← Back to Resources

KYC · Responsible AI

Responsible AI in KYC: let the machine prepare, let people decide

The AMLR allows AI in customer due diligence, with conditions attached. What responsible AI looks like in day-to-day KYC work, and how to keep every decision defensible.

28 September 2026 · David Zwarts · 7 min read

Compliance reviewer comparing an ownership network with a paper case file
Illustrative photograph

Most compliance teams no longer ask whether to use AI in KYC. They ask whether they can defend every decision the AI touched. From 10 July 2027, the EU Anti-Money Laundering Regulation (AMLR) makes that the question that counts.

What the AMLR actually allows

The AMLR does not ban AI. Article 76(5) allows decisions resulting from automated processes, including profiling and AI systems, subject to three cumulative conditions:

  • Stay within due diligence data. The system may only use data collected under the regulation’s customer due diligence chapter.
  • Keep a human meaningfully involved. Decisions to accept or refuse a client, carry out or refuse a transaction, or increase or decrease due diligence measures need meaningful human intervention.
  • Be able to explain it. Customers can request an explanation and challenge the outcome, except where the decision relates to a suspicious transaction report.

AML systems are not named as a high-risk use case in Annex III of the EU AI Act. That is not a free pass: whether another AI Act category applies depends on the system’s actual use, and data-protection and AMLR safeguards still matter.

Institutions in the Caribbean that work with EU banks, fund administrators or parent companies can also expect counterparties to ask about comparable safeguards.

Four principles that work in practice

  1. Explain at the right level. A reviewer needs to know why a particular screening hit was flagged. A compliance officer needs to know how the setup behaves. A supervisor needs documentation, test results and version history. Design explanations for each audience.
  2. Test for fairness, not just accuracy. Name matching across alphabets, country-based risk factors and historic case data can skew results. Measure how false positives are spread across client groups and act on what you find.
  3. Make human oversight real. The reviewer needs the evidence, time and authority to disagree. A reviewer approving 200 recommendations an hour is not meaningfully intervening. Make it as easy to disagree as to agree, and record both.
  4. Keep the audit trail complete. Show what the system produced, what the reviewer changed and why. That record helps answer questions from customers and supervisors months later.

Where AI adds the most value in KYC

The biggest gains are in preparation, not judgement:

  • reading and extracting data from company documents and ID;
  • untangling layered ownership structures and calculating indirect interests;
  • summarising adverse-media results with links to the original sources;
  • drafting a first assessment for a reviewer to challenge.

This is where analysts lose many hours. It is also where AI can assist within the AMLR’s safeguards without taking over the decision.

How Vitauri approaches it

Vitauri KYC is built around a simple rule: AI-powered clarity, human-led decisions. AI prepares the ownership analysis, links findings to source documents and drafts the assessment. Suggestions appear beside the evidence, where an analyst can accept, edit or override them. Nothing is applied automatically. Prompts, outputs and analyst decisions are logged separately, so the reasoning behind each file can be shown later. The platform is hosted in the EU (Frankfurt).

See it in practice

Want to see responsible AI in a live KYC file? Book a 30-minute demo and we’ll walk through a case built around your client types.

Book a demo ↗

Sources

Last reviewed 28 September 2026. General information, not legal advice.