In an authorised push payment scam the real customer logs in on their own device, passes strong authentication and sends the money themselves, so every control built to stop impostors waves the payment through. The only signals left are in how the session behaves and how the payment differs from the customer's normal pattern: a person being coached on the phone types, pauses and navigates differently from the same person paying a known bill. This guide covers transaction monitoring and behavioural biometrics as a pair, from the regulatory baseline in the EU strong customer authentication rules to the UK reimbursement regime that now makes banks pay for missed scams. It is written for fraud strategy, fraud analytics and digital banking security teams. It goes deeper than the Device Intelligence and Fingerprinting Guide, which covers the question "is this the right device", and does not repeat it.
Key takeaways
- Behavioural biometrics answer a different question from device intelligence. Device signals ask whether the device is expected; behavioural signals ask whether the person on the device is acting freely and like themselves, which is the only question that matters in a coached scam.
- Transaction monitoring is a legal baseline, not an option. EU Delegated Regulation 2018/389 Article 2 requires payment service providers to run transaction monitoring that considers compromised credential lists, transaction value, known fraud scenarios, malware signs and abnormal use of the access device.
- Reimbursement has changed the economics. Since 7 October 2024 UK payment firms must reimburse in-scope APP scam victims up to £85,000, split 50:50 between the sending and receiving firm, so a missed coached payment is now a direct loss to both banks.
- Purpose decides the privacy regime. The ICO treats biometric data as special category when it is processed to uniquely identify a person, and lists keystroke and gait recognition as behavioural biometric techniques, so a model that recognises a customer is regulated differently from one that only scores session risk.
- Detection is worthless without a good intervention. The model's output must route to the right friction: a hold, a scam-specific warning, a call-back by a trained agent, or a payment block, measured by how many scams are stopped against how many genuine payments are delayed.
Two control families and what each one sees
Transaction monitoring looks at the payment: amount, payee, payment rail, time, velocity and how the transaction compares with the customer's history and with known fraud scenarios. Behavioural biometrics look at the session: how the user types, swipes, holds the phone, moves the mouse and moves through the app. Each has a blind spot the other covers. A large first payment to a new payee is suspicious on its own but common for house deposits and car purchases; the same payment made after a long hesitation on the payee screen, with an active phone call and text pasted into the reference field, is a far stronger scam indicator.
| Dimension | Transaction monitoring | Behavioural biometrics | Device intelligence |
|---|---|---|---|
| Core question | Is this payment unusual for this customer or typical of a known scam? | Is the person acting like themselves, and acting freely? | Is this the expected, genuine device? |
| Main inputs | Amount, payee age, payee account risk, rail, velocity, time, location | Keystroke timing, touch pressure and swipe shape, device orientation, navigation path, hesitation, copy and paste | Fingerprint, emulator and root signals, app attestation, IP and network |
| Best at | Mule-bound payments, unusual values, rapid draining | Coached victims, remote access sessions, bots, account takeover by a different human | New device takeovers, emulators, device farms |
| Weak at | Genuine customers making rare large payments | Short sessions with little interaction data, first sessions with no profile | Victim on their own trusted device |
| Typical timing | At payment submission, real time | Throughout the session, before submission | At login and at sensitive actions |
The practical design is a single decision point at payment submission that receives all three scores plus the payee-side risk. See the Payment Fraud Guide for the wider payment fraud picture.
The regulatory baseline for transaction monitoring
For EU and EEA payment service providers the minimum is set by Commission Delegated Regulation (EU) 2018/389, the regulatory technical standards on strong customer authentication under PSD2. Three articles matter for monitoring design.
- Article 2, transaction monitoring mechanisms: providers must have mechanisms that detect unauthorised or fraudulent payments, and must take into account, at a minimum, lists of compromised or stolen authentication elements, the amount of each transaction, known fraud scenarios, signs of malware infection in any session of the authentication procedure, and, where the provider supplies the access device or software, a log of its use and any abnormal use.
- Article 18, transaction risk analysis exemption: a remote payment can skip strong customer authentication only if the provider's fraud rate for that payment type is at or below the Annex reference rate, the amount is under the exemption threshold, and real-time analysis has found no abnormal spending or behavioural pattern of the payer, no unusual device or software access information, no malware in the authentication session, no known fraud scenario, no abnormal payer location and no high-risk payee location.
- Article 21, monitoring of exemptions: providers using the exemptions must record and monitor, at least quarterly and separately for remote and non-remote transactions, the value of fraudulent transactions, total value and resulting fraud rate, average transaction value, and the number and share of transactions under each exemption, and make the results available to competent authorities and the EBA on request.
| Exemption threshold value (Annex, RTS 2018/389) | Reference fraud rate: remote card payments | Reference fraud rate: remote credit transfers |
|---|---|---|
| EUR 500 | 0.01% | 0.005% |
| EUR 250 | 0.06% | 0.01% |
| EUR 100 | 0.13% | 0.015% |
Article 18 explicitly names "abnormal spending or behavioural patterns of the payer", which is why behavioural analytics often enter a bank through the strong customer authentication programme rather than the scam programme. The key point for APP scams is that strong customer authentication succeeds in a coached scam, because the customer authenticates themselves. Article 2 monitoring and behavioural signals are therefore the controls that carry the load. See the Authorised Push Payment (APP) Fraud Guide for the scam typologies.
