AI and Fraud in Payments
Summary: AI is simultaneously the biggest new fraud threat and the most effective fraud defense in payments — fraudsters use it to scale attacks while businesses use it to detect them.
Sources: raw/articles/stripe-state-of-ai-fraud-2025.md, raw/articles/avant-va-launch-2026-05-10.md
Last updated: 2026-05-17
The Dual Dynamic
AI is a tool for both attackers and defenders. The rise of generative AI has coincided directly with rising fraud attempts:
- Businesses experiencing fraud: 65% in 2022 → 79% in 2024
- 30% of business leaders say AI is making merchant fraud worse (sophistication or volume)
- 47% of businesses use AI to detect and prevent fraud — the #1 use of AI in payments
AI as an Attack Vector
Card Testing Attacks
- Fraudsters use AI to gather vast stolen card credential datasets, then run thousands of test transactions per business per day to validate cards
- 1 in 4 large-scale card testing attacks blocked by Stripe involve 1M+ transactions against a single business
Fake Identities Bypassing KYC
- Generative AI creates fake identities + falsified documents convincing enough to pass KYC verification
- Fraudsters pair fake identities with AI-generated business websites to expand schemes (fake marketplace sellers, license resellers)
- Stripe’s Radar for Platforms blocked 500,000 fake account creation attempts (Jan–May 2025)
AI as a Defense
Who Is Adopting AI for Fraud Defense
| Industry | Adoption Rate |
|---|---|
| Insurance | 62% |
| SaaS | 58% |
| Travel & hospitality | 54% |
| Platforms / marketplaces | 76–78% |
| All businesses (average) | 47% |
Platforms and marketplaces lead because they face both transaction fraud and merchant fraud simultaneously, and traditional manual approaches can’t scale.
Real-World Results
FreshBooks + Stripe Radar
- Linked merchant risk and transaction monitoring (previously separate, periodic processes)
- Dynamic, risk-based decisioning with real-time portfolio health view
- Blocked 300 fraudulent accounts from onboarding in 3 months
DoorDash + Stripe Radar
- Separate in-house fraud models per payment workflow (checkout, gift cards, etc.)
- Uses behavioral + transactional signals + historical user context
- Added Stripe Radar risk scores to incorporate network-scale data
- 10% decrease in chargeback costs
Avant Virtual Agent
- Deployed VA (built with Replicant) across inbound Avant Card customer service, January 2026
- 62% of calls completed without agent transfer
- 4.6 / 5 customer rating on post-call surveys
- Every VA interaction produces consistent structured data — a feedback loop difficult to build at human scale
Stripe Payments Foundation Model
- Trained on tens of billions of transactions
- Detection rate for card testing attacks on large users: 59% → 97%
- Stripe Radar reduced dispute rates for users by 17% last year, even as industry-wide ecommerce fraud grew 15%
The Future
Three predictions from Stripe:
- Agentic commerce — universal wallets required for AI agent payments may create new account takeover vectors
- AI agents in fraud ops — detect and interpret patterns teams miss; eliminate manual reviews during investigations
- Identity verification evolves — as AI-generated fake documents improve, verification shifts toward biometrics and passkeys
What This Means for FDE
- Fraud is a perennial high-priority problem with clear ROI — catches translate directly to cost savings and revenue
- Enterprises running fraud on every transaction are paying the inference bill billions of times per day (see nvidia-finance)
- The PRAGMA result of +65% fraud recall confirms foundation models can dramatically outperform bespoke ML models (see pragma-revolut)
- There’s a wedge opportunity: enterprises want AI fraud tools but face InfoSec restrictions on data access (see ai-adoption-in-finance)