Banks are examining how to scale AI at lower cost while retaining control over its use. Crédit Agricole expanded its employee AI programme into agents connected to internal systems, while Santander and bunq reported comparisons showing how specialised and open-weight models could reduce costs while maintaining performance.

Operational resilience and oversight remained central to those ambitions. Google Cloud’s London outage exposed shared infrastructure dependencies and drew regulatory scrutiny, while State Bank of India investigated suspected customer-data exposure. The UK Financial Conduct Authority illustrated why human oversight of AI requires independent challenge and clear authority to intervene.

Read more on the week’s key developments:

1. Crédit Agricole extends employee AI into agents connected to bank systems

Crédit Agricole announced a Mistral partnership and the rollout of Sécur’IA on 20 October. Its new Crédit Agricole Artificial Intelligence unit is working with Prisme.AI to let business teams create, use and share agents connected to internal systems. The Mistral partnership covers a shared AI platform intended to serve all group customers and employees, alongside planned model adaptation using bank data.

Sécuri’Chat, introduced at Crédit Agricole S.A. in 2024, was already used regularly by nearly nine in ten trained employees. Sécur’IA takes this further by allowing business teams to develop agents connected to bank systems, supported by shared models and centrally coordinated infrastructure. Computing will be allocated across the group’s private cloud, European partners and other providers according to the sensitivity and criticality of each workload. The announcement did not specify which entities would receive Sécur’IA first or the rollout timetable.

2. Google Cloud outage exposes shared dependencies in banking recovery

A power failure in Google Cloud’s London region on 20 October disrupted services at Revolut, Lloyds and HSBC, including Revolut’s card payments and cash withdrawals. Google’s preliminary report, published on 22 October, said primary and backup generator controllers failed, exhausting battery reserves and taking one availability zone offline. Google recorded six hours and ten minutes of principal cloud disruption; individual applications recovered at different times.

The impact spread as redirected traffic overwhelmed capacity in neighbouring zones and some regional services continued to depend on components in the failed zone. Reuters reported that the European Banking Authority had contacted Google to assess the implications for EU financial institutions, while the UK Financial Conduct Authority was also engaging with the provider. For banks, the outage highlighted the need to test whether backup infrastructure can handle additional workloads and function when another zone is unavailable.

3. bunq tests open models against live support conversations

bunq’s Finn assistant was running open-weight models on Nebius infrastructure, the provider disclosed on 21 October. Benjamin Kleppe, bunq’s Head of AI innovation, described a week-long live comparison using Kimi K3. He reported quality within 1% of the best proprietary model at around 60% of its cost per answer. The models ran on dedicated Nvidia graphics processors in the EU.

Kleppe said bunq first evaluated more than 20 models against roughly 1,000 real support cases in multiple languages, using three AI judges to score answers. The selection was therefore informed by the bank’s own service workload. A January technical account, co-authored by Kleppe, had described Finn using Anthropic’s Claude through Amazon Bedrock. The October disclosure adds a model and hosting option beneath an established assistant.

4. Barclays links planned role reductions to technology deployment

Barclays said on 22 October that it was reducing around 5,000 full-time roles across the group during 2026, with about 1,000 already removed and the remainder expected by year-end. Group Finance Director Anna Cross said about 60% of the reductions concerned support functions, including fewer developers as AI raised productivity. She put the group’s 2027 gross efficiency-savings target above GBP 1.1 billion ($1.5 billion), compared with around GBP 0.7 billion ($0.9 billion) annually in the preceding two years. Efficiencies would offset inflation and business growth, keeping run-the-bank costs flat or lower in 2027 and creating capacity for GBP 0.4 billion ($0.5 billion) of additional investment.

Within Barclays UK, improved technology had reduced call-centre volumes by 10% over the past year, while branch staffing changes cut employees per branch by around 10%. However, accelerated restructuring meant the UK business no longer expected costs to decline in 2026, illustrating the upfront expense of securing later efficiencies.

5. SBI discloses customer records found on the dark web

State Bank of India (SBI) disclosed on 21 October that preliminary analysis of approximately 375,000 dark-web records had identified 3,897 customer records containing personal information. Reported fields included contact details, branch and customer identifiers and loan-related metadata. The bank activated incident-response and containment procedures and appointed an independent cybersecurity agency to investigate the exposure.

