
Two fraud patterns dominate wallets: account takeover (ATO), where a criminal gets into someone else's account, and authorised push payment (APP) scams, where the genuine user is manipulated into sending money themselves. The first is an identity problem. The second is harder, because every credential check passes. The right person, on the right device, with the right passkey, is sending money to the wrong place.
That is why fraud cannot be a phase-two item in eWallet app development. In our experience the first scam wave tends to arrive within weeks of launch. This post outlines the pipeline we design for wallets: what to capture, where to score, how to respond and how to avoid freezing honest customers on payday.
The shape of the pipeline
client signals -> event stream -> feature service -> scoring (rules + model)
|
allow / step-up / hold / block
|
review queue -> labels -> retraining
Scoring runs synchronously on a small set of high-risk events and asynchronously on everything else.
Where to score
Not every tap needs a model. Focus real-time scoring on the moments where money or identity changes hands:
-
Login and new-device login. The core ATO checkpoint.
-
Payee creation. New beneficiaries are where APP scams start.
-
Top-up. Stolen card testing and mule account funding.
-
Transfer and payout. The last chance before money leaves.
-
Profile changes. New phone number or email often precedes takeover.
Signals worth capturing
Device and session
-
Device attestation results from Play Integrity and App Attest.
-
New device, reinstall or emulator indicators.
-
Session age and authentication strength (passkey, biometric, fallback).
Behavioural
-
Typing cadence and hesitation on amount and payee fields.
-
Navigation path: did the user go straight to transfer, or browse first?
-
Time between payee creation and first transfer.
-
Copy-paste of account details, which can indicate instructions from someone else.
-
Activity during an ongoing phone call, where the platform permits detecting it.
Account history
-
Typical transfer amounts and times for this user.
-
Payee novelty compared with past behaviour.
-
Recent profile changes or failed logins.
Network
-
Whether the receiving account has received from many unrelated senders recently, a mule pattern.
-
Links between accounts through shared devices or identity attributes.

Rules first, then models
On day one you have no labelled fraud cases from your own users, so you cannot train a meaningful model. Start with transparent rules tuned conservatively, and a human review queue. Every reviewed case becomes a label. Once you have enough labelled cases to train and evaluate properly, introduce models alongside the rules rather than replacing them.
def score(event, features):
reasons = []
if features.new_device and features.minutes_since_login < 5 and event.type == "payee_create":
reasons.append("new_device_new_payee")
if features.payee_age_minutes < 30 and event.amount > features.p95_amount:
reasons.append("fresh_payee_large_amount")
model_risk = model.predict(features) if model.ready else 0.0
return decide(reasons, model_risk)
decide maps rule hits and model risk to an action. Keep thresholds in configuration so the risk team can tune them without a release.
Responding without punishing honest users
Blocking is a blunt instrument. A graded response protects users without driving them away:
-
Allow. Low risk, no friction.
-
Step up. Ask for a passkey or biometric confirmation on a risky action. Good for ATO.
-
Interrupt with context. For likely APP scams, show a specific warning about the scam pattern and add a short delay to the first transfer to a new payee. Scam victims often need a moment outside the pressure of a call.
-
Hold for review. Funds held, not lost, while an analyst checks.
-
Block. Reserved for high-confidence cases.
The step-up path depends on strong identity foundations. Passkeys and device-bound keys beat SMS codes, because SIM-swap remains a leading takeover route.
The human review queue
Analysts make the final call on held transactions. AI helps them move faster rather than deciding for them. A language model can summarise the case: what the user normally does, what changed, which signals fired and related accounts. The analyst confirms or overrides, and that decision becomes a label. If you are building this kind of analyst copilot, strict tool permissions and audit logs are non-negotiable; that is the core of how we approach LLM integration.
Governance
The rule we apply on every wallet: models score and recommend; they do not move or release money on their own. Decisions are logged with the signals and model version that informed them, so you can explain any outcome to a partner bank, a regulator or the customer.
Measuring it
Track fraud losses, false-positive rate on legitimate transfers, step-up completion rate and the time held funds wait for review. A fraud stack that stops scams but freezes payday transfers is not working.
This pipeline is one part of a much bigger system. TechCirkle's full guide to eWallet app development covers the ledger, compliance and cost, and our AI development services team builds fraud models and evaluation harnesses for wallet operators.
Frequently Asked Questions
What is the difference between account takeover and an APP scam?
Account takeover is when a criminal gains access to someone else's account. An APP scam is when the genuine user is manipulated into authorising a payment, so credential checks pass and behavioural signals matter more.
Can rules alone stop wallet fraud?
Rules are the right start, but they struggle with social-engineering scams. Behavioural models, trained on your own labelled cases, catch patterns rules miss.
When should a wallet introduce fraud models?
Once you have enough labelled cases from your own users to train and evaluate a model properly. Before that, use conservative rules and human review.
How do you avoid blocking legitimate customers?
Use graded responses such as step-up authentication, contextual warnings and short delays on new payees, and reserve outright blocks for high-confidence cases.
Why are new payees a key checkpoint?
Most APP scams begin by adding a new beneficiary and sending money quickly. Scoring payee creation and the first transfer to that payee catches many scams early.
Should AI decide whether to release held funds?
No. AI can summarise cases and recommend, but an analyst should decide, with the decision logged and used as a training label.