Revolut's 2024 annual report disclosed that the company runs thousands of machine learning models across its fraud, credit, and personalization stacks, processing transactions for 52.5 million customers while generating $4 billion in revenue. That single data point tells you more about where neobanking is headed than any product launch or press release: the winners in digital banking are not building prettier apps. They are building better ML infrastructure.
I spend a good deal of time studying how the major digital banks deploy machine learning, and the pattern is remarkably consistent. The banks that have reached profitability at scale have done so by turning ML into a core competency rather than a bolt-on feature. Revolut, Nubank, and Monzo each took different paths to get there, but the underlying thesis is the same: when you own the full data stack from transaction ingestion to model inference, you can undercut legacy banks on cost while matching or exceeding them on risk management.
The Big Three: How They Compare
Revolut has built what is arguably the most diversified ML operation among European neobanks. The company's "Sherlock" fraud detection system monitors every transaction in real time, analyzing location data, device fingerprints, and behavioral patterns to flag anomalies. Since deploying its AI-powered card scam detection, Revolut reported a 30% reduction in fraud losses from card scams related to fake investment opportunities, and the company claims its models flag over 95% of fraud in real time. But fraud is only one piece. Revolut's wealth management revenue grew 298% in 2024 to $647 million, driven in part by ML-powered personalization that recommends products based on spending behavior and financial profile. Their average revenue per user (ARPU) hit approximately $75, and over 65% of new customers arrived via word-of-mouth, keeping acquisition costs low enough to sustain a 26% net profit margin.
Nubank operates in a fundamentally different market context. With over 122 million customers across Brazil, Mexico, and Colombia, Nubank's ML stack is optimized for a population that traditional banks systematically underserved. The company built a transformer-based foundation model trained on trillions of raw transactions using self-supervised learning. Instead of hand-crafting features for each use case, the model learns general-purpose user embeddings that feed into credit scoring, product recommendation, and fraud detection simultaneously. This architecture matters because it dramatically reduces the engineering overhead of maintaining separate feature pipelines for each business function. Nubank's "Precog" system applies language model principles to predict customer intent during service interactions with over 50% accuracy improvement. The results show up in the financials: monthly ARPAC (average revenue per active customer) reached $13.40 in Q3 2025, with a cost to serve of just $0.90 per customer per month. That spread is the economic engine of the entire business.
Monzo has taken a more targeted approach, with ML investment concentrated in five primary domains as of 2025: fraud prevention, operational efficiency, product intelligence, customer experience, and financial crime compliance. Their fraud team moved from maintaining many small, specialized models to shared multi-task deep learning architectures that learn across related fraud problems. This shift improves generalization and helps catch new attack vectors before they're explicitly modeled. Monzo's reactive fraud prevention platform processes transactions through a three-layer control network: detectors (typically ML models), action controls (determining the intervention type), and action selection controls (aggregating all requested actions into a final decision). The results are measurable: only $213 per $1 million in transactions was lost to authorized push payment scams in 2023, and the system prevents an estimated 700 fraud attempts monthly. Monzo reached $1.6 billion in revenue for FY2025, up 48% year-over-year, with adjusted pre-tax profits soaring eightfold to about $145 million.
The AI-Native Personal Finance Layer
Below the full-service neobanks sits a different category: AI-native personal finance apps that lead with ML rather than banking infrastructure. Cleo and Albert represent this model.
Cleo has demonstrated that an AI-first approach to personal finance can produce serious revenue. The company reached an estimated $280 million in annual recurring revenue by mid-2025, roughly doubling year-over-year, while becoming profitable. The average annual spend per user sits around $329, with engagement reportedly running 20x higher than legacy banking apps. What makes Cleo interesting from an ML perspective is its subscription model; 59% of revenue comes from premium features like automated savings and credit building, with the remaining 41% from transaction fees on cash advances. The July 2025 launch of Cleo 3.0 introduced two-way voice conversations and long-term memory, turning the app from a chatbot into something closer to a persistent financial advisor that learns user habits over time.