Why UK reimbursement changes the business case
The Payment Systems Regulator's reimbursement requirement applies to payments made on or after 7 October 2024 over Faster Payments and CHAPS between UK accounts. The PSR states that the maximum claim is £85,000, which it says covers over 99% of claims, that victims should report within 13 months of the payment, that firms may apply an optional excess of up to £100 (not to vulnerable consumers), that most victims should be reimbursed within five business days, and that sending and receiving firms split the cost 50:50.
- Sending firm: now funds half of every reimbursed scam it failed to stop, so the value of an intervention that stops a coached payment is the full payment amount, not a share of reputational cost.
- Receiving firm: funds the other half, which makes inbound mule detection a direct financial control. See the Money Mule Detection Guide.
- Cost of friction: reimbursement does not remove the cost of false positives. Delayed genuine payments generate complaints and abandonment, so the business case has to price both errors.
UK Finance reports more than £1.1 billion stolen through fraud in 2024 in its Annual Fraud Report 2025, and states that APP fraud fell in both value and case numbers. Use the report's category tables for your own baseline rather than headline totals.
Behavioural signals that indicate coaching or remote control
A scam victim is a genuine user under someone else's direction. The signals that matter are therefore deviations from the customer's own baseline and patterns typical of being directed, not of being a different person.
| Signal | What it can indicate | Caveat |
|---|---|---|
| Long hesitation or repeated back-and-forth on the payee or amount screen | Uncertainty, waiting for instructions | Also normal for first-time users and older customers; compare with the customer's own history |
| Typing cadence that changes mid-session, or segmented typing of an account number in dictated chunks | Details being read out on a call | Needs enough keystrokes to be reliable |
| Pasted payee details or payment reference | Details sent by a scammer by message or email | Many genuine invoices are pasted; weight by payee age |
| Active phone call during a banking session (mobile, where the platform exposes it) | Live coaching, impersonation of bank or police | Platform permissions and privacy limits vary; disclose it |
| Screen sharing or remote access software running, or input that looks injected rather than touched | Remote access scam, attacker driving the session (ATT&CK T1219 Remote Access Tools) | Some accessibility tools look similar |
| Unfamiliar navigation path, going straight to limits or new payee setup | A different human or a script | App redesigns reset navigation baselines |
| Raising a payment limit, then a large payment to a new payee in the same session | Classic safe-account and investment scam sequence | Strongest when combined with payee risk |
| Inhumanly regular timing or no touch variation | Bot or automation, credential stuffing follow-on | Covered better by bot defences |
Avoid treating age or disability as a proxy for vulnerability. The aim is to detect a change from the customer's own behaviour and context typical of coaching. See the Elder Financial Abuse and Scam Prevention for Banks and Care Providers Guide for the vulnerability angle.
Privacy, consent and fairness
Behavioural data is personal data, and in some designs it is special category biometric data. The ICO's biometric recognition guidance says biometric data is special category when processed for the purpose of uniquely identifying someone, that you process it from the moment you collect it for that purpose even if no match is found, and that singling a person out is enough without knowing their name. It lists keystroke, gait and gaze recognition among behavioural techniques.
- Identification design: a model that builds a per-customer behavioural template and confirms "this is Jane" is recognition, so plan for special category conditions and an explicit DPIA.
- Risk-scoring design: a model that scores session patterns typical of coaching or bots, without a per-person template, may fall outside special category, but it is still profiling. Document the reasoning and keep the DPIA either way.
- Accuracy and discrimination: the ICO guidance has dedicated sections on statistical accuracy, fairness and the risk of discrimination. Test error rates by device type, accessibility settings and customer age band before go-live.
- Transparency: say in the privacy notice that session behaviour is analysed for fraud and scam prevention, and say what signals (for example, an active call) are collected.
- Retention: keep raw interaction streams briefly; keep derived features and decisions as long as dispute and reimbursement evidence requires.
For the wider legal picture of biometric processing in the US see the Illinois BIPA and Biometric Privacy Guide.
Decisioning and interventions
Scores are only useful if they map to interventions that stop scams without teaching genuine customers to click through warnings.
| Risk band | Intervention | Notes |
|---|---|---|
| Low | Allow; standard confirmation | Under RTS Article 18 only if all conditions are met |
| Elevated, generic | Step-up authentication | Stops impostors, does nothing against a coached victim |
| Elevated, scam pattern | Dynamic scam warning specific to the selected payment purpose, with forced pause | Generic warnings are ignored; ask purpose questions and tailor the warning |
| High | Hold the payment and call the customer from a trained scam team, on a known number | Agents must ask open questions; scammers coach victims to lie to the bank |
| Very high or confirmed | Block, restrict new payees, flag the payee account to the receiving firm | Feed outcome back to models and consortium data |
Choosing build or buy: large banks with data science teams can build transaction models in-house and buy behavioural SDKs; smaller firms and fintechs usually buy both as a managed service and should insist on explainable reason codes, raw feature export for disputes, and per-segment error reporting. A firm with mostly web traffic gets less from mobile-only signals such as call detection, and should weight payee-side risk and transaction context more heavily. See the Fraud Prevention Platform Buyer's Guide.
