One Big Idea

The biometric matched. The device was trusted. The customer approved the payment themselves. The money left irreversibly in under a second — and no control failed. On an irreversible rail, a control that passes is not the same thing as a bank that is protected.

Last week this letter asked whether the answer your assistant just handed a customer still applies today. This week the question moves from what a system says to what it can see — and there is no attacker in it at all, which turns out to be the harder problem. A customer authorises a payment on their own phone, with their own biometric, at an ordinary hour. Every control fires. Every control passes. This issue is about why that is not a detection failure, and about the half of a payment defence most banks have never built.

The Insight · A control is worth what it can see

Authentication answers one question: is this person who they claim to be? On an irreversible rail, the expensive question is different — has that person been directed by someone else? No impostor control, at any budget, can answer it.

Money leaves an account wrongly in only two ways. In the first, an impostor acts without consent — stolen credentials, a hijacked device, a swapped number. The deception targets the bank's systems, so the bank's systems can find it. There is a lie inside the transaction, and fraud detection is built to find lies.

In the second, the owner acts, deceived. The deception targets the customer, not the bank, and it succeeds before the payment exists. By the time the payment reaches any control, the lie has already landed somewhere the bank cannot observe — a phone call, a message, a relationship. The device is genuine, the credentials are genuine, and the intent, in the only sense a system can measure, is genuine. The transaction itself is clean, so no guardrail watching it has anything to catch.

Instant settlement is what made this expensive. On an older rail a transfer sat somewhere for a few hours; that gap was never a design feature, but every fraud control a bank built quietly depended on it. Somebody could look, ask, hold, or reverse while an answer came back. Banks responded to instant rails reasonably — controls earlier, controls faster — and the losses kept climbing anyway. That is not a failure of execution. It is a failure of diagnosis.

Across the European Economic Area in 2024, on credit transfers, payment service users themselves bore approximately eighty-five percent of the total fraud loss — "mainly as a result of scams that tricked users into initiating fraudulent transactions", in the supervisors' own words (joint EBA–ECB report, 15 December 2025). Read that precisely: it describes who absorbed the loss, not a claim that any single fraud type is eighty-five percent of the total, and it is not a rate any individual bank should expect. Strong customer authentication did not fail in those cases. It passed, confirming accurately that the right person was instructing the bank.

Their finding on authentication is worth quoting rather than paraphrasing, because the careful version is the useful one: SCA "remains effective against the fraud types it was designed to mitigate... especially for card payments", while for credit transfers "this effect was less clear". Not bypassed — less clearly effective, against a fraud type it was never built to see.

Framework of the Week · The Payments Defense Grid

This is the author's proposed decomposition — not a validated standard, not six products to buy, and not a maturity ladder. One axis is what a control can see: an impostor, or a deceived owner. The other is what it can still do by the moment it fires — prevent before value moves, interrupt at the moment of intent, absorb after finality. Six cells, and every control a bank operates sits in exactly one:

  • D1 · payer and device trust — biometrics, presentation-attack detection, device binding, malware and remote-access detection. The regulatory floor
  • D2 · payee intelligence — not who is sending but where the money is going: payee-name verification, destination and mule scoring, account age, first-time-payee status. The highest-leverage cell, and no single bank can build it properly
  • D3 · the in-flight decision — scoring inside the authorisation path, within the scheme's latency budget, with a defined behaviour when the model is unavailable
  • D4 · customer-in-the-loop friction — specific warnings rather than generic ones, questions that interrupt a rehearsed story, holds and cooling-off on new payees and new device bindings. The only cell that can reach a person while they are being manipulated, and the only one with a direct revenue cost
  • D5 · post-finality recovery — cancellation, freezing, tracing. Be honest: on an instant rail this is largely a formality, because funds cascade onward within minutes
  • D6 · network signal and shared liability — cross-institution mule intelligence, and an allocation of loss that gives a receiving bank a reason to act. The cell that changes the economics of the other five

Two disciplines keep it from becoming decoration. Nobody progresses from D1 to D6 — the columns are different problems, not stages, and reading them as a sequence rebuilds the error the grid corrects. And owning a product is not operating a cell: mark only the cells where something fires and somebody is accountable for the outcome. Most control inventories shrink considerably under that test.

Most banks run the entire impostor column, often to a high standard, and then count it as coverage. The deceived-owner column is mostly unbuilt. That is not a decision anyone made — it is the half that was measurable, purchasable and required.

Use Case · What moved when one regulator changed who pays

The obvious objection is that the second column is soft — that friction and payee checks are what you recommend when you have run out of engineering. So look at what happened when a regulator stopped treating this as a detection question.

From 7 October 2024 in the United Kingdom, reimbursing a scam victim became mandatory rather than discretionary, and the cost was split evenly between the sending and the receiving firm. Before that, the fourteen largest banking groups reimbursed about sixty percent of APP scam losses by value, with a sixty-one point spread between the best and worst firms — your protection depended largely on where you happened to bank. Eighteen months in, about eighty-eight percent of in-scope losses have been returned — three hundred and sixteen million pounds between 7 October 2024 and 31 March 2026, against one hundred and seventy-three million at the twelve-month mark, the same eighty-eight percent at both (Payment Systems Regulator, reimbursement dashboard updated 30 July 2026).

