AI Foundation Models in Payments: History, Trends, and Future Impact

Artificial intelligence (AI) and machine learning (ML) have been pivotal in the payments industry for decades, primarily for combating fraud and streamlining transactions. The modern era of payment fraud analytics began around 1992-1993. In 1992, HNC Software introduced the FICO Falcon Fraud Manager, a neural-network-based system for real-time credit card transaction scoring. Concurrently, Visa pioneered the deployment of AI models for risk and fraud management in 1993. These early systems significantly improved fraud detection, with U.S. payment card fraud falling from about 0.18% of card volume in 1992 to around 0.05% in recent years. Such tools executed thousands of risk calculations in milliseconds, learning from trillions of transactions to enhance payment safety without adding friction.

While building payment systems, I've watched AI evolve from basic fraud rules to something far more sophisticated. What started as simple "if-then" logic has become systems that genuinely understand the context, intent, and patterns behind how money moves. This isn't just an incremental improvement—it's a fundamental shift in how we think about financial intelligence.

The rise of e-commerce in the 2000s saw companies like PayPal innovate with ML-driven techniques. PayPal's early fraud team developed methods like CAPTCHAs and an algorithm nicknamed "Igor" to identify abnormal behaviour, marking one of the first commercial big-data fraud applications. These efforts were vital, as unchecked fraud nearly bankrupted the company. By combining data-driven models with human review, PayPal reduced its fraud loss rate to approximately 0.32% of revenue, well below industry averages. Throughout the 2010s, card networks and banks heavily invested in AI. Visa reported spending over $3 billion on AI and data infrastructure in the last decade and by 2023, had hundreds of AI models powering over 100 products, including advanced deep learning systems for real-time fraud scoring and "stand-in" transaction approvals during outages. Mastercard also acquired AI platforms, such as Brighterion, to bolster network-wide fraud prevention. Traditionally, these AI models were specialized, each optimized for a single task like fraud detection or credit scoring, delivering steady improvements over the years. Prior to "foundation models," AI in payments relied on many task-specific models, large historical transaction datasets, and focused on risk mitigation, approval rate improvement, and personalization.

The Moment Things Changed

I remember the moment I realized we weren't just detecting fraud anymore—we were predicting user behavior. It was watching how modern AI models could distinguish between a legitimate late-night purchase and suspicious activity, not just by transaction amount or time, but by understanding the user's entire financial story.

Traditional fraud systems asked: "Does this transaction look suspicious?"

Foundation models ask: "Does this transaction make sense for this specific user, in this specific context, given everything we know about how people actually use money?"

That's a completely different question, and it requires a completely different kind of intelligence.

Rise of Foundation Models in Payments

Recently, the industry has begun transitioning from narrow models to foundation models purpose-built for payments data. A foundation model is a large-scale model, often based on transformer neural networks, trained on broad, unlabeled data, and adaptable to many tasks. In payments, this involves training on massive volumes of transaction data to learn underlying patterns. The objective is to capture subtle, high-dimensional signals from transactional behavior that specialized models might miss.

In May 2025, Stripe announced its Payments Foundation Model, described as the world's first AI foundation model for payments. This transformer-based model was trained on tens of billions of payment transactions, leveraging over $1 trillion in payments processed annually by Stripe. The model automatically learns "hundreds of subtle signals" about each payment, which previously required separate models or were too complex to detect. By analyzing patterns across diverse transaction attributes (card type, merchant, device, location, timing, historical behavior), the foundation model develops a holistic understanding of legitimate versus fraudulent activity. Stripe reported dramatic performance boosts, such as a 64% improvement in detecting "card testing" fraud overnight when the foundation model was deployed. This single model now underpins multiple Stripe services, replacing many siloed ML models.

Around late 2023, UK-based Featurespace introduced TallierLTM, described as a "Large Transaction Model" (LTM) – effectively a foundational generative AI model for payments. TallierLTM models genuine versus fraudulent behavior in a generalized way. Featurespace’s CTO stated, "What OpenAI did around language and words, we've created for the payments environment, modeling what genuine behaviors and transactions look like". The LTM uses generative AI to simulate normal transaction patterns, thereby spotting anomalies indicating scams or money laundering. In a pilot, this approach improved fraud value detection by up to 71%. Integrating the generative LTM into a UK payments network pilot identified an additional £40 million in fraud, boosting detection rates to 56%.

Foundation models differ from traditional predictive models by learning a broad representation of transaction data without being limited to one task. They operate in a self-supervised manner, learning from patterns in transaction sequences, metadata, and user behavior, resulting in a high-dimensional "knowledge" of typical transaction flows. This can be adapted for various purposes like fraud detection, risk scoring, and compliance. Stripe's model, for example, conceptually maps payments into a multi-dimensional feature space, where legitimate payments cluster in stable regions and fraudulent ones are outliers. This ability to automatically find structure in transaction data is their key strength.