SBI received alerts from its threat-intelligence provider, the Defence Cyber Agency and CERT-In on 16 October. The bank said core operations continued normally, but further investigation was needed to establish the data’s authenticity and completeness, trace its source and assess the risks to customers. On 23 October, SBI said it had individually notified customers whose data was identified and reiterated that account balances and funds remained safe.

6. FCA shows how independent human review can test AI decisions

The UK Financial Conduct Authority (FCA) published examples of good and poor AI practice on 19 October, drawing on supervision, industry engagement and testing programmes. The publication addressed consumer outcomes, accountability, human oversight and supplier dependencies under existing requirements, including the Consumer Duty and the Senior Managers and Certification Regime. It introduced no new rules.

In one affordability-assessment case, staff made decisions while the AI system was being established, without seeing its proposed outcome. This preserved an independent benchmark for comparing human and machine judgements during validation. At another firm, reviewers increasingly accepted recommendations without sufficient scrutiny, and unclear escalation criteria allowed uncertain outputs to proceed. Effective human review requires an independent basis for challenge and clear authority to intervene; an approval step alone establishes neither.

7. Santander reports lower serving costs from specialised AI models

Santander reported on 19 October that selected experimental banking applications had reduced model-serving costs from around EUR 5 ($5.6) to approximately EUR 0.03 ($0.03) per interaction while maintaining performance. Its AI Lab developed Model Arena to evaluate open-weight models against banking tasks and was moving towards smaller, specialised models. The bank also agreed a three-year collaboration framework with Nvidia.

Model Arena addresses which model is adequate for a banking task. Santander separately used Nvidia’s Dynamo-Triton software to improve how models run within its own architecture, explicitly linking that approach to the serving-cost reduction. The comparison therefore reflects both model selection and the computing used to produce responses. It measures one component of an AI service’s economics. Simulation and adversarial testing formed part of the lab’s assessment before wider deployment.

8. FABMISR adds on-premises AI agents to its automation programme

First Abu Dhabi Bank Misr (FABMISR) announced on 18 October that it had implemented an on-premises AI platform using IBM watsonx. The release, published by IBM the following day, described two production-ready applications: an agent that retrieves internal knowledge to support answers and a Financial Analyst agent designed to support financial analysis. The platform keeps AI workloads and data within the bank-controlled environment.

FABMISR’s earlier IBM programme, announced in February 2024, used Cloud Pak software on Red Hat OpenShift to connect enterprise applications, manage their interfaces and automate processes. The watsonx platform adds tools for staff to retrieve and interpret internal information, extending the programme from system connectivity into the preparation of answers and financial analysis. Chief Technology Officer Shaarawy Mohamed said the aim was to embed AI in the bank’s operations and management decision-making.

9. Maybank expands regional technology monitoring and response

Maybank agreed a five-year programme with ServiceNow to expand its Resilience Operations Centre across Malaysia, Singapore and Indonesia, according to a joint release published on 19 October. The platform will combine operational data with a common view of applications, infrastructure, technology assets and service dependencies, supported by analytics and automated workflows.

The planned design links technical events to the banking services that depend on affected systems, helping operations teams prioritise incidents by service impact. Rollout will begin in Malaysia and Singapore before expanding to Indonesia. The agreement supports Maybank’s broader technology modernisation agenda, for which it has earmarked MYR 10 billion ($2.48 billion) through 2030. An internal centre of excellence, supported by ServiceNow University, will build expertise in platform engineering, automation and AI.

10. Borgo prepares to take over platform services from Bank of Åland

Bank of Åland and Swedish mortgage provider Borgo agreed on 19 October to phase out their platform-services arrangement ahead of schedule. Migration of data, processes and operational functions will begin in the fourth quarter of 2026 and continue through 2027. Bank of Åland expected a one-off operating-profit benefit of approximately EUR 8 million ($9.0 million) in the fourth quarter, followed by the gradual loss of recurring platform income.

Borgo gained the capacity to bring these services in-house through its acquisition of Hypoteket’s technology platform, announced in August 2025. That agreement included a separate option to acquire Hypoteket’s mortgage portfolio, distribution and brand. Chief Executive Pehr Olofsson described the transition as a move towards a more integrated operating model. Bank of Åland will remain a shareholder and mortgage distributor, retaining its commercial relationship with Borgo as its technology and mortgage administration fees gradually cease.