One honest note on that pair, since this letter keeps asking you to check denominators. They are not like-for-like: "in scope" is a category the mandate itself created, so no pre-mandate number can be an in-scope rate, and the periods do not overlap — the sixty percent stops on 7 October 2024, the day the requirement began, while the eighty-eight percent runs from there. The regulator declines the same comparison, noting that it "can't make direct comparisons, due to difference in scope". Read it as direction, firmly established, rather than as a precise delta. What is not in question is the mechanism: the detection technology did not change. Who carried the cost did — and behaviour followed the liability, quickly.

The same data carries a structural lesson, and it is the finding most likely to redirect a roadmap. Non-directed payment providers — those outside the fourteen largest banking groups — took about a third of all scam value while handling roughly a fifth of consumer Faster Payments by value, in the period before mandatory reimbursement, and the regulator separately reports their fraud rate by value as forty-one times higher than the largest groups. That multiple is its own published measure, not the ratio of the two shares, which is nearer twofold. Criminals were not defeating the biggest banks' models — they were shopping for the weakest place to land the money. The leverage is at the destination, not the sending side.

One caution, reading this in Asia. That concentration comes from a market with a long tail of small payment firms and e-money institutions. A rail dominated by licensed commercial banks under a common identity mandate may distribute mule accounts differently — the problem may concentrate inside institutions rather than between them, which changes the control from shared intelligence to internal account-cluster monitoring. That changes the control, not the conclusion, and it is worth answering with your own data rather than borrowing another market's structure.

Risk Note · The mandate is one half, and blocking is not free

Get the Vietnamese position right rather than dramatic, and get the dates right rather than approximately right. Thông tư 77/2025/TT-NHNN, issued 31 December 2025, amends Circular 50/2024 and requires both presentation-attack detection and automated device-integrity checks. Its commencement is staged, and the staging does not fall where a skim puts it. PAD sits in Article 7 of the amending circular — certified to ISO 30107 Level 2 or equivalent — and the device checks in Article 5, so both take effect on the general date, 1 March 2026. What is deferred is only Articles 3 and 10 of that circular, the identity-change confirmation methods and the revised Appendix: to 1 July 2026 for providers serving both individual and organisational customers, and 1 October 2026 for organisation-only providers. If you read the July date as covering your biometric programme, you will defer the wrong work by four months. Confirm against the gazette text rather than against a newsletter.

Either way, that is a mandate on one half of this grid. The other half is where the loss is, which is why the mandated cell and the expensive cell have to be funded as two decisions rather than one.

Two honest limits on the argument. The evidence that friction works is mandate-based rather than outcome-measured — no source behind this issue priced customer abandonment or conversion cost against loss prevented, so a finance director can fairly call D4 an unpriced control. And a control that blocks a customer's payment can itself bring that decision within scope of Vietnam's AI Law 134/2025 and the personal-data regime — conditionally, depending on whether the control is an in-scope AI system and whether the block is automated decision-making with significant effect, so a deterministic rules engine may sit outside it entirely. The point is that the cost is not zero, which the fraud literature routinely assumes. No enforcement precedent on payment interdiction exists yet under either instrument.

Before you fund more detection, establish which half of the grid is missing.

One page for your next payments committee, and do not accept a blended loss number to start it: split last quarter's losses into authorised and unauthorised, because a blended figure hides the whole problem and flatters the control estate. That split is your baseline. Then mark the cells you actually operate — not the ones you have bought a product for, but the ones where something fires, somebody owns the outcome, and the audit trail would satisfy a supervisor. If the loss sits in one column and your controls sit in the other, you have an investment case, and it is not a faster model.

Latest Video

This week's briefing — Video 16 The control worked The money still left — walks the two axes and all six cells through on screen in 8 minutes 16 seconds, including the illustrative payment that has nothing wrong in it: the call, the persuasion, the biometric passed, and the transfer approved on a trusted device.

Watch: youtu.be/RzMZD4zwnvQ

The full deep-dive — the two-failure reframe, all six cells with their owners and residual risks, the UK before-and-after in full, the receiving-side concentration on both readings, and every figure with the market, period and definition it was measured on — is in The Payments Defense Grid: The Control Worked. The Money Still Left..

The free five-page playbook in the Frameworks library turns it into a worksheet for your own estate — the six-cell grid with owners, the decision table for which cell to fund next, and the five-step coverage audit.

Reply and tell me which cell of that grid has a named owner in your bank — not which product you have bought, which cell has somebody accountable for the outcome when the payment is clean and the money is gone anyway. In most banks the whole right-hand column comes back blank. I read every response. Forward this to a banking executive about to sign off another detection uplift.

Was this forwarded to you? Subscribe to The AI Architect Letter — free, every Saturday.

Minh Tran · AI Business Architect · LinkedIn · Workshops & advisory: aibusinessarchitect.ai