Albert occupies adjacent territory, offering budgeting, savings, and cash advances with an AI assistant called "Genius." The company charges $8.99 monthly and focuses on delivering personalized financial guidance. Albert's approach is less about building novel model architectures and more about applying off-the-shelf ML techniques to spending categorization, savings optimization, and cash advance underwriting.
MoneyLion takes a broader approach, bundling investing, credit-building, and banking into a single platform. Its Instacash feature provides cash advances up to $500, and the company uses ML models for credit decisioning and personalized product recommendations across its product suite.
For incumbent banks looking to add similar personalization without building from scratch, Personetics provides an interesting case study. The company serves over 150 million users across 130 banks in 35 markets, offering an AI-powered "cognitive banking" platform that generates personalized insights and product recommendations. Their reported results are notable: Huntington Bank delivers 14 million insights monthly across 96 use cases with a 4.7/5 customer satisfaction rating, and Synovus saw an 80% boost in product cross-sell among engaged users. I view Personetics as evidence that the ML-driven personalization playbook works even when layered onto traditional banking infrastructure.
The Data Flywheel: Why ML Infrastructure Is the Moat
The core advantage these companies share is not any individual model but the data flywheel that feeds continuous improvement. Every transaction processed, every fraud alert reviewed, every credit decision made generates labeled training data. More users produce more data; more data trains better models; better models attract more users. This flywheel dynamic explains why Nubank can serve customers at $0.90 per month while traditional Brazilian banks spend multiples of that on per-customer servicing.
Revolut's ARPU of roughly $75 might look modest compared to traditional bank revenue per customer, but when your marginal cost to serve is near zero and your customer acquisition is 65% organic, the unit economics work. Nubank's ARPU of about $100 annually is roughly one-quarter of what incumbent Brazilian banks generate, but the gap is closing as the company cross-sells lending, insurance, and investment products into its base.
The models themselves matter less than the infrastructure supporting them. Nubank's foundation model approach, where a single architecture generates embeddings reusable across dozens of downstream tasks, is an example of the kind of engineering that creates durable advantage. When you can deploy a new credit product and immediately benefit from the representations learned across your entire transaction history, your time-to-market advantage over banks that hand-engineer features for each use case is significant.
Challenges That Should Not Be Dismissed
I want to be direct about the headwinds, because the neobank story is not all upward trajectories.
Profitability pressure is real. Monzo only achieved its first annual profit in 2024 after eight years of losses. Revolut has been profitable for four consecutive years, but a meaningful portion of that comes from interchange fees and foreign exchange spreads that face regulatory compression in multiple jurisdictions. The UK's Payment Systems Regulator, the EU's PSD3 framework, and Brazil's evolving open banking rules all have the potential to squeeze margins.
Regulatory complexity scales with geographic ambition. Revolut is pursuing banking licenses in over 10 countries simultaneously. Each jurisdiction brings its own capital requirements, consumer protection rules, and data residency obligations. Monzo received a $27 million fine from the FCA in 2025 for historical anti-money laundering failures, a reminder that ML sophistication in fraud detection does not automatically satisfy regulatory expectations around financial crime controls.
Credit risk is the elephant in the room. As these banks move beyond payments and savings into lending, their ML credit models face their first real stress test. Nubank's credit portfolio has grown rapidly, but default cycles in emerging markets can be brutal. Revolut's customer lending portfolio grew 86% in 2024 to $1.2 billion; the question is how those models perform when the economic cycle turns.
Looking Ahead
I expect the gap between ML-native neobanks and traditional banks to widen over the next three to five years, but not uniformly. The neobanks with the deepest data flywheels and the most disciplined approach to credit risk will likely continue gaining market share. The ones that chase growth through aggressive lending without sufficiently stress-tested models could face painful corrections.
The AIFI directory tracks several of the companies mentioned here, and what strikes me about the sector is how quickly the competitive dynamics have shifted from "who has the best app" to "who has the best data infrastructure." That shift is not likely to reverse. Whether any individual neobank sustains its current trajectory depends on execution across ML engineering, regulatory compliance, and credit discipline simultaneously. Getting all three right is harder than it looks from the outside.