Key Players and Pioneers

Several companies are advancing AI foundation models or analogous large-scale systems for payments:

• Stripe: This fintech platform's transformer-based Payments Foundation Model, trained on over 100 billion data points, significantly boosted fraud detection for large merchants. It underpins Radar fraud prevention, payment authorization optimization, and automated dispute handling, benefiting from Stripe's $1.4 trillion annual payment volume that includes diverse transaction types like e-commerce and crypto. Stripe claims its system captures signals that "specialized models can't".

• Featurespace: Known for its ARIC™ Risk Hub, this financial crime AI firm launched TallierLTM™, a Large Transaction Model, in late 2023. TallierLTM is generative, simulating realistic transaction patterns to improve the detection of anomalies. It improved fraud value detection by 71% in testing and utilizes collaborative data to learn genuine customer behavior, helping catch scams like authorized push payment fraud. Featurespace processes over 50 billion events a year, protecting more than 500 million consumers.

• PayPal: With a long history in AI, PayPal processes over $1.2 trillion in payment volume and 15+ billion transactions annually. They use machine learning and graph neural networks to link entities and find suspicious relationships in their two-sided network. This graph AI approach helps surface hidden links, like multiple accounts funding the same device. This dropped PayPal's transaction loss rate to 0.12% of volume in 2020. PayPal's AI also improved approval rates by up to 2.4% for some merchants by preventing false declines. The 2023 launch of its stablecoin, PYUSD, indicates potential future leverage of on-chain transaction data with AI.

• Visa: A pioneer since 1993, Visa has AI deeply embedded in its network. Visa Advanced Authorization (VAA) evaluates over 500 transaction attributes in milliseconds using deep learning, preventing an estimated $27 billion in fraud in 2022. "Smarter STIP" uses AI to approve legitimate transactions with over 95% accuracy if a bank's system fails. Visa is also active in blockchain analytics, analyzing stablecoin and crypto transaction trends.

• Mastercard: Similar to Visa, Mastercard has infused AI across its ecosystem, notably via its 2017 acquisition of AI software company Brighterion. Brighterion’s platform provides models for fraud detection, AML, and credit risk, used by large U.S. banks. Mastercard's Cyber & Intelligence division uses AI for scam and transaction laundering detection. The acquisition of crypto analytics firm CipherTrace in 2021 aims to augment AI-driven compliance for blockchain transactions.

• Other notable entities include Sardine, an AI risk platform; Feedzai, providing AI-driven risk scoring; Sift and Forter, applying large-scale ML for e-commerce fraud detection; and Chainalysis and TRM Labs, using ML for blockchain analytics. Cloud providers like AWS and Google also offer specialized AI for financial anomaly detection.

Stablecoins and AI: Transaction-Rich Data as Fuel

Stablecoins, pegged to fiat currencies, have seen explosive growth and offer unique opportunities for AI models due to their operation on public blockchains, creating open, transaction-rich datasets. In 2024, stablecoin on-chain transfer volume reportedly surpassed $27.6 trillion. This data can fuel foundation models.

• Training Data: Stablecoin networks generate millions of transaction records (addresses, timestamps, amounts, smart contract interactions), ideal for training AI to understand financial flows at scale. Patterns in stablecoin circulation can help distinguish normal liquidity from suspicious activity. A model trained on this data could anticipate market fluctuations or detect signs of money laundering.

• Improved Compliance and Fraud Monitoring: While posing new compliance challenges, blockchain transparency is an advantage. AI models can monitor stablecoin transactions in real-time. Firms like Chainalysis use AI to trace illicit funds. A foundation model could correlate on-chain behaviors across multiple dimensions and chains to identify masked transactions or rate transaction risk instantly. Issuers like Circle or PayPal could use AI for early warnings of abnormal activity. This provides an unprecedented radar for financial integrity.

• Better Financial Products and Inclusion: Analyzing stablecoin payment flows in emerging markets can help fintechs identify underserved regions. Stablecoin data could also train models for alternative credit scoring in cash-based economies. Foundation models could enable autonomous finance, like algorithms managing liquidity pools dynamically.

• Stablecoins Benefiting from AI: AI can audit stablecoin smart contracts for vulnerabilities or optimize reserve management. AI chatbots can improve user interface with stablecoin wallets. As stablecoins integrate into fintech (e.g., Stripe's stablecoin accounts), AI, such as Stripe's foundation model, can extend to monitor both fiat and crypto transactions seamlessly.

