Attackers don't have a compliance deadline. Banks now have two.

Two clocks run in every European bank.

The first is regulatory. The Digital Omnibus on AI came into force on 27 July 2026 and moved the high-risk deadline from 2 August 2026 to 2 December 2027, with product-embedded systems pushed out to 2 August 2028. It slipped before it ever took effect. Regulations get postponed; this isn't a scandal.

The second belongs to whoever shipped this week’s build of the remote-access toolkit now pointed at your mobile channel. It files no conformity assessment, observes no transitional period, and keeps perfect time. 

Cleafy sits on both. Our work fuses cyber and fraud intelligence, which puts us where the two meet. From there, the delay is not the part worth your attention.

Recital 58 and Annex III point 5(b) carve fraud detection out of the creditworthiness item explicitly. It was never in scope.

There is a catch, and it matters: The exception holds only where fraud detection is the main intended use. Bolt a fraud module onto a model that also scores creditworthiness, and you can lose it.

So the compliance deadline that was meant to push banks into fixing their AI decisioning has, for fraud, quietly gone.

Sixteen months is a long time in which nobody has to answer a hard question about how their scoring works.

The vendors selling black-box decisioning are the clearest beneficiaries. If you are in an evaluation now, some of them are counting on it. Expect "we're not high-risk under the AI Act" to be the most-quoted line in the category by Q1. It is usually true, and it settles nothing you need to know.

Logged, timestamped, replayable, human-authored decisioning earns its place on defensibility. It did so before the Act existed. A customer disputes a blocked payment. An ombudsman asks why an account was frozen. Your own second line asks what changed between March and June. A score and a confidence value answer none of them.

CERTFin's 2025 data puts payer manipulation in 76% of effective retail banking frauds, and 86% at the execution stage. The technique is cyber, the loss is fraud, and the adversary works to no calendar.

A bank that upgrades its decisioning only when a regulation forces it has just been handed the time to fall further behind.

The regime you have just been exempted from still spells out the standard, in operative text. Article 26(6): logs kept for six months or more. Article 26(2): human oversight by people with real authority. Article 27: a documented account of the risk to affected people. DORA Article 17: structured incident records, on a four-hour clock.

None of that binds your fraud stack in December 2027. All of it describes what a defensible fraud decision looks like.

The next liability regime will be less forgiving. The Payment Services Regulation attaches refund liability to monitoring failure. When it applies, "the model said so" becomes a position you defend with money.

And this was the extension, not the first of several. The conditional trigger in the Commission's original proposal, which would have tied the start date to the readiness of standards, did not survive trilogue. The date’s fixed now.

Adopt it as a procurement standard now, while no deadline is distorting the decision.

Ask any vendor for the decision record on one specific alert: the inputs, the reasoning path, the reviewer, the time, exportable. Ask which provision they rely on for their classification, and for the intended-purpose documentation behind it. Ask for known limitations and failure modes in writing, which Article 25 as amended says they owe you. Ask who becomes the provider if you retrain on your own book, because under Article 25(1)(b) the answer may be you.

Most vendors manage three of those. A black box cannot answer the first one at all, which is why you lead with it.

None of this means ripping out your behavioural analytics or your transaction monitoring. Those layers produce signal, and more signal helps. The question is what happens where signal becomes a conclusion, and a conclusion becomes an action against a customer's account.

Explainability bought in 2026 is an architecture decision. Demanded in 2027, it is a migration.

The Act’s calendar is no longer your reason to move. What you can prove about a single decision, to a customer, an ombudsman or a court, on any day of the year, is the reason. Choose for that now, while the choice is calm.