Starling plans UK-hosted AI model trained on 15 billion transactions

Starling is developing an AI model for banking, trained on more than 15 billion anonymised transactions and hosted in a UK cloud. The bank’s first planned use is fraud detection, with production targeted for early 2027.

Watch Desk analysis

What happened

Payments Industry Intelligence reports that Starling is building the model with NVIDIA and UK cloud provider Civo. The report says the project uses NVIDIA’s Transaction Foundation Model blueprint and is intended to remain within UK jurisdiction for both training and hosting.

Starling plans to use the model first to strengthen fraud detection, then extend it into its existing AI assistant. Its longer-term ambition is a wider intelligence layer, which could support services such as proactive budgeting guidance.

Why it matters

This is a concrete banking-AI project built around a bank’s own transaction data, rather than a general-purpose model offered as a ready-made service. Keeping the infrastructure in the UK also gives the project a resilience and data-control angle: Starling says it wants more choice beyond a small number of global infrastructure providers.

The near-term test is narrower than the grand vision. Fraud detection is the first intended application; the broader assistant and intelligence-layer plans remain future ambitions, not features customers can use today.

Our read

The interesting bet is that specialised data and control over where a model runs can matter as much as raw model capability in financial services. Fifteen billion transactions is a substantial training pool, but the useful measure will be whether the model improves fraud detection without making false alarms or customer support worse. Banking AI gets its report card from what happens to people’s money, not from a large number in a headline.

What to watch

• Whether the model reaches production in early 2027 as planned.
• What Starling says about measured fraud-detection performance and false positives.
• Whether the planned assistant and wider intelligence layer become real customer-facing services.

Discussion spark: Would you trust a bank-specific AI model more because it is trained on the bank’s own data and hosted in the UK, or would you need independent evidence of better fraud detection first?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.