Free executive playbook.

Your authentication worked. The biometric matched, the device was trusted, and the customer approved the payment themselves. The money is gone, irreversibly, in under a second. Nothing failed — and that is the problem. Instant settlement removed the pause every legacy fraud control quietly depended on, banks reasonably responded by moving the same controls earlier and making them faster, and the losses kept climbing, because the dominant loss on an instant rail is made by a genuine customer on a trusted device. Money leaves an account wrongly in only two ways: an impostor acts without consent, where the deception targets the bank's systems and the bank's systems can find it; or the owner acts, deceived, where the deception targets the customer before the payment exists and the transaction itself is clean. The consequence most banks have not priced is that no amount of investment in impostor controls produces any coverage of the second kind at all — you can double the budget and buy nothing. The Payments Defense Grid puts the two axes that matter on one page: what a control can see, and what it can still do by the moment it fires. Six cells, each with an accountable owner and a stated residual risk. It is the author's proposed decomposition, not a ratified standard.

What's inside

  • The reframe — impostor vs. deceived owner, compared across where the deception lands, what the bank can actually observe, what detects it, the primary mitigation, and the residual risk each column's controls leave against the other. The short version, spelled out in the table: against a deceived owner the bank can observe nothing — device, credentials, session and intent are all genuine, and no transaction-level signal exists to detect it · plus the worked walkthrough of a payment with nothing wrong in it: the call, the persuasion over an hour or several weeks, the customer's own phone at an ordinary hour, the biometric passed, the confirmation screen read, the transfer approved, settled in under a second and not coming back. ILLUSTRATIVE throughout — no bank, no figure, no loss, and none is needed, because the shape is the point: the checkpoint fired, the gate opened, and the audit trail records clean evidence of a correct instruction
  • The grid — all six cells named, with an owner and a residual risk each. Down one side what a control can see; across the other what it can still do — prevent before value moves, interrupt at the moment of intent, absorb after finality. D1 payer & device trust (biometrics, presentation-attack detection, device binding, malware and remote-access detection — CISO and Head of Retail) · D2 payee intelligence (payee-name verification, destination risk and mule scoring, account age, first-time-payee status — Head of Payments and Fraud Risk) · D3 the in-flight decision (scoring inside the authorisation path, within the scheme's latency budget, with a defined fail-open or fail-closed behaviour — CRO and Head of Fraud Risk) · D4 customer-in-the-loop friction (specific warnings, interdiction questions, holds and cooling-off on new payees and new device bindings — Fraud Risk and Customer Experience, jointly) · D5 post-finality recovery (cancellation, freezing, tracing — Payments Ops and Disputes, and on an instant rail largely a formality, because funds cascade onward within minutes) · D6 network signal & shared liability (cross-institution mule intelligence and an allocation of loss that gives the receiving bank a reason to act — CCO and General Counsel) · plus the two disciplines that stop the grid becoming decoration: it is not a maturity ladder, nobody progresses from D1 to D6 because the columns are different problems rather than stages; and owning a product is not operating a cell, so mark only the cells where something fires and somebody is accountable for the outcome — most control inventories shrink considerably under that test
  • What moved when liability moved — the UK before-and-after table: mandatory reimbursement effective 7 October 2024, the cost split fifty-fifty between sending and receiving firm, reimbursement by value rising from sixty percent (all APP scam losses at the fourteen largest banking groups, pre-mandate) to eighty-eight percent of in-scope losses and one hundred and seventy-three million pounds, roughly two hundred and sixty-nine thousand claims reported of which one hundred and eighty-eight thousand were in scope — against detection technology that did not change. Coverage followed liability, not capability, though the two percentages sit on perimeters the mandate itself redefined, so the direction is the finding rather than the delta · and the receiving-side concentration on both readings, because the two are not interchangeable: non-directed providers took thirty-four percent of total fraud value while handling nineteen percent of consumer Faster Payments by value, and forty-eight percent of fraudulent transactions on ten percent by volume, with the regulator separately reporting their fraud rate by value as forty-one times the largest groups — its own measure, not the ratio of those shares. Criminals were not defeating the biggest banks' models; they were shopping for the weakest place to land the money · plus the scam-mix trap for anyone measuring fraud performance by case count: purchase scams are seventy-three percent of UK cases but twenty-seven percent of value, investment scams four percent of cases and twenty-five percent of value
  • The honest caution, stated as a caution rather than buried — 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 not distribute mule accounts the same way: the problem may concentrate inside institutions rather than between them, which changes the control from shared intelligence to internal account-cluster monitoring without changing the conclusion that the leverage sits at the destination. It is a genuinely open question and the playbook says so, because the alternative is borrowing another market's structure instead of answering it with your own data
  • The coverage audit and the decision that follows — five steps with why each one is where it is (split the loss authorised vs. unauthorised, because a blended number hides the problem and flatters the estate · mark the cells you operate, not the cells you bought a product for · compare the two · fund mandate and loss as separate decisions, because the mandated cell and the expensive cell are not the same cell · instrument the inbound ledger, because under a fifty-fifty split mule-hosting is a direct cost centre) · the decision table for which cell to fund next given what your losses actually look like, including the row most likely to describe your bank — takeover and SIM-swap losses point at D1 then D3 and explicitly not friction, scam losses point at D2 and D4 and not a faster in-flight model, a predominantly receiving institution points at D2 and D6, an arriving reimbursement duty points at D6 first because it re-prices the rest, and a bank that cannot split authorised from unauthorised loss should buy nothing yet and instrument the split · and the one-afternoon diagnostic: put the grid in front of the COO, the Head of Payments and the CRO, ask what fires here, who owns the outcome, and what would we show a supervisor — any cell with three blanks is not a control