How to implement
- Baseline: tag historical APP scam cases by typology (impersonation, purchase, investment, romance) and measure where in the journey each was visible. This tells you which signals would have helped.
- Instrument: deploy the behavioural SDK in web and mobile channels in collection-only mode, with the DPIA and privacy notice update done first.
- Shadow score: run behavioural and transaction scores in parallel with the existing rules for 6 to 12 weeks, recording what each would have decided, without customer impact.
- Calibrate: join shadow scores to confirmed scam and reimbursement outcomes, set thresholds per payment type and segment, and check error rates by device, accessibility settings and age band.
- Pilot interventions: turn on scam warnings first (low cost), then holds and call-backs for a small share of traffic or one product, with a control group.
- Integrate payee risk: add receiving account signals and inbound mule detection; share confirmed mule data through approved channels.
- Roll out and keep a rollback: keep the previous rule set deployable, with a feature flag per intervention so a threshold error can be reversed in minutes rather than via a release.
- Govern: report Article 21 monitoring data quarterly where it applies, review model drift monthly, and retrain after app redesigns.
Detection and measurement
- Scam detection rate: value of confirmed APP scams stopped before funds left, divided by stopped plus reimbursed scams.
- False positive ratio: genuine payments held or warned for each scam stopped, by risk band.
- Intervention effectiveness: share of customers who abandon after a scam warning, against share who proceed and later report a scam.
- Time to decision: latency added at payment submission, and median time to complete a call-back.
- Reimbursement cost: reimbursed value as sending firm and as receiving firm per month.
- Segment fairness: false positive rate by age band, device class and accessibility setting, against the overall rate.
Log every decision with model version, reason codes and the intervention shown, so that each reimbursement claim can be reviewed against what the bank saw at the time. See the Fraud Metrics and KPIs Guide.
Framework mapping
| Control area | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 | Other |
|---|---|---|---|
| Session and payment monitoring | DE.CM-03, DE.AE-02 | SI-4 System Monitoring, AU-6 Audit Record Review | RTS 2018/389 Art. 2 |
| Step-up at risky actions | PR.AA-03 | IA-11 Re-authentication | RTS 2018/389 Art. 18 |
| Risk assessment of the model | ID.RA | RA-3 Risk Assessment | DPIA (UK GDPR Art. 35) |
| Lawful processing of behavioural data | GV.OC | PT-2 Authority to Process PII | ICO biometric guidance |
| Remote access and voice coaching | DE.CM-09 | SI-4 | ATT&CK T1219, T1566.004 Spearphishing Voice |
Common failure patterns
- Relying on strong authentication against scams: the coached customer completes every factor. Step-up is the wrong response to a scam score.
- One generic warning for every payment: customers learn to dismiss it, and scammers tell victims in advance what the warning will say.
- Population baselines instead of personal ones: flagging all slow typists punishes older and disabled customers and misses fast, coached ones.
- No feedback loop: reimbursement claims that never reach the model team mean thresholds are tuned on stale data.
- Silent privacy assumptions: launching call detection or behavioural templates without a DPIA or privacy notice change, then switching it off after a complaint.
- Ignoring the receiving side: treating inbound mule accounts as somebody else's problem, now that the receiving firm pays half.
Practitioner checklist
- Map your transaction monitoring inputs to each of the five RTS Article 2 factors and close gaps.
- Confirm every Article 18 exemption decision checks behavioural, device, malware, location and fraud-scenario conditions.
- Decide whether your behavioural design is recognition or risk scoring, record why, and complete the DPIA.
- Run new models in shadow mode against reimbursement outcomes before any customer impact.
- Route scam scores to scam-specific interventions and trained call-back staff, not to step-up.
- Test error rates by age band, device class and accessibility settings before go-live.
- Keep decision logs with reason codes long enough to defend reimbursement disputes.
- Add receiving account risk and mule detection to the same decision point.
- Keep a feature-flagged rollback for each threshold and intervention.
Standards and references
- EUR-Lex: Commission Delegated Regulation (EU) 2018/389 on strong customer authentication
- Payment Systems Regulator: APP fraud reimbursement protections
- Payment Systems Regulator: APP scams
- ICO: Biometric data guidance, biometric recognition
- UK Finance: Annual Fraud Report 2025
- NIST: SP 800-53 Rev. 5
- NIST: Cybersecurity Framework 2.0
- MITRE ATT&CK: Remote Access Tools
Related NHI Mgmt Group resources: Device Intelligence and Fingerprinting Guide · Authorised Push Payment (APP) Fraud Guide · Money Mule Detection Guide · Account Takeover Prevention Guide