Every sector below generates years of behavioural records — transactions, sequences, graphs, and tables — that current models barely scratch. A foundation model trained on that data produces shared representations that transfer across every downstream decision the business needs to make.
Most financial institutions run separate, siloed models for fraud detection, credit scoring, AML, and churn — each requiring its own feature engineering, retraining pipeline, and maintenance overhead. Stripe took a different approach: a single payments foundation model trained on billions of transactions, whose embeddings are reused across every risk problem in the business. Card-testing fraud detection moved from 59% to 97% with no increase in false positives. Any institution sitting on years of transactional data has the raw material to do the same — replacing brittle, task-specific models with a single behavioural foundation that improves every downstream decision simultaneously.
Law firms and litigation funders make expensive decisions — which cases to take, how to price risk, when to settle — using experience and intuition rather than systematic pattern recognition. Yet the data to do this properly already exists: case type, jurisdiction, opposing counsel, claim value, duration, outcome, settlement terms. A foundation model trained on a firm's own case history produces embeddings that can be reused across outcome prediction, pricing, funding decisions, and business development. The same model that identifies likely winners also flags underpriced risk and surfaces the factors that consistently drive settlement — without rebuilding from scratch for each question.
Farming generates more structured, sequential data than almost any other sector — soil readings, weather patterns, input costs, yield records, machinery telemetry — and almost none of it is used predictively. The Stripe parallel is instructive: rather than building separate models for yield forecasting, disease risk, and equipment failure, a single foundation model trained on a farm or cooperative's historical data produces representations that transfer across all of these problems. For operations where a wrong decision on inputs or timing costs tens of thousands of pounds, the performance uplifts from this approach are not marginal. They are transformative.
Every project a contractor completes is a data point — costs, labour, materials, subcontractor performance, defects, delays, weather conditions. Individually these are records. Collectively they are a training dataset. The insight from payments AI is that a single foundation model trained on this sequential project data can produce embeddings that transfer across bid accuracy, delivery risk, safety prediction, and subcontractor scoring — problems that today each have their own spreadsheet or crude model, if they have anything at all. In a sector where margins are typically 2–3% and overruns are endemic, that level of systematic prediction would be a genuine competitive advantage.
Housing associations hold rich, longitudinal data on repairs, rent payments, tenancy behaviour, void patterns, and property condition — often going back decades. The challenge is that this data is used reactively rather than predictively, and the analytical tools applied to it haven't meaningfully advanced. A foundation model trained on tenancy and property histories produces a shared representation that transfers across void prediction, maintenance scheduling, arrears risk, and tenant vulnerability — the same architecture solving multiple operational problems from a single trained model. Given the regulatory pressure to move from reactive to proactive management, the case for building this capability is both financial and compliance-driven.
Network operators collect continuous structured data — inspection records, pressure readings, fault histories, maintenance logs, sensor telemetry — but the models applied to it are typically narrow and task-specific. The more powerful approach, demonstrated in financial services, is to train a foundation model on the full history of asset behaviour, producing embeddings that transfer across failure prediction, maintenance scheduling, and capital allocation decisions. A network asset that looks unremarkable in isolation may carry a recognisable pattern of pre-failure behaviour when the full sequence is modelled. That distinction — between point-in-time snapshots and learned sequential behaviour — is what separates current practice from what is now possible.
A freight forwarder's operational history — routes, carriers, timings, customs outcomes, delay patterns, cost variances — is a sequential record of how their specific network behaves under real conditions. Generic benchmarks and industry averages don't capture it. A foundation model trained on that proprietary history learns the latent structure of their network: which carrier-lane combinations degrade under certain conditions, which customs routes carry hidden delay risk, which customers generate systemic exceptions. Those learned representations transfer across delivery prediction, carrier scoring, cost forecasting, and customer risk — multiple business problems addressed by a single model that no competitor can replicate because they don't have the data.
Every programmatic campaign generates a continuous stream of structured data — impressions, bid prices, win rates, click-through rates, conversion events, audience segments, contextual signals, and timing — across thousands of simultaneous auctions happening in milliseconds. Most advertisers and agencies model these as separate optimisation problems: budget allocation in one tool, audience targeting in another, bid strategy in a third. A foundation model trained on a brand or agency's full auction history learns the latent structure of how their specific audiences, creatives, and placements interact — representations that transfer across bid optimisation, budget forecasting, audience discovery, and incrementality measurement from a single trained model. Just as Stripe's foundation model learned that certain transaction sequences signal fraud before any individual signal becomes obvious, a bidding foundation model learns which combinations of contextual and audience signals consistently precede conversion before a human analyst would notice the pattern. For brands spending at scale, the compounding effect of that predictive edge — applied simultaneously across every auction decision — represents a step change in media efficiency that rule-based bidding strategies cannot match.
If you're sitting on a transaction history, an event log, or a table that your current models aren't doing justice to, we'd like to hear about it.