Five pages · free with a member sign-in · PDF.

Two honest limits are carried in the downloadable edition as well as here, because they bound what the grid can claim. The evidence that friction works is mandate-based rather than outcome-measured — no source behind this playbook priced customer abandonment or conversion cost against loss prevented, so D4 is fairly called an unpriced control, and raising it to high confidence needs an independent measurement of both sides of that trade. 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 altogether. The point is that the exposure is not zero, which the fraud literature routinely assumes it is.

Four notes on the downloadable edition, in the interest of the same measurement discipline this playbook argues for. The first three were found by checking the PDF against the primary sources on 2026-09-21 and the fourth on 2026-09-26, and the on-site pages carry the corrected and current reading.

The page-4 reimbursement row. It sets a "reimbursement by volume" figure of seventy-three percent before the requirement against "eighty-two percent of claims closed within five business days" after it, in one row. Those are two different measures on two different denominators — a reimbursement rate and a claims-handling speed — so they are not a before-and-after pair, and the PSR publishes no comparable post-mandate by-volume rate in that report to complete the row with. The reimbursement movement that belongs beside the sixty percent is the eighty-eight percent by value of in-scope losses, stated correctly one row above.

The sixty-to-eighty-eight comparison itself, on page 1 and page 4, needs a perimeter caveat the PDF does not carry: the sixty percent is all APP scam losses at the fourteen largest banking groups before 7 October 2024, while the eighty-eight percent covers losses in scope of the reimbursement requirement, a category the mandate created. Both figures are the regulator's own, on periods that do not overlap. The direction is solid; the delta is not like-for-like.

The page-1 "41×" stat card. The figure is the regulator's own and is correctly attributed, but as the card's headline it invites the reader to derive it from the thirty-four-percent and nineteen-percent shares printed beside it. It is not that ratio — that one is nearer twofold — but a separately reported rate on a different basis.

The reimbursement absolutes have since been superseded. Both editions carry eighty-eight percent of in-scope losses on one hundred and seventy-three million pounds, with roughly two hundred and sixty-nine thousand claims reported and one hundred and eighty-eight thousand in scope. Those were the regulator's twelve-month figures. The PSR's APP scams reimbursement dashboard for Q1 2026, updated 30 July 2026, now reports the same eighty-eight percent across the requirement's first eighteen months — 7 October 2024 to 31 March 2026 — on three hundred and sixteen million pounds, with around four hundred and thirty-eight thousand claims reported and three hundred and one thousand five hundred in scope. The rate is unchanged; the absolutes and the counts are larger because the window is longer, not because performance moved. The insight page carries the current set.

One figure now differs between the two editions — the reimbursement absolutes in the note above, where the regulator has published a longer window since the PDF was cut. Everything else that differs is framing, in a piece about framing, so it is disclosed rather than smoothed over, and the corrected render is owed.

Watch: youtu.be/RzMZD4zwnvQ — Video 16 The control worked The money still left, 8 minutes 16 seconds.

Want the weekly version? Subscribe to The AI Architect Letter · Read the full analysis · Read the latest insights.