A Specialized Generative AI Model for the Banking Industry

Organisations involved

Main Participant: FairMind Srl develops an Agentic Platform      that support software-development teams by improving quality, productivity and governance. Credit Agricole Group Solutions provides domain expertise, validating datasets and evaluating banking-specific use cases. 

The challenge

The banking sector must adopt AI responsibly while protecting sensitive customer data and meeting strict GDPR and compliance rules. Public access general-purpose frontier AI models cannot be used, as they risk exposing confidential information, and lack the domain knowledge required in regulated software-development environments. Financial institutions also need AI that can work within established internal processes, while safeguarding data sovereignty.

FairMind identified that generic language models, openweights and with a low number of parameters do not understand effectively banking-specific compliance terminology, coding standards or risk controls. To address this, the company set out to build BankGPT, a custom model      post trained on banking data and tailored to tasks in software modernization, development and operations. The key challenges included choosing a suitable open     weight      base model, gathering high-quality training data, and running experiments at scale while remaining cost-effective.

Training banking-specific models required far more compute than FairMind could access internally. HPC resources enabled the company to train models up to 32B parameters, explore multiple architectures, run controlled bias tests, and iterate rapidly. Utilizing HPC resources allowed FairMind to optimize BankGPT for on-premise use.

The Solution

FairMind developed BankGPT by training specialized open     weight      models that deploy 4-32 billion parameters. The development followed a full lifecycle of data preparation involving:

•    Curated banking-specific datasets with version control through Hugging Face collections, ensuring full traceability;

•    Model tuning utilizing optimized balancing throughput and memory management;

•    Model evaluation that incorporated systematic bias assessment using established benchmarks, plus a custom banking Q&A benchmark validated by Credit Agricole experts.

     BankGPT serves as the intelligent core of one of our agents, acting as its "brain" dedicated to processing banking-specific content. Designed as a compact language model, BankGPT can be seamlessly integrated within customer IT environments, ensuring data sovereignty and high protection standards that set it apart from cloud-based alternatives. Thanks to this integration, the agent can efficiently support other FairMind agents in all key activities in the banking software development lifecycle - from ideation, requirement drafting, and code generation to validation - while handling information and regulatory requirements unique to the financial context.

Impact 

BankGPT shows that domain-specific AI can meet banking security and GDPR standards while delivering strong technical performance. FairMind’s evaluation framework gives financial institutions confidence to adopt open-source AI, while on-premise deployment aligns with data-sovereignty rules and reduces legal and operational risks. This positions FairMind to enter highly regulated markets, including banking and insurance. The company expects to achieve the following goals:

• Year 1: 3–5 pilots (€180k–600k)
• Year 2: 10–15 clients (€600k–1.8m)
• Year 3: 20–30 clients (€1.2m–3.6m)

Finally, BankGPT helps professionals focus on higher-value tasks while the model automates compliance-driven coding and documentation work. It also promotes transparent, open     source AI adoption, reducing dependence on proprietary cloud platforms.

Benefits

• Successfully post-trained open-weight models of up to 32B parameters on the Leonardo supercomputer.
• Created a custom banking Q&A benchmark validated by domain experts.
• Achieved full data traceability through structured dataset versioning.
• Built a robust checkpointing and training workflow for large-scale experimentation.