Armundia Group CIO: How Banks Can Secure Customer Data While Using External LLMs

Banking institutions are navigating a critical conflict between the productivity gains of generative AI and the stringent data privacy requirements of regulated environments. As legislative frameworks like the EU AI Act and DORA loom, the industry is shifting focus from contractual promises to architectural safeguards that prevent sensitive data from ever leaving internal servers.

What was announced

The industry is moving toward a "Hybrid Symbolic-Generative RAG" (Retrieval-Augmented Generation) architecture to bridge the gap between external Large Language Models (LLMs) and internal data security. This approach addresses the risks associated with sending regulated data—including personally identifiable information (PII) and proprietary financial documents—to commercial models. While contractual protections like private tenants exist, they do not prevent data from physically leaving the institution’s perimeter, leaving a legal instrument as the only protection.

The proposed framework utilizes an anonymization layer that sits between the bank’s application and the external model. This system identifies sensitive entities and replaces them with neutral tokens, such as specific identifiers for persons or IBANs, allowing the model to reason on the structure of the information without accessing real values. Once the processed information returns to the institution’s "zone of trust," the real values are restored based on user permissions. This methodology addresses specific failure points in standard PII filters, which often fail to protect business-critical secrets like target company names or valuations in closed-ended fund documents.

The architecture relies on a sophisticated combination of five core components:

• A deterministic engine for structured identifiers to provide certainty rather than probability.
• A statistical named-entity recognition (NER) engine for free-text prose.
• A domain engine for business entities that off-the-shelf models do not recognize.
• Symbolic guards to control what information is permitted to leave the environment.
• An in-house medium language model (MLM) acting as the sole gateway to the outside.

This development is timed against a tightening regulatory calendar, including the EU AI Act (August 2026) and the UK’s AI Regulation Bill, alongside existing mandates like GDPR and the Digital Operational Resilience Act (DORA).

The companies involved

Armundia Group is a fintech organization focused on the digital evolution of the financial services sector. The company operates in an environment where the "first-mover" advantage of generative AI is being balanced against the caution required for regulatory compliance. As financial institutions move beyond the initial experimentation phase with external LLMs, the market for "harness" technologies—systems that wrap around commercial models to provide a zone of trust—is becoming a distinct sub-sector of the fintech landscape.

The group’s focus on the intersection of artificial intelligence and regulatory compliance reflects a broader market trend where financial institutions are seeking specialized partners to navigate the complexities of data sovereignty. In the context of private equity and closed-ended funds, the need for such specialized oversight is acute. These sectors handle highly confidential documents, such as offering memoranda and capital account statements, which contain sensitive business intelligence that generic security tools might overlook. By developing frameworks that understand domain-specific secrets, technology providers like Armundia Group are attempting to solve the "black box" problem of external AI integration for high-stakes financial entities.

What this means

The move toward Hybrid Symbolic-Generative RAG suggests that the "all-in" approach to cloud-based LLMs is hitting a regulatory ceiling in banking. The industry is effectively building a "firewall for meaning," where the external model is treated as a utility for logic but is denied access to the underlying truth of the data. This puts significant pressure on "off-the-shelf" AI providers to prove they can operate within these symbolic guards without losing contextual utility. The real challenge for the sector will be maintaining the "intelligence" of the response when the external model is only seeing tokens rather than raw data. For banks, the competitive edge may no longer lie in which model they use, but in the sophistication of the deterministic engines they use to mask their proprietary knowledge.