Opportunities Unlocked by AI Foundation Models in Fintech

Foundation models tailored to payments unlock numerous opportunities:

• Revolutionized Fraud and Scam Detection: Foundation models dramatically improve fraud detection accuracy and reduce false positives by holistically modeling legitimate behavior. As seen with Stripe and Featurespace, detection rates jumped significantly (e.g., +64% in card-testing fraud for Stripe; up to 71% fraud value detection for Featurespace ). This means fewer losses and fewer false declines, boosting revenue and customer satisfaction. Sophisticated scams, such as authorized push payment fraud, may be detected by AI, which can identify unusual user payment patterns.

• Real-time, Contextual Compliance and AML: Foundation models analyze transactions for compliance with greater context than traditional rule-based systems, which generate many false alerts. An AI model can ingest transaction details and unstructured data (news, geographic risk) to assign nuanced risk scores instantly, turning raw data into live probabilities for fraud or laundering risk. This enhances regulatory compliance and focuses human resources on true positives.

• Higher Authorization Rates and Superior UX: AI models can analyze why valid transactions are declined and predict when to retry or route differently, improving approval rates (PayPal saw 0.3-2.4% gains ). Foundation models can take this further by considering all contextual signals to maximize approval chances, meaning fewer declines for users and smoother checkouts. AI can also personalize security, dynamically escalating it only when a transaction appears genuinely risky, thereby enhancing the user experience without compromising security.

• New Product Innovation:

  • Automated Financial Insights: AI can analyze spending and provide tailored advice, like benchmarking business expenses against peers.

  • Conversational Finance: Generative AI enables natural interactions, with chatbots assisting users with queries about transactions or account reconciliation.

  • Dynamic Services: Fintechs can offer foundation-model-driven fraud detection as an API, democratizing advanced risk management.

  • Advanced Credit/Lending: Lenders can use foundation models to analyze a merchant's payment history for dynamic credit decisions, potentially offering more inclusive credit.

  • Optimized Cross-border/Treasury: AI can optimize global payment routing and currency conversions in real-time.

• Smarter Fraud Response: AI can automate chargeback disputes, as Stripe has done, by predicting which disputes are genuine and reducing friction for legitimate users by avoiding overly restrictive security measures.

Foundation models are enabling fintechs to simultaneously increase security and improve user experience, opening doors for real-time, intelligent financial services.

Challenges and Risks

Despite the potential, applying foundation models in payments involves significant challenges:

• Data Privacy and Security: Training models on sensitive transaction data raises concerns about PII leakage. Strict data governance and techniques, such as differential privacy, are necessary to prevent models from memorizing or revealing private details. Any use of external AI tools must comply with regulations like GDPR, and models should be in secure environments.

• Model Bias and Fairness: AI models can perpetuate or amplify biases present in historical financial data, potentially leading to unfair targeting in credit or fraud decisions. Careful monitoring, threshold adjustments, and human oversight are essential, as AI isn't neutral and can reinforce biases.

• Explainability and Regulatory Compliance: Foundation models are often considered "black boxes," yet financial regulations (e.g., the EU AI Act, GDPR's "right to explanation") require transparency. Companies explore techniques like SHAP values or LIME for reason codes, but it's an ongoing challenge to balance model power with interpretability.

• Model Governance and Oversight: Robust AI risk committees and continuous monitoring are vital. Liability questions arise if an AI model causes critical errors. Clear governance, regular audits, stress-testing, and fallback systems are necessary for these "not set-it-and-forget-it" systems.

• Adversarial Risks: Fraudsters may attempt to exploit AI models through adversarial attacks or use AI themselves to craft convincing scams (an AI vs. AI scenario ). Maintaining data integrity against poisoned data is also critical.

• Regulatory Uncertainty and Legal Risk: The evolving regulatory landscape means rapid AI deployment could outpace laws, leading to penalties for violating fair lending or privacy regulations. Many financial AI systems may be classified as "high-risk" under frameworks like the EU AI Act, requiring extensive documentation and oversight. Authorities like the US Federal Reserve have emphasized that AI/ML models need the same rigor as traditional ones.

• Operational and Technical Challenges: These models require massive computing resources for training and real-time serving (Stripe partnered with NVIDIA for this ). This might create a competitive divergence between large players and smaller fintechs. Integrating models into legacy systems and managing their lifecycle (continuous learning and updates) is complex and costly.

• Critiques of Hype vs. Reality: Experts caution that foundation models might be overkill for some tasks, and simpler models could suffice with less complexity. There is a risk that investments in "AI hype" may not yield proportional returns or create a false sense of security. Over-reliance on a single model could be a "monoculture" risk; a hybrid approach might be wiser.

The Convergence

Here's my contrarian take: the real value isn't in making traditional payments smarter. It enables AI systems to understand and operate programmable money.

