What proprietary AI has Santander built in-house?
Santander's in-house AI is built at group level and is moving from vendor models toward its own specialised ones. On October 19, 2026 the group said Santander AI Lab had built Model Arena to evaluate open-weight models against banking tasks, is specialising models with NVIDIA NeMo Customizer and serving them 'within Santander's own architecture', and in selected experimental use cases cut serving costs from about €5 per interaction to about €0.03. In production the group runs more than 280 process automation agents across credit, fraud, KYC and operations, Openbank AI models that process about 100,000 anti-money laundering alerts a year, and an in-house evaluation framework that scores AI products on performance, safety and fairness (June and July 2026). The AI Lab has open-sourced 11 projects (June 25, 2026) and the A2K agent protocol (September 23, 2026). Staff tools come from outside: Microsoft Copilot, OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini, under what the bank calls a 'secure, multi-provider approach'. In the United States the bank's own decisioning is older: Santander Consumer USA assigns every auto applicant a 'proprietary custom score' that places them in one of 100 pricing tiers, per the 2025 10-K, while its machine-learning credit-risk models run on FICO Platform.[1][2][3][4][5][6][7]
Data as of Oct 22, 2026 · Santander profile updated Oct 21, 2026 · 16 sources cited on this page · 140 banks profiled on this site · Markdown
Which AI systems has Santander built itself, and on what?
Group systems unless marked US. 'Built on' names outside models or platforms where Santander has said so; Santander does not say which group systems US staff use.
| Platform or model | What it does | Built on | Since | Status |
|---|---|---|---|---|
| Model Arena | Evaluates open-weight models against banking tasks to pick the right model for each workload; being extended to design, simulate, evaluate and supervise banking agents | Open-weight models; built with NVIDIA support | Announced October 19, 2026 | Experimental[1] |
| Specialised banking models | Smaller models specialised for specific banking tasks; serving cost cut from about €5 to about €0.03 per interaction in selected experiments; one customer-service model cut costs by more than 50% with better conversation quality | NVIDIA NeMo Customizer; NVIDIA Kumo open-weight models for structured data; NVIDIA Dynamo-Triton for inference in Santander's own architecture | October 2026 | Experimental use cases[1] |
| Process automation agents | More than 280 agents in production automating manual tasks across credit, fraud, KYC and operations; in Brazil card fraud claims around 95% faster, up to 90% automated, error rate below 1% | Not disclosed | Reported June 22, 2026 | In production[2] |
| Openbank AML models | AI models that process approximately 100,000 anti-money laundering alerts a year; investigations that took hours now take minutes | Not disclosed | Reported June 22, 2026 | In production[2] |
| AI product evaluation framework | One common way to evaluate AI products on performance, behaviour, safety and fairness; generates test cases from each product's requirements and monitors for drift in production | Built by the Data & AI Science team | Described July 30, 2026 | Bank-wide capability[3] |
| Harness Engineering for agents | Santander's term for the loop around a model that manages state, tool calls, evaluation, boundaries and human escalation; Autoguardrails, released as open source, tunes guardrail policies against adversarial inputs | Model-agnostic | July 27, 2026 | Used in agent builds[8] |
| Open-source AI Lab projects | 11 projects on GitHub under Apache 2.0, including gen-fraud-graph (synthetic fraud networks), mech-gov-framework (LLM governance) and mutatis-mutandis (discrimination testing); no real customer data | In-house | June 25, 2026 | Published[4] |
| A2K protocol | Open specification for AI agents to discover, query, cite and audit governed enterprise knowledge, over MCP, A2A and HTTPS | In-house | September 23, 2026 | Published[5] |
| Proprietary custom score (US) | Every Santander Consumer USA auto applicant gets a 'proprietary custom score' from FICO scores, debt-to-income and loan-to-value ratios and over 30 other predictive factors, placing them in one of 100 pricing tiers | In-house; not called AI in the filings | 2025 10-K (March 2, 2026) | In production[6] |
| Credit-risk machine learning (US) | Approve or decline decisions, adverse-action reasons and probability of default for pricing across a $60 billion portfolio and more than three million customers | FICO Platform | May 7, 2025 | In production; 2025 FICO Decisions Award[7] |
Which Santander teams build its AI?
