Payment fraud and APP scam detection

• Every outbound payment is scored in milliseconds by a fast, explainable model, alongside the rules you already run. Generative AI is never in the payment path.
• Only payments in the highest-risk band are held automatically, under the payment delay power and rules a person approved. The customer is told why and what happens next, the same day.
• Every other alert goes to an analyst before any hold. The payment isn't held while they look: the customer sees a scam warning in the app, tailored to the pattern that scored, and must confirm before it goes. The analyst decides whether anything more is needed: contacting the customer, holding a later payment, or trying a recall.
• Each alert, and each automatic hold, reaches an analyst with a short summary: the main reasons it scored high, the scam pattern it most resembles, and the evidence behind each point.
• The analyst gets a suggested set of questions for the customer conversation, drawn from approved scripts for that scam type, so calls are consistent and nobody is accused.
• Incoming payments are watched too. Accounts that look like money mules are flagged to a person before the money moves on.
• Reimbursement claims arrive with the evidence already gathered and the deadlines tracked. A person makes every decision on a payment, an account or a claim, and every step is logged.

• PSR (moving to the FCA)APP scams reimbursement requirement, live since 7 October 2024
  →

  Victims of APP scams over Faster Payments are reimbursed up to £85,000 per claim, split equally between the sending and receiving firms, with matching rules from the Bank of England for CHAPS. A firm may apply an excess of up to £100, but not to vulnerable customers. Claims can be made up to 13 months after the payment. Decisions are due within five business days, or 35 if the clock is stopped. The AI gathers the evidence and tracks the deadlines. A person decides every claim, and every finding of gross negligence, which is a high bar and doesn't apply to vulnerable customers.

• HM Treasury and the FCAPayment Services Regulations 2017, as amended in 2024 (the payment delay power), and FCA guidance FG24/6
  →

  A firm may delay an outbound UK sterling push payment where it has reasonable grounds to suspect fraud or dishonesty, established by the end of the next business day. The delay must be no longer than necessary, and never beyond the fourth business day. The customer must be told about the delay, the reason and what they need to do, unless telling them would be unlawful. The firm is liable for interest and charges the delay causes, even if the payment is then sent. So only the highest-risk band is held automatically, under rules a person approved: the notice goes out automatically, and an analyst decides to release or refuse. Every other alert goes to an analyst before any hold, and only the analyst can apply one.

• PSRConfirmation of Payee (Specific Direction 17)
  →

  Name checking is now permanent: the PSR removed the direction's expiry date in October 2026. The CoP result (match, close match, no match) is a strong signal for the model, and the warning the customer saw is key evidence in any claim, so both are stored with the payment.

• FCAConsumer Duty (PRIN 2A)
  →

  There's no AI-specific rulebook. The FCA applies the Duty to AI like any other change. Firms must avoid foreseeable harm, which cuts both ways: scams that should have been caught, and genuine payments blocked without good reason. Outcomes, including hold rates and release times, are monitored by customer group, including vulnerable customers.

• ParliamentEconomic Crime and Corporate Transparency Act 2023
  →

  Section 188 (in force since 15 January 2024) lets regulated firms share customer information with each other to prevent economic crime, without breach of confidence claims. It doesn't remove UK GDPR duties or the tipping-off offence. The failure to prevent fraud offence (in force since 1 September 2025) covers fraud by staff or agents intended to benefit the firm, so claims handling needs reasonable procedures too: no misleading a customer to avoid paying a valid claim.

• ICOUK GDPR and automated decisions (as amended by the Data (Use and Access) Act 2025)
  →

  Only the temporary hold on the highest-risk band is automatic, because it has to happen in milliseconds, and it follows rules a person approved. Every other hold is an analyst's decision. Since 5 February 2026, decisions like this need safeguards rather than a ban, unless special category data is used: telling the customer, letting them respond and contest it, and a person who can step in. The delay notice and the analyst review provide exactly that. Keep special category data, such as health details from a vulnerability note, out of the scoring model. A DPIA comes before the pilot.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• A fraud strategy owner approves the hold rules and thresholds, including the narrow highest-risk band that is held automatically, and signs off every change. The model suggests; the approved rules decide what is held automatically, and analysts decide every other hold.
• An agreed pick-up time for alerts and an agreed release target, so alerts are reviewed quickly and held payments aren't left waiting to the end of the fourth business day by default.
• Approved call scripts per scam type, owned by fraud and compliance, with a weekly review of a sample of real calls.
• A named owner for the model, the typology library and each prompt, with change control and model risk review like any other model.
• A weekly look at scam types the model is missing, including where scams moved to: crypto, international payments and other routes outside the reimbursement rules.
• Claims QA on a sample of decisions each week, checking consistency, deadlines and how vulnerability was handled.

