Summary

• Santander reports reducing serving costs from approximately €5 to €0.03 per interaction in selected experimental AI workloads.
• A separate customer service experiment improved conversation quality while reducing costs by more than 50%.
• A new three-year NVIDIA collaboration framework covers specialised models, agent evaluation, financial crime applications, and advanced simulation.

Banco Santander is testing smaller, specialised artificial intelligence models as an alternative to using large general-purpose systems for every banking task, reporting substantial reductions in the cost of running selected applications.

Working with NVIDIA, the Spanish banking group has reduced model serving costs from approximately €5 to €0.03 per interaction in particular experimental workloads. A separate customer service experiment improved conversation quality while reducing costs by more than 50%, according to results released on 19 October.

The work is being carried out through Santander AI Lab, which evaluates emerging technology before wider deployment. The bank has developed its own environment for comparing models against defined business tasks and is extending its NVIDIA relationship through a three-year collaboration framework.

Although the reported improvements illustrate how different infrastructure and model choices can affect AI expenditure, they are not independently verified savings across Santander’s operations. The two cost comparisons also describe different experiments and cannot be treated as interchangeable measures of performance.

Matching models to the work they perform

Large general-purpose AI models are designed to handle a broad range of requests, but their flexibility comes with computing requirements that may be unnecessary for narrower business processes. Banks process substantial volumes of structured information and repeat many defined activities, creating opportunities to use systems selected for particular tasks rather than maximum general capability.

Santander’s Model Arena compares open-weight models against banking workloads to determine which offers an appropriate balance of performance, cost, and operational suitability. The bank is also using NVIDIA NeMo Customizer to develop more specialised models, allowing the evaluation process to extend beyond selecting an existing system.

Open-weight models give developers access to parameters that can be adapted or deployed in different environments, subject to the relevant licence. They may provide greater flexibility over model configuration and infrastructure, although access to those parameters does not establish that a system is secure, accurate, or suitable for regulated work.

A model handling a routine customer enquiry does not necessarily require the capabilities of a system intended to analyse complicated financial documents. Similarly, an application producing predictions from structured transaction information presents a different technical problem from one interpreting natural language.

Santander is developing specialised intelligence using NVIDIA Kumo, a family of open-weight foundation models designed for structured data. That work targets predictive capabilities within core banking processes, where useful information may depend on relationships between records rather than conversational fluency.

The bank has not published an independently assessed comparison of every model evaluated through Model Arena. The system’s immediate purpose is to establish a controlled way of making those comparisons before successful applications are considered for wider use.

Inference spending becomes an operating expense

AI costs continue after the initial development and training work. Each time an application requests a model output, computing resources must execute the necessary operations, a process known as inference.

That expenditure can become material when an organisation handles large numbers of requests. A difference of a few cents per interaction may appear relatively small, but repeated across customer services and internal processes it can affect the economics of deploying AI.

Santander is using NVIDIA Dynamo-Triton to improve the efficiency of inference within its own architecture. The bank associates its reported reduction from approximately €5 to €0.03 per interaction with the model specialisation and serving improvements achieved in selected experimental cases.

The comparison is striking, although its commercial interpretation depends on what each interaction involved, how output quality was measured, and whether the same workload could be reproduced under normal operating conditions.

A cheaper answer does not necessarily mean a cheaper completed business process. If the response requires additional employee review, generates errors, or fails to resolve the original task, some of the apparent savings may be absorbed by further work.

The bank’s separate customer service experiment takes that distinction further by comparing both conversation quality and cost. Santander reports improving the former while reducing the latter by more than half, suggesting that model efficiency does not inevitably require weaker performance in narrowly defined applications.

However, the bank has not disclosed enough information about experimental volumes, quality measures, or production operating conditions to extrapolate those savings across its customer service operations.

Controlled evaluation remains part of deployment

Financial applications require closer examination than whether an AI model can generate a convincing response. Incorrect information, inappropriate access to records, or unreliable automated actions can create customer, financial, and regulatory consequences.

Santander AI Lab is developing tools to design, simulate, evaluate, and supervise banking agents. It is also carrying out adversarial testing of conversational systems within controlled environments to identify weaknesses before broader implementation.

Agents introduce additional responsibilities because they may do more than prepare information. Depending on their configuration, they can interact with applications and undertake sequences of actions using permissions granted by the organisation.

An agent reviewing a transaction for possible fraud, for example, needs access to relevant information and a clear distinction between identifying suspicious activity and taking an action that affects a customer. Decisions with financial or legal consequences may require additional controls, escalation, or human approval.

The bank’s evaluation framework provides a way to examine those requirements, although its existence does not establish that every proposed agent has passed testing or entered live operation.

For regulated organisations, the ability to explain why a model was selected and how its performance was assessed can also be useful during internal risk reviews and technology procurement. A system may achieve favourable benchmark results without being suitable for the data, processes, or operational requirements of a particular business.

The partnership reaches beyond customer service

Santander and NVIDIA have agreed a three-year framework covering foundation models, AI agents, fraud and anti-money laundering capabilities, real-time decision-making, and advanced simulation.

The partners also intend to explore a banking digital twin, an approach that could use simulations to assess operational scenarios before changes are introduced into actual processes.

Such applications would require detailed representations of the systems and conditions being examined. Their value would depend on how closely simulated behaviour corresponds to real operational outcomes, particularly when transaction patterns or regulatory requirements change.

The agreement establishes areas for future collaboration rather than a timetable for rolling out every proposed technology across Santander’s international businesses.

NVIDIA supplies the computing infrastructure and software used in the experiments, giving it a commercial interest in demonstrating that its technology can improve model economics. Santander’s own testing results provide more concrete evidence than a general partnership announcement, but remain company-reported findings.

Other financial institutions assessing comparable investments face a similar calculation. The most capable model may not be the most economical choice for every task, especially when serving costs, evaluation requirements, data integration, and ongoing supervision are taken into account.

Santander has demonstrated that specialised models and different serving arrangements can produce sharply lower costs in selected experiments. Whether those improvements translate into sustained financial benefits will depend on deployment performance, reliability, and the additional work required to operate AI within established banking controls.