How I choose between classic ML, deep learning, foundation models, and LLMs benchmarked against Revolut's PRAGMA, grounded in PAYG Mobile Device Finance

tl;dr

"Should we use ML or an LLM?" is the wrong question. It is a category error dressed up as a strategy. Uhm, yes, this was inspired by an intellectual debate with my mate.

The right question is the one product leaders are supposed to ask about everything: what is the job to be done, and what is the cheapest, most reversible, most defensible way to get it done well enough to win? Method follows job. It never leads.

When Revolut, one of the most AI forward consumer fintechs on the planet, sat down to build a model that learns from card transactions, transfers, in-app navigation, and customer communications to do credit scoring, fraud detection, lifetime value, and product recommendation, they did not reach for a large language model. They built PRAGMA, an encoder foundation model with a custom key–value–time tokeniser and they wrote down, in black and white, why feeding their data to an LLM would have been the wrong tool.

That single decision is the most important thing in this prose. A team that could have used any model in the world looked at structured, high velocity, numerically dense financial event streams and concluded: this is not a language problem, so do not use a language model.

This piece unpacks that decision into a usable framework, then stress-tests it across the domains we actually live in: consumer app personalisation, Mobile Device Finance (MDF), thin-file credit risk in emerging markets, fraud and anomaly detection in PayGo, unit-economics modelling, and even the unglamorous places; internal tool ops, people ops, the emails. I'll show you where classic ML still quietly wins, where foundation models earn their keep, where LLMs are genuinely transformative, and most usefully where PRAGMA itself fails, because the failures are where the real lessons live.

Picking the wrong model class is the most expensive cheap looking decision a product org makes, and it cannot be made in a corridor.

Part 1; The wrong question, and why everyone keeps asking it

Walk into almost any fintech product review in 2026 and you'll hear some version of this exchange:

"Can't we just use an LLM for that?"

It's a reasonable sounding sentence. It is also, nine times out of ten, the moment a roadmap quietly goes wrong.

The phrase smuggles in three assumptions that don't survive contact with reality:

1. That "an LLM" and "ML" are two points on a single quality ladder, where the LLM is simply the newer, better rung.
2. That generality is free; that a model trained to do everything will, naturally, do your thing well.
3. That the bottleneck is the model. It almost never is. The bottleneck is the data exhaust, the control surface, the latency budget, and the regulatory constraint.

Let me dismantle the first assumption directly, because the other two collapse once it's gone.

LLMs are not the top of a ladder. They are one corner of a landscape. They are extraordinary at one family of jobs; understanding and generating language, and reasoning over it and merely adequate to bad at a different family of jobs that classic machine learning has owned for two decades: prediction over structured, tabular, high cardinality, numerically precise data. A gradient boosted decision tree from 2017 will still, today, out predict a frontier LLM on a well specified credit default table, run a thousand times cheaper, return an answer in single digit millisecond, and hand you a feature importance chart your risk committee can actually read.