On-device coached-scam check inside the payment flow
A small model on the phone, inside the bank app, flags coached payments before they send. Neobank fraud teams pay.
Crowded · Singapore ships on-device Gemma scam detection
Written for: Small startup team
Next step
Call the head of fraud at a UK neobank such as Monzo or Starling, ask what share of APP scam losses began with a message the customer could screenshot, and ask whether an on-phone check with no data egress clears their data protection officer.
Why now
What changed since the earlier SDK card: a government deployment proves the no-egress design that bank privacy officers asked for, and phone-sized models can now read screenshots and documents. 89% of card issuers plan to adopt or expand AI fraud tools. The honest limit is that a bank app can't read WhatsApp, so signals must come from inside the app: payee details, the stated purpose, screen-sharing, and screenshots the user chooses to share.
What became possible
Google DeepMind and GovTech Singapore are using lightweight on-device Gemma models to catch conversational scams in private chats with no data leaving the phone. Liquid's LFM2.5-24B-A2B is built for fast, private on-device agents. MiniCPM-V 4.7 runs a 35B-A3B vision model in about 3GB of phone memory and reads real-world documents.
Three ways in
- Purpose interview
On a high-risk transfer, a local model asks two short questions and scores the answers for signs of coaching. Only the score leaves the device.
- Customer
- Head of fraud at a neobank
- Model
- Per-active-user annual license
- Screenshot check
The user shares a screenshot of the 'investment' or 'tax office' message, and an on-device vision model classifies it before the payment goes ahead.
- Customer
- Neobank fraud and customer-protection teams
- Model
- Bundled add-on
- Validation pack
Test sets, false-positive rates and drift reports for the bank's model risk team.
- Customer
- Model risk management
- Model
- Annual fee
Competition
The detection signals are commoditised. Active calls, screen sharing and remote-access tools are offered by Fingerprint, Darwinium, OneSpan, ASEE, ThreatFabric and Cleafy.
Full competition notes (12 competitors)
The detection signals are commoditised. Active calls, screen sharing and remote-access tools are offered by Fingerprint, Darwinium, OneSpan, ASEE, ThreatFabric and Cleafy. BioCatch owns the behavioural 'coached customer' story at UK banks, and Google ships a free OS-level warning plus 30-second delay on Android across most major UK banks. A standalone risk-score SDK would struggle. Narrow openings: (1) cheap, quick-to-integrate packaging for small UK/EU EMIs and neobanks that BioCatch and ThreatMetrix price out; (2) owning the response layer rather than detection: reason codes linked to dynamic safety questions for each scam type, and audit logs that support decisions under the UK reimbursement rules (from October 2024), with detection signals bought in; (3) coverage Google's feature lacks: iOS, coaching over WhatsApp/Teams, and scammers spoofing saved contacts, though Darwinium and Fingerprint are moving into these too.
- Google Android in-call scam protection
Built into Android. Fires when a banking app is opened during a call with an unknown number while the screen is shared. Offers one-tap end call/stop sharing; if the user doesn't act, forces a 30-second pause. UK pilot with Monzo, NatWest and Revolut, now at most major UK banks, with pilots in Brazil, India and the US (including Cash App).
- BioCatch
Incumbent for coached-scam and vishing detection. Behavioural biometrics such as hesitation and typing, plus active-call detection. Works with more than a dozen UK banks covering over 110M customers; has a Scams360 product.
- Fingerprint Active Call Detection
Launched August 2026. Yes/no signal for cellular and internet calls on iOS and Android, no extra app permissions. Makes the basic signal cheap.
- Darwinium
May 2026 SDK update detects live Teams/Meet/Slack calls and tells a TeamViewer session apart from harmless TV casting.
- ThreatFabric Fraud Risk Suite
Mobile SDK with remote-access-tool detection and vishing/APP fraud behavioural checks feeding a fraud risk score; large EU bank footprint.
- Cleafy
Session risk scoring that detects remote-access fraud and voice-phishing transfers; raised €12M.
- OneSpan Mobile SDK
Android remote-access detection (port scan, permissions, screen sharing, accessibility services) with probability scores.
- ASEE Ongoing Call Detection
iOS/Android SDK feature that detects a call in progress during login or payment and can block one-time-code generation during a call.
- Outseer Fraud Manager
Added active-call detection for instant payments and APP fraud in April 2024.
- LexisNexis ThreatMetrix
Active Phone Call Detection feature; vendor case study claims a 40% fraud reduction for one bank.
- XConnect APP Fraud Check / GSMA Scam Signal
Uses mobile network data to spot a live call during a transaction, no app SDK needed; Vodafone/UK bank pilot.
- Sardine
Remote-access scam detection via browser and mobile instrumentation, plus the Sardine Shield product for fintechs.
Checked 2026-10-07
Risks
- Heavy competition from BioCatch, Feedzai and in-house teams
- App size, battery use and older Android phones
- Only a score leaves the device, which limits the bank's audit trail
- Friction at payment time hurts conversion
Best on market pull and team fit; open space is the weak point (4).
tech 6 · pull 7 · space 4 · fit 7 · build 6
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detectionGoogle DeepMind
- Run fast, private agents on-device with LFM2.5-24B-A2BHugging Face Blog
- LFM2.5-24B-A2B: Built for fast on-device agentsHacker News
- A 35B multimodal model running on a phone with ~3GB memory? MiniCPM-V 4.7 demo is pretty wildr/LocalLLaMA
- Nearly 9 in 10 Card Issuers Turn to AI to Stop Fraud Without Blocking SalesPYMNTS
What happened next
- +7d · 14 Oct 2026 nothing yet. No product launches, funding, acquisitions, or significant press coverage related to coached-scam detection SDKs for neobanks were found after October 7, 2026. Relevant activity in this space predated the flag by several weeks to months.
Status history
- 2026-10-07 Flagged as buildable