Machine Learning at Monzo in 2025

It’s been a few years since we last wrote about machine learning at Monzo. Back then, the discipline was still young — we were putting the right data, infrastructure, and foundations in place.

Three years on, those foundations have matured into a core capability that spans the full spectrum of our work: defending against financial crime, supporting responsible credit decisions, scaling operations, and powering intelligent, personalised experiences in our products.

Machine learning now underpins how we protect customers, enable growth, and anticipate their needs — and its role is only expanding.

In 2025, our ML investment focuses on five primary areas that reflect the discipline’s growth and direction:

Each of the above areas are at a different stage of maturity, but together they show how we’re embedding intelligence into the fabric of Monzo. And while this post focuses on customer-facing applications, we’re also increasingly using AI internally to make our people and processes more effective — driving productivity and impact across Monzo.

Machine learning is at the core of how Monzo protects customers from fraud and financial crime — preventing millions in losses each year from unauthorised transactions, scams, and money-muling.

Our early systems were built to detect the behaviour of fraudsters — identifying known attack patterns and flagging suspicious activity. These defences remain vital but must continuously evolve to counter adaptive, adversarial behaviour. As attackers learn and adapt, staying one step ahead and maintaining defensive advantage demands rapid iteration and tight integration between models, data pipelines, and monitoring.

Our reactive fraud-prevention platform provides that foundation, enabling near real-time detection and response. Building on it, we’ve advanced our modelling through multi-task deep learning, moving from many small, specialised models to shared architectures that learn across related fraud problems — improving generalisation and spotting new attack vectors sooner.

We’re now investing in representation learning to capture behavioural structure directly from data. Self-supervised embeddings of customers and transactions act as durable, reusable features for supervised models, improving detection in low-label, high-imbalance settings. This shift — from identifying known patterns to modelling normal behaviour — marks a major step forward: still reactive when needed, but increasingly adaptive, data-efficient, and resilient.

Operations at Monzo have always been a collaboration between people and technology. Machine learning is now amplifying that partnership — helping us operate at scale while keeping our service fast, accurate, and human. Our research shows that customers want service that’s fast, frictionless, and empowering — handling simple tasks quickly and independently while knowing support is there when it matters. Speed remains the strongest driver of satisfaction, and that’s why we’re investing in automation: to make everyday interactions instant, effortless, and reliable, while keeping humans at the heart of complex or sensitive situations.

We’re building in-house solutions using foundational large language models (LLMs) to power both customer-facing and back-office workflows. These systems are already delivering strong results: automated handling of a subset of customer queries, faster decisioning workflows, measurable gains in customer satisfaction, and faster, more consistent, and context-aware support.

This automation doesn’t just make us more efficient — it elevates customer experience. Issues are resolved more quickly and consistently, friction is reduced, and our teams can focus on the complex cases where human judgment and empathy matter most.

Automation also extends into broader predictive and assistive ML: models that understand intent, classify and route requests, and infer context from past interactions. Alongside this, we’re developing systems that combine operations research (OR) with predictive modelling to match the right customer service agent to the right work. These systems balance task importance, adherence to SLAs, and forecasted agent availability and capacity. It’s a genuinely novel approach in this problem space.

We’re now working on automating complex operational processes, which involves agents that reason, build and execute plans, and securely call a range of tools. We then apply multi-stage evaluation strategies that ensure correctness, specificity, and tone of voice — enabling these solutions to be safely deployed into production with humans in the loop. Our work focuses on making  these systems learn safely from real interactions through controlled feedback loops, rigorous evaluation, and subject matter expert oversight at every stage. As the models evolve, we can steadily extend the boundaries of what can be safely automated, while expert input remains essential to ensure accuracy, quality, and fairness. This balance allows our systems to improve continuously while maintaining the transparency and trust that underpin everything we do.

Combining AI with Monzo’s deep operational context creates a powerful opportunity to redefine how human and machine intelligence work together — improving both how we serve customers and how our teams work behind the scenes.

Machine learning plays an important role in our borrowing products. We use machine learning to predict credit risk, understand utilisation, and estimate product propensities so we can offer the right products within our risk appetite. Improved predictive performance lets us responsibly lend to more customers at appropriate prices and limits, expanding access to manageable credit. 

Our current work focuses on developing stronger feature representations, applying modern algorithms, and building a clearer understanding of different customer segments. As our portfolio evolves — including new Flex variants and expanded business lending — we continually monitor model performance amid shifting behaviours and economic conditions, while maintaining strong standards of explainability and fairness.

Monzo’s mission is to make money work for everyone. That means building experiences that are tailored to individual context — understanding each customer’s needs, context, and goals, and helping them make better financial decisions without friction.

Personalisation is our newest and most ambitious area of investment, turning machine learning from a supporting capability into a core part of how we design products. Our early focus is on three fronts:

These capabilities extend across product verticals — from lending and credit to savings and investments — where personalisation helps tailor pricing, limits, and advice to each individual’s situation. The opportunity is to make intelligence feel invisible: turning the complexity of machine learning into the simplicity of a product that just works.

The machine learning platform is the backbone of how we build intelligent products at Monzo. It lets teams experiment quickly, deploy safely, and integrate seamlessly with the rest of the Monzo stack.

Supporting both predictive models and LLM-based systems demands different strengths. Predictive ML needs robust feature management, versioning, and continuous monitoring to stay accurate and fair as behaviour shifts. LLM applications require orchestration, retrieval, evaluation, and human feedback — all built with the rigour, observability, and compliance expected in financial services.

The platform unifies these capabilities so practitioners can move from idea to production with confidence. It’s what turns research into running systems — and makes machine learning a natural part of how we ship software at Monzo.

Machine learning at Monzo has grown from a handful of models into a discipline that shapes how we protect customers, run our operations, and design intelligent experiences in our products. What began as fledgling discipline is now a foundation for how we build.

In the year ahead, we’re scaling across every dimension: advancing adaptive fraud detection, expanding LLM applications in operations, deepening personalisation, driving innovation in credit decisioning, and strengthening the platform that supports them all. These are complex, high-impact problems that sit at the intersection of technology, product, and trust — the same principles that have guided our growth from foundation to maturity. We’re building systems that are not only intelligent but also safe and transparent — supported by strong governance,  rigorous monitoring, and a deep commitment to customer confidence. These principles let us move fast while keeping trust at the centre of everything we build. Together, engineers, data scientists, and machine learning scientists are shaping this future — solving hard problems that matter. If you’re excited by the idea of applying modern ML to help make money work for everyone, now is the time to help us build it.