The opinions expressed here are those of the authors. They do not necessarily reflect the views or positions of UK Finance or its members.

The grooming call happened last week. The new payee was added on Tuesday. The receiving account went quiet for nine months and woke up on Monday. By the time a payment reaches the rail, the decision that mattered has already been made, and the person who made it had been worked on for a month.

UK Finance put payment fraud losses at £1.28 billion in 2025, up four per cent, with APP fraud accounting for £576.4 million of that across 248,070 cases, up 19 percent. Each of those cases is someone who was persuaded. Detection determines how many there are.

Five shifts in detection are worth understanding.

1. Detection moves ahead of the payment

Legacy controls flag activity at the point of transaction or after it clears. That leaves a decision window measured in seconds, followed by an expensive cycle of recovery and reimbursement. Continuous analysis of account behaviour changes the timing. Risk factors surface as they emerge, which gives a fraud team days rather than milliseconds.

Earlier detection is only half the problem. What a bank can lawfully and proportionately do with a signal five days out is the other half, and it is a harder question than the modeling. Part two takes it on directly.

2. The whole journey, not sampled points

Fixed API calls at login and at payment sample a customer journey. They do not describe it. Sixty-six per cent of APP fraud cases in 2025 started on online platforms and a further 17 percent through telecoms, which means the signal a bank needs often sits in the app session or the call centre, not the payment message. Correlating device, behaviour, channel, and account history into one continuous view is what makes that signal visible.

It is also a data protection question before it is a modeling question. Continuous behavioural monitoring across an entire customer base needs a lawful basis, proportionality, and retention settled with your DPO at design time, not at review.

3. Normal is a baseline, and authorised fraud breaks it quietly

Models that learn how an individual actually uses their bank build baselines that are hard to spoof. Bots and synthetic identities move through forms differently from people, and that shows up fast.

Authorised fraud is the harder case. The customer is genuine, on their usual device, passing every authentication step, doing exactly what the criminal coached them to do. There is no artificial behaviour to isolate. What changes is the account: a new payee added after a long phone call, a login pattern that drifts over weeks, hesitation where there was none. Detecting a coerced customer is a different discipline from detecting a bot.

4. The receiving account is your exposure too

Every scam needs somewhere for the money to land. Under the reimbursement rules, the cost of an APP claim splits evenly between sending and receiving institutions, so inbound mule activity sits squarely on your own book.

Mule accounts are difficult to catch at the transaction, because the first fraudulent credit looks ordinary. The signals arrive earlier: dormancy ending, changed device and login patterns, small test credits, rapid onward transfers. Reading the receiving account as a sequence allows it to be classified before the credit lands. A rule waiting on the payment cannot do that.

5. Coverage, not just speed

The advantage generative models hand criminals is volume. One operator can hold hundreds of grooming conversations in parallel, and cloned voice extends the same economics to the phone. UK Finance found telecoms behind 17 percent of APP cases in 2025 but 28 percent of losses, which is what better-prepared, higher-value social engineering looks like in the data.

That problem does not sit at the payment. It sits in the number of accounts somewhere inside a scam's preparation phase on any given day, where nothing has yet happened that would trigger a rule. What changes operationally is the queue itself: instead of a list of flagged payments, the fraud team works a ranked list of accounts whose behaviour has moved, re-ranked as it keeps moving. Whether that is more work than a payment alert queue or less depends entirely on how well the ranking holds up, which is the question to put to anyone selling it.

Speed still matters once a payment is in flight. Coverage decides whether anyone saw it coming.

That coverage is what Acoru builds. Continuous Account Intelligence watches accounts across digital, call centre, and payment channels, and classifies each one as regular, victim, or mule while a scam is still being set up. Learn more at acoru.com.