Foundation models trained on both traditional payment data and blockchain transaction patterns will create financial intelligence that can:

• Automatically execute complex multi-step financial strategies

• Predict optimal timing for DeFi interactions

• Generate custom financial products based on individual behavior patterns

• Orchestrate cross-chain transactions with perfect risk management

The companies building this convergence today - AI-native payment processors, stablecoin issuers with sophisticated transaction analysis, and DeFi protocols with intelligent automation—are positioning themselves for the programmable money era.

Sources:

1. KDnuggets. (2014). Evolution of Fraud Analytics - An Inside Story. https://www.kdnuggets.com/2014/03/evolution-fraud-analytics-inside-story.html

2. Visa. (2023, September). 30 years of AI and counting. Visa Perspectives. https://corporate.visa.com/en/sites/visa-perspectives/innovation/thirty-years-of-ai-and-counting.html

3. Fin (Plaid). (2016). PayPal's history of fighting fraud. https://fin.plaid.com/articles/paypals-history-of-fighting-fraud/

4. Stripe Newsroom. (2025, May). Stripe accelerates the utility of AI and stablecoins with major launches. https://stripe.com/newsroom/news/sessions-2025

5. TechCrunch. (2025, May 7). Stripe unveils AI foundation model for payments, reveals 'deeper partnership' with Nvidia. https://techcrunch.com/2025/05/07/stripe-unveils-ai-foundation-model-for-payments-reveals-deeper-partnership-with-nvidia/

6. AI Street. Stripe Built a Payments LLM to Fight Fraud. https://www.ai-street.co/p/stripe-built-a-payments-Ilm-to-fight-fraud

7. PYMNTS. (2024, January). Putting 'Scam Dens' Out of Business Means Using AI to Fight AI. https://www.pymnts.com/news/security-and-risk/2024/putting-scam-dens-out-of-business-means-using-ai-to-fight-ai/

8. Visa. Generative AI in payments. https://corporate.visa.com/content/dam/VCOM/global/services/documents/vca-global-generative-ai.pdf

9. PYMNTS. (2025). Stripe Launches Stablecoin Accounts and AI Model for Payments. https://www.pymnts.com/digital-payments/2025/stripe-launches-stablecoin-accounts-and-ai-model-for-payments/

10. Lex Substack. AI: How Stripe, Block, and PayPal are using Model Context Protocol. https://lex.substack.com/p/ai-how-stripe-block-and-paypal-are

11. Parsers VC. FeatureSpace - Funding, Valuation, Investors, News. https://parsers.vc/startup/featurespace.com/

12. Emerj Artificial Intelligence Research. (2022). Artificial Intelligence at PayPal - Two Unique Use-Cases. https://emerj.com/artificial-intelligence-at-paypal/

13. World Economic Forum. (2025, March). Stablecoin surge: Reserve-backed cryptocurrencies are on the rise. https://www.weforum.org/stories/2025/03/stablecoins-cryptocurrency-on-rise-financial-systems/

14. Mastercard. (2023). Market-Ready AI Offers a Faster Path to Fraud Mitigation. https://b2b.mastercard.com/media/x3rbuvd3/american-banker-market-ready-ai-offers-faster-path-fraud-mitigation-whitepaper.pdf

15. Business Wire. (2024, March 4). G2 Risk Solutions and Mastercard Combine AI and Merchant Insights for Superior Transaction Laundering Detection. https://www.businesswire.com/news/home/20240304550784/en/G2-Risk-Solutions-and-Mastercard-Combine-Al-and-Merchant-Insights-for-Superior-Transaction-Laundering-Detection

16. PaymentsJournal. (2023). How AI can Help Manage Payments Risk in 2023. https://www.paymentsjournal.com/how-ai-can-help-manage-payments-risk-in-2023/

17. Finance Magnates. (2023). Mastercard and Network International Extend AI-Powered Fraud Protection. https://www.financemagnates.com/fintech/mastercard-and-network-international-extend-ai-powered-fraud-protection/

18. International Banker. (2023, February 10). Mastercard and Network International launch AI fraud-prevention technology solution. https://intlbm.com/2023/02/10/mastercard-and-network-international-launch-ai-fraud-prevention-technology-solution/

19. X. Brighterion, a Mastercard Company. https://x.com/brighterion?lang=en

20. Chainalysis. (2024). Stablecoins 101: Behind crypto's most popular asset. https://www.chainalysis.com/blog/stablecoins-most-popular-asset/

21. Dwayne Gefferie Substack. Programmable Finance: What the Convergence of AI and Stablecoins Means. https://dwaynegefferie.substack.com/p/programmable-finance-what-the-convergence

22. Rapyd. (2023). AI and Compliance in Payments: Addressing Regulatory Challenges. https://www.rapyd.net/blog/ai-and-compliance-in-payments/