Santander has announced work with Nvidia on tools to evaluate AI models and agents for banking tasks. The bank's AI lab has built a Model Arena that tests open-weight models against real banking tasks and specialises them using Nvidia's NeMo Customizer. Santander says that in selected experiments, moving to smaller specialised models cut serving costs from €5 to around €0.03 per interaction. The two companies will also work together on agents, fraud and anti-money laundering, and banking digital twins.
What it means: Santander's announcement shows big banks moving away from the idea of one general model for everything. For specific tasks such as customer service or document classification, a smaller model trained on the bank's own data can be good enough and far cheaper. Going from €5 to a few cents per interaction, even if only in selected experiments, changes the economics of any large-scale application. The second important element is systematic evaluation. An internal environment where each new model is tested on real tasks before going live is exactly what supervisors expect for model risk management. For Greek banks and fintechs, the approach also works at smaller scale. Open-weight models can be hosted on European infrastructure, which eases GDPR and DORA compliance, and smaller sizes reduce computing needs. The key is good evaluation data in Greek, because many models perform noticeably worse in our language than they do in English.
Related service: myDATA integrations
Related stories
- World unveils human-verification suite for banks to fight deepfakes
- Truist shifts investment toward deposits and digital banking
- Salmon secures $140 million for digital banking and lending in the Philippines
- Marqeta and Coinbase offer stablecoin-backed cards to enterprise clients
- ISX Financial and Mastercard extend account-to-account payments across borders