1. 3–4 weeks

   Discovery

   Map today's alert and claims process, measure the baseline, check the data the payments hub can send in time, and agree the scam types to start with.

2. Gate · must be true to passBaseline agreed, latency budget confirmed with payments engineering, DPIA started, model risk and compliance involved.

3. 8–10 weeks

   Pilot

   Shadow mode: the model scores every payment alongside the existing engine, summaries are produced for real alerts, and analysts compare them with their own work.

4. Gate · must be true to passBetter precision at the same or lower losses, no customer group worse off, summaries rated useful by analysts, and model risk sign-off.

5. 4–6 weeks

   Production

   The model joins the live alert rules for the first scam types, with automatic holds limited to the approved highest-risk band, summaries in the case tool, monitoring and a rollback plan.

6. Gate · must be true to passMonitoring, support model, rollback switch and analyst training in place.

7. Ongoing

   Scale

   Add the claims assistant, inbound mule detection, then card and account takeover fraud.

8. Gate · must be true to passEach new use passes the same shadow-mode gate.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   A language model is put in the payment path, and payments slow down or get inconsistent answers.

   How the design avoids it

   A small, fast scoring model with reason codes raises alerts, and only the approved highest-risk band is held automatically. Generative AI only explains and drafts, after the alert.

2. 2

   Goes wrong

   The automatic-hold band quietly widens as thresholds are tuned, and more customers are held without an analyst looking first.

   How the design avoids it

   Only the fraud strategy owner can change the band, every change is logged, and its size and precision are reported weekly.

3. 3

   Goes wrong

   The model learns last year's scams, and the criminals have already moved on.

   How the design avoids it

   A weekly typology review, fresh labels from claims, a regular retraining cycle, and drift checks owned by a named person.

4. 4

   Goes wrong

   Holds go up, losses go down, and good customers are delayed for days.

   How the design avoids it

   Measure good payments delayed alongside losses, agree a release target, and tune thresholds with the business, not just the data team.

5. 5

   Goes wrong

   A drafted message sounds like an accusation, or hints at a suspicious activity report.

   How the design avoids it

   Approved scripts as the only source, guardrails on every draft, and the analyst sending nothing they haven't read.

6. 6

   Goes wrong

   Claims are decided differently by different handlers, and the regulator notices.

   How the design avoids it

   The same evidence pack and checklist for every claim, a second review of every decline, and a weekly consistency sample.

1. December 2026 to May 2027
   The PSR's consultation on changes to APP reimbursement, following the independent evaluation published on 1 July 2026. It covers claims that run past 35 business days, guidance on the consumer standard of caution, civil disputes and me-to-me payments, and annual data on the platforms fraudsters use. A decision and revised directions are planned for May 2027, with up to six months to implement.

   Keep the claim evidence structured and the checklist configurable, so new guidance changes a rule, not the system. Start recording where each scam began: social media, a marketplace, a phone call.

2. Late 2026 to 2028
   The Financial Services and Markets Bill abolishes the PSR and moves its powers, including APP reimbursement, to the FCA. The Lords stages finished on 15 September 2026 and the Bill is now in the Commons. Commentators expect firms to feel the change from 2028.

   No redesign is needed. Map each control to the requirement, not the regulator, and watch for FCA consultations on a combined payments rulebook.

3. From 31 December 2026 (proposed)
   Confirmation of Payee is now permanent. The PSR also proposed directing firms that offer CoP voluntarily, as a new group, from 31 December 2026.

   Make sure the CoP result and the exact warning shown are stored with every payment, ready for scoring and for claims.

4. 2026 to 2029
   The Home Office Fraud Strategy 2026 to 2029 (9 March 2026) set up a public-private Online Crime Centre and plans to tackle barriers to information sharing. Banks and payment firms are pressing for tech and telecoms firms to share reimbursement costs. The government hasn't adopted that.

   Build one controlled interface for sharing fraud data, with a DPIA and clear lawful basis, so new sharing schemes plug in rather than being bolted on.

5. 2026 to 2027
   The ICO's code of practice on AI and automated decision-making, required by regulations in force since May 2026.

   Keep the hold notices, the analyst decisions and the reasons in a form you could show the ICO.