Santander AI Lab, led by José Manuel de la Chica, 'explores and tests emerging AI technologies before broader deployment across the bank' and built Model Arena; Iñaki Bernal is chief artificial intelligence officer, per the October 2026 NVIDIA announcement. The Lab also authors research with Spain's CSIC, including a July 2026 report on small and specialised models.[1][9][10]
Delivery runs through the Data & AI operating model, created in 2025 with more than 4,500 employees 'who work to ensure the widespread rollout of AI in every command chain'. The evaluation framework comes from the Data & AI Science team, and Head of Responsible AI Luis Ibáñez's teams test conversational systems for over-refusal before release.[11][3][12]
The US business has no named AI team on the record. Its disclosed models sit in Santander Consumer USA's auto originations, where an automated process assigns custom scores and a team of underwriters reviews exceptions and deal structures.[6][13]
When does Santander build AI rather than buy it?
Santander buys general-purpose models for staff and builds the platform around them. In August 2025 its chief data and AI officer Ricardo Martín Manjón described the architecture as 'modular and vendor-agnostic, integrating seamlessly with OpenAI and other large language models', with the rule 'build once, reuse everywhere'. In June 2026 the group listed Microsoft Copilot for everyday productivity and OpenAI, Anthropic, Google and G42 for specialised capabilities, and said it does not share customer data externally to train third-party models.[14][2]
For core banking tasks the direction is toward its own models. The AI Lab says Santander is 'progressing from open-weight models towards specialised models and governed banking agents designed for greater control, safety and lower operating costs at scale', and its report with Fundación General CSIC argues that small and specialised models give 'greater control over data, weights and traceability'. In US auto lending the split is explicit: the custom score is the company's own, while the machine-learning credit-risk models run on FICO's platform.[1][10][6][7]
What has Santander not disclosed about its own AI?
Santander has not named the open-weight models it specialises, said which of the Lab's experimental models are in production, or broken down the 280-plus agents or €84 million of first-half 2026 AI business value by country. No US-built AI system, US AI team or US AI budget has been published, and the group does not say which of its in-house tools Santander US or the acquired Webster business use.[1][2][15]
No patent protects a named Santander AI system on the public record; the only Santander US filing found is a 2023 Santander Consumer USA application on simulating loan payments that does not describe AI. The 10-K does not call the proprietary custom score AI or machine learning.[16][6]
What does the Santander profile record on this topic?
The rows below are the entries of the Santander AI strategy profile that bear on this page, as of Oct 21, 2026; the profile holds the full record of 37 dated moves, 16 use cases and 49 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Machine-learning auto credit risk on FICO Platform | Approve/decline decisions, adverse-action reasons and probability-of-default for pricing in the US auto business, with faster model development and governance. | In production | [7] |
| A2K: governed agent-to-knowledge protocol (Santander AI Lab) | An open specification for AI agents to discover, query, cite and audit enterprise knowledge bases with scoped authority and audit controls. | Announced | [5] |
| Synthetic fraud-network data (Santander AI Lab) | gen-fraud-graph generates artificial networks of fraud transactions and behaviours, with no real customer data, for testing detection models. | Announced | [4] |
| Automated custom score and 100 auto pricing tiers | An automated originations process assigns every auto applicant a proprietary custom score (FICO scores, DTI, LTV and over 30 other predictive factors) and one of 100 pricing tiers that are continuously monitored and adjusted; a team of underwriters reviews exceptions. Not described as AI in the filings. | In production | [6][13] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Sep 23, 2026 | Santander AI Lab launches A2K, a protocol for agent-to-knowledge access | An open specification, running over MCP, A2A and HTTPS, that lets enterprise AI agents discover, select, query, cite, audit and reconcile knowledge bases with scoped authority and lifecycle management. | [5] |
| Jun 25, 2026 | Santander AI Lab publishes 11 open-source projects | Apache 2.0 repositories including gen-fraud-graph (synthetic fraud networks) and mech-gov-framework (language-model governance mechanisms); no real customer data. | [4] |
| Mar 2, 2026 | 2025 10-K: every auto applicant automatically scored into one of 100 pricing tiers | Santander Holdings USA says each auto applicant is automatically assigned a proprietary custom score built from FICO scores, debt-to-income and loan-to-value ratios and over 30 other predictive factors, which sets one of 100 continuously monitored pricing tiers, with underwriters handling exceptions; the filing does not call the score AI or machine learning. | [6] |
| May 7, 2025 | FICO award for machine-learning credit risk | The US auto business wins a 2025 FICO Decisions Award for AI, Machine Learning & Optimization for rebuilding credit-risk modelling on FICO Platform. | [7] |
Does Santander build its own AI models?
