Group — Principles for Using AI and ML in Financial Crime Compliance (2022). The closest thing to an authoritative guide, written by twelve global banks. Five principles: legitimate purpose; proportionate use; design and technical expertise; accountability and oversight; openness and transparency. Principle 4 is the commercially decisive one — a vendor cannot absorb the client’s accountability. Whatever the tooling, the institution owns the outcome.

  • Wolfsberg Group — Effective Monitoring for Suspicious Activity, Part II: Transitioning to Innovation (Aug 2025). The freshest tier-one material in the field. Three pillars — transition and validation; balancing model risk against financial-crime risk; explainability — developed with FIUs, policymakers and supervisors, and explicitly calling for innovation-friendly regulation. This is the strongest answer to “our regulator won’t like it.”

  • FATF — Opportunities and Challenges of New Technologies for AML/CFT (2021). Dated on the technology, but it is the FATF position. Cite it for standing, not currency.

2.2 Layer 2 — the independent evidence base

The CCAF Cambridge 2026 Global AI in Financial Services Report (with the BIS, IMF, WEF, IDB, CGAP and Arab Monetary Fund) surveyed firms, vendors and 130+ regulators:

  • Firms adopting AI at some level: 81%

  • Firms at advanced adoption: 40%

  • Firms actively adopting agentic AI: 52%

  • Fintechs vs incumbents (advanced): 47% vs 30%

  • Regulators piloting/deployed AI for AML/CFT supervision: 27%

  • Top barriers to scaling: Data quality, talent, legacy systems

  • Top risk concerns: Data privacy (74%), hallucination (70%)

Two numbers do the work. 27% of regulators are already piloting or deploying AI for AML/CFT supervision — the people examining you are adopting the same tools. And the top barrier to scaling is data quality — which is precisely what a remediation fixes.

2.3 Layer 3 — the regulatory position across the three markets, one direction

The mechanics differ by market, but they point the same way: the duty that bites is continuous — ongoing monitoring of the relationship — not a one-off re-verification deadline; and each supervisor treats AI as legitimate if it is governed.

  • Australia — AUSTRAC. There is no blanket deadline to re-verify the back book. Pre-commencement customers can generally continue without initial CDD; initial CDD is event-driven (an SMR obligation arises, or a significant change pushes the relationship to medium/high ML-TF risk). Ongoing CDD applies to everyone. Customers who completed an applicable identification procedure before 31 Mar 2026 are treated as compliant; new professional-services and real-estate categories run from 1 Jul 2026.

  • Singapore — MAS. The binding rule for banks is MAS Notice 626; its amendment of 30 June 2025 (in force 1 Jul 2025) sharpened the obligation to identify, assess and understand money-laundering, terrorism-financing and proliferation-financing risk, alongside CDD, ongoing monitoring and five-year record retention. On AI, MAS is supportive within governance: its FEAT principles (fairness, ethics, accountability, transparency) and the Veritas toolkit set the expectation that AI in finance is governed and explainable, not prohibited.

  • United Kingdom — FCA / Bank of England. Both are technology-agnostic — they neither prescribe nor prohibit AI, applying an outcomes-focused standard equally to it. Concretely, the FCA has opened AI Live Testing (FS25/5): firms trial AI in live markets alongside the regulator, with applications open to 24 March 2026, and Parliament’s Treasury Committee has reported on AI in financial services. The underlying AML duty (the Money Laundering Regulations) already requires ongoing monitoring of the business relationship.

  • EU & US — context. EU AI Act: AML and fraud detection are not automatically high-risk (Recital 58, Reg (EU) 2024/1689; Commission clarification, May 2025), though AI-literacy and transparency duties still apply. US: the FinCEN-led 2018 Joint Statement confirms an unsuccessful pilot will not, on its own, draw supervisory criticism.

The through-line: every market points the same way — AI is legitimate if it is governed, the binding constraint is data, and the obligation that bites is continuous, not a one-off deadline. And the supervisors are leaning in, not back: the FCA is testing AI in live markets, and 27% of regulators are already piloting or deploying it for AML/CFT supervision.

3 · What to be sceptical of

Five claims circulate in this market that don’t survive checking:

  1. “AI removes the compliance burden.” It does not. Accountability is non-delegable — Wolfsberg Principle 4. AI changes the economics of the work; it does not move the responsibility off the institution.

  2. “Everyone is already doing this.” 40% at advanced adoption means most are not. “Everyone’s doing it” is a sales line, not a fact.

  3. “The reform requires re-verifying every customer by a deadline.” In none of the three markets is that the rule. Australia has no blanket re-verification deadline (event-driven initial CDD plus continuous ongoing CDD); Singapore (MAS Notice 626) and the UK (the Money Laundering Regulations) likewise impose ongoing monitoring, not a one-off re-verification cliff. Anyone selling you a fictional deadline is selling fear — and a competent compliance head will catch it.

  4. “The regulator has approved this approach.” No regulator approves vendors or methods. A claim of endorsement is the fastest way to lose a reader who checks — and they all check.

  5. “The EU AI Act doesn’t apply to us.” Too neat the other way. AML/fraud detection isn’t automatically high-risk, but AI-literacy and transparency duties still apply, and adjacent credit or access decisions can pull a system back into scope.

4 · What to do about it

Four things, none of which require buying anything:

  1. Map your programme to Wolfsberg’s five principles, and write the remediation-specific answer under each. For accountability: state plainly what the institution retains, what evidence it keeps, and what the audit trail looks like when a file is cleared. If you can’t answer under a principle, that’s your gap — found before a supervisor finds it.

  2. Treat the data-quality finding as your business case. The independent barrier to scaling AI is data and legacy systems (CCAF, with the BIS and IMF). That is the honest argument for a remediation — not that AI is magic, but that it makes clearing an unreliable book economically viable where manual review never was.

  3. Get your jurisdiction’s trigger model right before you plan anything. Australia is event-driven initial CDD plus continuous ongoing CDD; Singapore runs on MAS Notice 626’s risk-based CDD and ongoing monitoring; the UK on the MLRs’ ongoing-monitoring duty. Different mechanics, the same continuous obligation — plan against the actual rule, not a headline.

  4. Keep accountability in-house, and watch the supervisors. 27% of regulators are adopting AI for supervision themselves; the examination is changing. The right posture is to govern your own use to a standard you’d be comfortable having examined — because increasingly it will be.

There is no authoritative guide to running an AI-assisted KYC remediation; this is a reading of what the governed sources actually support, from the public record, so a reader can draw their own line. It is the map of a field you are being sold into — and the standard the rest of this series holds itself to.

Sources

Published by Dligence as part of the “What Good Looks Like” series. Dligence is a compliance-data company; this paper is a reading of the public record and standard-setter guidance, is not legal advice, and no part of it implies regulatory endorsement of any provider or method.

bottom of page