Why Pix Made AI Fraud Prevention an Infrastructure Requirement

Brazil’s Pix processed 63.4 billion transactions worth $4.6 trillion in 2024 — 53% year-over-year growth, according to the Banco Central do Brasil. On a single day in June 2025, 276.7 million transactions were processed. The system is now the default payment method for 75% of the Brazilian population, surpassing cash, debit cards, and bank transfers as the primary way money moves across the country. It is, by almost any measure, the most successful instant payment infrastructure in the world.

It has also created a fraud problem with no precedent in any other banking market.

Three pressures are converging on Latin American banks as a direct consequence of Pix’s success and the financial inclusion transformation it represents. The first is a fraud surface that grows with every transaction and that is architecturally incompatible with rule-based or batch-scoring defences. The second is a credit opportunity that Pix’s inclusion success has created but that traditional credit models cannot serve. The third is a regulatory framework — now in active enforcement — that requires AI capabilities that were discretionary investments eighteen months ago.

The Pix Paradox: infrastructure success created the fraud crisis

The mechanism is straightforward. Pix operates on irrevocable, instant settlement. When a payment is initiated, the funds move in seconds and cannot be recovered from the receiving institution in the ordinary course. A bank that scores fraud decisions on a batch basis — assessing transaction patterns periodically rather than at the point of initiation — is making a prevention decision after the funds have already moved. Detection at that point produces an alert, not a prevention. Recovery depends on whether funds remain in the first receiving account, which in the majority of fraud cases they do not: Brazil’s original MED refund mechanism had an 89% denial rate because fraudsters routinely moved funds through multiple intermediate accounts within minutes of receiving them, according to analysis by Silverguard.

The consequence is visible in the chargeback and fraud loss data. Brazil’s chargeback rate runs at 3.48 to 3.55% — compared to 0.47% in the US and 0.51% in the UK. According to the Global Anti-Scam Alliance’s State of Scams in Brazil 2024 report, total estimated fraud losses reached R$297.7 billion — approximately $54 billion — equivalent to 2.5% of Brazil’s GDP. The Banco Central do Brasil’s data specifically on Pix fraud put losses at R$6.5 billion in 2025. Fraud cases in Brazil quadrupled between 2018 and 2023, from 426,799 to 1,965,353 reported cases, according to the Brazilian Public Security Forum — fraud is now Brazil’s most common property crime, having overtaken physical robbery. The speed of the fraud cycle compounds the problem: 61% of scams in Brazil are completed within 24 hours of initial contact, leaving banks with a window measured in minutes, not days, to prevent losses that cannot be reversed.

The Banco Central do Brasil’s response was MED 2.0, the enhanced Special Refund Mechanism that became mandatory for all Pix participants on February 2, 2026. Where the original MED could only block funds in the first receiving account, MED 2.0 enables tracing and automatic blocking across chains of intermediate accounts — up to five layers deep — with enforcement penalties for non-compliant institutions that began in May 2026. The BCB estimates the upgrade could reduce successful Pix scams by up to 40%. Implementing MED 2.0 effectively is an AI problem, not a rules problem: mapping real-time multi-hop transaction chains across thousands of accounts, identifying which chains represent fraudulent flows, and triggering automatic blocks within the 11-day settlement window requires pattern recognition at a scale and speed that static rules cannot provide.

The irrevocable settlement model, the 3-to-5 times higher chargeback rate than Western markets, and now the mandatory MED 2.0 infrastructure together create an unambiguous architecture requirement. Inline AI scoring before payment initiation is not an enhancement to the fraud prevention system. In the Brazilian Pix context, it is the fraud prevention system.

The financial inclusion opportunity: 48% informality means traditional scoring excludes half the workforce

Pix’s success has been inseparable from a dramatic expansion of financial access. Adult banking penetration in Brazil reached 84% in 2024, up from less than 50% in 2017. Nubank alone has over 105 million users across Latin America. Mercado Pago serves 50 million, PicPay 36 million. The scale of digital financial inclusion achieved in Brazil over seven years is without parallel globally.

