top of page

MiCA · AIFMD II ready

AMLR 2027 ready

+370 600 26147

Vilnius · Lithuania

Pink Poppy Flowers

AML Teams · Audit · MiCA

· Backlogs · Payment Rails

MLRO · MSB Canada

· Access EU Market

MiCA · Casino · Bank & EMI

· iGaming · Immigration

Agentic AI in AML: Lessons from Hong Kong for EU Firms

1 day ago
6 min read

Most AML teams already use AI that writes: case summaries, adverse media digests, a first draft of a suspicious activity narrative. A person reads it and decides. Agentic AI works differently. You give it an objective, and it picks its own steps, calls other systems and acts on what comes back.

In November 2026 the Hong Kong Monetary Authority (HKMA) published a Discussion Paper on Agentic A.I. Governance in the Banking Sector. It is not EU law, and the HKMA says plainly that the paper “does not itself introduce binding supervisory requirements”. It is still worth an EU compliance officer’s afternoon. It is written entirely about agents, it goes into operational detail that general AI guidance skips, and its answers sit close to what the EU Anti-Money Laundering Regulation (AMLR) requires from 10 July 2027.

Our reading of agentic AI in AML, in one line: let the agent do the legwork, keep the decisions with people, and be ready to prove who owns the agent, what it is allowed to do and how you would stop it.


What the HKMA counts as an agent

The paper defines agentic AI as “a system that pursues an assigned objective by planning and executing a sequence of actions, using tools and other systems, and adapting as conditions change.” It deliberately leaves out robotic process automation whose steps are fully prescribed, single fixed-task models, and drafting or retrieval assistants that produce content for a person to act on.

That boundary is useful in AML. A tool that summarises an alert for an analyst is an assistant. A tool that pulls registry extracts, applies your triage rules and then routes or closes screening alerts by itself is an agent, and it needs a different level of governance.


Where agentic AI is turning up in AML work

The paper’s illustrative use cases include three that anyone who runs an alert queue will recognise: an agent that “assembles alert case files for an analyst”, one that “collects evidence, applies triage rules and routes screening alerts”, and one that “forms part of the transaction-monitoring layer”. On the onboarding side it describes an agent that “collects, verifies and files onboarding documents”. The HKMA stresses that these are illustrations, not a measure of how common real deployments are.

For a sense of pace, a research report released on 17 November 2026 by the Hong Kong Institute for Monetary and Financial Research found that 13% of the financial institutions surveyed had already implemented agentic AI and 56% were planning to. On compliance specifically, over 70% expressed interest in exploring it over the next three to five years.


The EU rule that already covers agents: AMLR Article 76(5)

EU firms do not need to wait for an agent-specific rulebook. Article 76(5) of the Anti-Money Laundering Regulation, Regulation (EU) 2024/1624, allows obliged entities to adopt decisions resulting from automated processes, including profiling, or from processes involving AI systems, on three conditions:

  • the system only processes data obtained under the AMLR’s customer due diligence chapter (Chapter III);

  • any decision to enter into, refuse or maintain a business relationship, to carry out or refuse an occasional transaction, or to increase or decrease the due diligence applied to a customer “is subject to meaningful human intervention to ensure the accuracy and appropriateness of such a decision”;

  • the customer can obtain an explanation of the decision and challenge it, except where it concerns a report to the financial intelligence unit (FIU).

Now hold that list against an agent’s to-do list. Gathering documents, screening names, assembling the case: no problem. Downgrading a customer from high to standard risk, approving an onboarding, or quietly keeping a relationship open after a screening hit: those are exactly the decisions that need a person who has actually looked at the file.

Article 18 matters when the agent comes from a vendor. Firms that outsource AML tasks “remain fully liable” for what the service provider does, and some tasks can never be outsourced, including the decision on a customer’s risk profile, the decision to enter a business relationship, FIU reporting and the approval of the criteria used to detect suspicious transactions. Article 18 was written for service providers, not software, but its list is a sensible map of where an agent should stop and hand over.


Five ideas from the HKMA paper worth borrowing

One named owner per agent

“Each Agentic A.I. use case or deployment should have one designated accountable owner,” the paper says, and accountability should sit with “an identified person or function, rather than displaced by a vendor, system or technical component.” In practice that means a name on a page, with the business purpose and the limits written next to it.

Oversight that is more than an approve button

Human oversight “is not established merely by adding a formal approval step,” according to the HKMA. It has to be timely, informed, exercised by someone with the authority to act, and capable of preventing, pausing or reversing harm. The paper also warns against treating the existence of an approval step as oversight in itself. An analyst clicking through 200 pre-filled alert closures before lunch is not “meaningful human intervention” under Article 76(5).

Limits that do not depend on the model

The paper notes that “controls that rely solely on the model to identify and comply with critical constraints may fail”. It recommends deterministic checks on things like delegated authority, limits and approvals, plus velocity limits on “the number, value, frequency, customer reach or cumulative impact of actions”. For an alert-handling agent, a hard daily cap on automatic closures and a rule that nothing touching a sanctions hit closes without a person are cheap insurance.

A stop button you have tested

The HKMA expects an effective stop mechanism, the familiar “kill switch”, that can bring an agent “to a safe state promptly”, and it expects that mechanism and the fallback arrangements to be tested. The fallback is the part AML teams forget. If you switch the agent off on a Monday morning, who works its queue?

Records an inspector can follow

The paper talks about “maintaining sufficient records to support auditability, tracing and post-incident investigation”. It is also blunt that a fluent explanation from the model “may not provide a reliable account of how it was generated”. Logs of what the agent looked at, which tool it called and what it changed are evidence. The agent’s own account of its reasoning is not.


Before you switch an agent on

  1. Write down the agent’s purpose, its permissions and the systems it can touch, and name the accountable owner.

  2. Mark every point where its output turns into an Article 76(5) decision, and put a trained person there with enough time to review properly.

  3. Set hard limits outside the model: volumes, values, customer segments, and alert types it may never close.

  4. Sample its closed work through four-eyes review, heavily in the first weeks, and compare its error rate with your human baseline.

  5. Test the stop mechanism, and have people ready to take over the queue.

  6. Check the vendor contract against Article 18: your policies apply, you can audit, and the supervisor can retrace what happened.

  7. Keep logs good enough for an inspector to rebuild any single decision.


Frequently asked questions

Does the HKMA paper apply to EU firms?

No. It is a Hong Kong discussion paper and, in its own words, “not a statement of supervisory expectations”. The HKMA wants first-round comments by 1 February 2027. For EU firms it is a well-argued benchmark. The binding rules come from the AMLR and national law.

Does the AMLR ban AI agents in KYC or transaction monitoring?

No. Article 76(5) expressly allows decisions resulting from automated processes and AI systems, on the conditions set out above.

When does Article 76(5) start to apply?

Together with the rest of the AMLR, from 10 July 2027 for most obliged entities. Article 90 sets a later date, 10 July 2029, for two specific categories of obliged entity listed in Article 3.


Agents change the work, not who answers for it

Automation rarely removes AML work. It moves it from collecting documents to checking what a machine concluded, which needs more experience, not less. We wrote about that shift in why senior human analysts matter more as AI agents and instant payments spread, and the vendor side of the question is covered in our piece on EBA and ESMA outsourcing requirements for fintechs.

CAML’s analysts have 5+ years of AML experience, our Team Leads are ACAMS-certified, and critical tasks go through a four-eyes review. If you are piloting an agent in onboarding or alert handling and need people to review its output, or to take over when you press stop, get in touch with CAML.

Comments


bottom of page