Forrester says banks are training private models on their own transaction data. It argues the model itself gives little lasting advantage. The durable asset is proprietary behavioural data and the record of outcomes that follow customer actions. Forrester points to Revolut's PRAGMA, built with NVIDIA and trained on billions of events. It also names Stripe, Visa, Nubank, JPMorgan Chase, Mastercard and PayPal as investors in similar work. FICO sells a vendor-built alternative trained on industry-wide data. Forrester warns that few institutions have the data scale, skills, capital and infrastructure to pre-train from scratch. Most should fine-tune, enrich or combine existing models instead. The share planning foundation-model investment comes from Forrester's 2026 priorities survey.
What changed
Banks relied on separate models per task and general-purpose AI, with 'customer 360' profiles as the data goal.
What it unlocks
Using transaction history as one shared behavioural layer across fraud, credit, marketing and recommendations.
- 18% plan to invest in foundation models
- 24 billion events in Revolut's PRAGMA
- 111 countries covered
What you need to act on it
- large-scale proprietary transaction data
- AI expertise and infrastructure
- outcome feedback captured from operations
Sources