This creates an immediate and commercially significant problem for credit. A near-majority of those newly included customers entered the financial system without formal credit history, because they come from an informal economy. According to the International Labour Organization, labour informality across Latin America stood at 48% in mid-2023 — meaning roughly half of all workers in the region have income that does not appear in formal credit bureau systems. Traditional credit scoring models built on bureau data, employment records, and payroll information cannot assess the creditworthiness of these customers. They are not uncreditworthy. They are invisible to the models that would determine whether they are.

The implication for credit decisioning is direct. AI models trained on alternative signals — Pix transaction regularity, mobile payment patterns, utility payment history, telco data — can make creditworthy assessments on customers that bureau-based models exclude entirely. Nubank’s growth demonstrates that this population can be served profitably at scale. The opportunity is not to provide credit as a social service but to capture a commercially viable market that incumbents, relying on traditional scoring infrastructure, are structurally unable to reach. In a region where labour informality is structural and persistent, this is not a transitional opportunity. It is a permanent feature of the credit landscape.

Regulatory momentum: MED 2.0 is in force, and the Pix model is spreading

Brazil’s regulatory requirements are no longer approaching — they are active. MED 2.0 mandatory compliance from February 2026 and enforcement penalties from May 2026 mean that Pix-participating institutions without real-time multi-hop fraud tracing infrastructure are currently facing supervisory consequences. BCB Resolution 589 also required all Pix participants to offer self-service fraud dispute functionality within their apps by October 2025. These are current obligations, not future ones.

Brazil’s General Data Protection Law, LGPD — structurally equivalent to GDPR — adds a constraint specific to the region: data localisation and cross-border processing restrictions mean that AI models used in Brazilian financial services must be trained and scored on domestically resident data. This rules out the straightforward approach of running shared models across geographies. On-platform inference with locally trained models is the only compliant architecture. For institutions building AI fraud and credit infrastructure for Brazil, this is a design requirement from the outset, not an optimisation applied later.

The regional dimension extends the opportunity significantly. Colombia, Chile, and Peru have all initiated instant payment platforms directly modelled on Pix. The fraud challenges Brazil faces today — irrevocable settlement, thin-file exclusion, social engineering at scale — will reach these markets as their real-time payment volumes grow. Brazil is, in this sense, a preview of the regional challenge, and AI infrastructure built for the Brazilian context provides a template that extends across the region. First-mover advantage in AI fraud prevention and thin-file credit modelling in Brazil is not only a Brazilian competitive advantage — it is a foundation for regional expansion as other LATAM markets reach the same inflection point.

The addressable value is $12.7 to $23.5 billion across LATAM banking

The opportunity across five decision categories reflects both the scale of the fraud problem Pix has created and the credit market it has opened.

The thin-file credit category represents the largest single opportunity across the five — larger than Pix fraud prevention itself. This distinguishes LATAM from every other region in this series, where the value at stake is predominantly in fraud prevention and compliance efficiency. In LATAM, the dominant opportunity is a credit market that financial inclusion has created and that only AI can unlock.

Part 1 of 3.

Sources

Banco Central do Brasil. Pix Transaction Statistics. 2024 Annual Data. Global Anti-Scam Alliance (GASA) / ScamAdviser / Whoscall. State of Scams in Brazil 2024. October 2024. Brazilian Public Security Forum. Brazilian Public Security Yearbook. Fraud case statistics 2018–2023. Silverguard. X-Ray of Pix Scams 2024. LexisNexis Risk Solutions / Forrester Consulting. True Cost of Fraud Study: Latin America. 2023. International Labour Organization. Labour Informality in Latin America. Mid-2023 data. Banco Central do Brasil. MED 2.0 — Special Refund Mechanism, enhanced version. Mandatory compliance from February 2, 2026. ClearingPost. Brazil Enforces Pix MED 2.0 Multi-Hop Fraud Tracking After R$6.5 Billion in 2025 Losses. March 2026. QED Investors. The Frontlines of Fraud: How Brazil Is Becoming a Global Testbed for Financial Crime Prevention. 2025. BioCatch. Brazil’s Criminal Revolution. 2025.