Banco Santander said specialised artificial-intelligence models reduced serving costs from about €5 to approximately €0.03 per interaction in selected experimental uses. The Spanish bank announced a three-year collaboration framework with Nvidia on 19 October. In one customer-service experiment, Santander said conversation quality improved while costs fell by more than 50%.
The work is being conducted through Santander AI Lab before any broader deployment across the bank. The lab is moving from large general-purpose models towards smaller systems specialised for particular banking tasks. Computer Weekly reported that the bank is testing these models to choose the most suitable option for each workload.
SMALLER MODELS TARGET BANKING TASKS
Santander has built a Model Arena to assess open-weight models against banking tasks and is using Nvidia NeMo Customizer to specialise them. The bank said it is also developing tools to design, simulate, evaluate and supervise banking agents. Its controlled testing includes performance engineering and adversarial checks before capabilities move into wider use.
The new framework covers foundation models, AI agents, fraud and anti-money-laundering capabilities, real-time decisioning and advanced simulation. Santander said the planned work includes a banking digital twin. It is industrialising model inference within its own architecture using Nvidia Dynamo-Triton, which it linked to the measured serving-cost reduction.
VALUE DEPENDS ON CONTROLLED DEPLOYMENT
Computer Weekly reported that Santander's 2026–28 plan targets €1 billion of business value from AI through lower costs and higher revenue. The publication said the bank had reported €35 million of value in the first three months of 2026 and expected €200 million by year-end. Those broader targets are separate from the experimental per-interaction cost measure announced this week.
The immediate test is whether Santander can reproduce the laboratory gains across regulated production systems without weakening model governance, resilience or customer outcomes. The three-year framework provides time to move from experiments towards governed agents in fraud, anti-money laundering and service operations. Evidence of production-scale deployment and realised financial value will be the next material milestones.