Increasingly, at group level. Santander AI Lab specialises open-weight models for banking tasks with NVIDIA NeMo Customizer and serves them inside Santander's own architecture, cutting serving costs from about €5 to about €0.03 per interaction in selected experiments (October 2026). Staff productivity tools come from Microsoft, OpenAI, Anthropic and Google. In the US, Santander Consumer USA uses its own proprietary custom score for auto pricing.
What is Santander Model Arena?
A tool Santander AI Lab built to evaluate open-weight models against banking tasks and identify the right model for each workload, announced with NVIDIA on October 19, 2026. The Lab is developing it further to design, simulate, evaluate and supervise banking agents, and the group set a three-year collaboration framework with NVIDIA covering foundation models, agents, fraud and anti-money laundering and real-time decisioning.
Does Santander US have its own AI?
Its disclosed in-house models are in auto lending. Santander Consumer USA assigns each applicant a proprietary custom score from FICO scores, debt-to-income and loan-to-value ratios and over 30 other factors, placing them in one of 100 pricing tiers, per the 2025 10-K, which does not call it AI. Machine-learning credit-risk models run on FICO Platform. No US-built generative AI system has been announced.
What AI has Santander released as open source?
Santander AI Lab published 11 projects on GitHub under the Apache 2.0 licence on June 25, 2026, including gen-fraud-graph for synthetic fraud networks, mech-gov-framework for language-model governance and mutatis-mutandis for discrimination testing, all without real customer data. It added Autoguardrails for guardrail tuning and, on September 23, 2026, the A2K protocol for agents querying governed knowledge.
How many AI agents does Santander run?
More than 280 process automation agents in production across credit, fraud, KYC and operations, per the group's June 22, 2026 update. In Brazil AI makes card fraud claims around 95% faster with up to 90% automation, and Openbank's AI models handle about 100,000 anti-money laundering alerts a year. Santander has not broken the agent count down by country.
- Santander builds AI capabilities that improve performance and lower costs using NVIDIA AI infrastructure
- Santander turns its AI-first strategy into measurable impact and extends AI access to all 185,000 employees
- When speed meets quality: How we make AI products measurable at Santander
- Santander publishes AI projects under an Open Source licence to ramp up shared innovation
- A2K: a new protocol suite for connecting AI agents to governed, auditable enterprise knowledge
- Santander Holdings USA, Inc. Annual Report on Form 10-K for 2025
- Santander US Auto Business Wins 2025 FICO Decisions Award for Machine Learning Credit Risk Analysis
- Engineering the loop: how Santander makes AI agents reliable
- Santander and the CSIC join forces to spearhead research into advanced and responsible AI for the future of banking
- Europe's approach to new AI: Vertical specialization, data sovereignty and regulatory trust (Santander AI Lab and Fundación General CSIC report)
- AI at Santander: From learning to practical use
- Designing conversational AI that knows when to say no
- Santander Holdings USA, Inc. Quarterly Report on Form 10-Q for the period ended 30 June 2026
- Santander's data & AI-first strategy accelerates through OpenAI collaboration
- 2Q 2026 financial results: Santander adds 12 million customers as underlying profit rises 15% to €7.3 billion in the first half
- US20230377033A1: Computer system and method for simulating loan payment structure on an interactive interface (Santander Consumer USA)
When Santander moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning