Banking AI: From Capability to Regulatory Permission

Banking innovation is primarily constrained by regulatory permission rather than by the technical capacity of AI models. This creates a unique tension where advanced AI capabilities must first navigate strict compliance requirements before deployment.

The intersection of AI capability and regulatory compliance has emerged as the primary focus for engineering efforts within financial institutions. Developers must balance cutting-edge AI features with the need for robust oversight.

Javier Solano emphasizes that banks operate within strict regulatory guardrails, which dictates the pace and scope of AI adoption. This environment necessitates a cautious, structured approach to integrating new technologies.

"that tension capability versus permission is where most of interesting engineering is right now"

AI agents fundamentally differ from standard software due to their non-deterministic nature, making traditional pass/fail testing methods inadequate. This unpredictability poses a significant challenge for financial institutions.

Financial institutions must implement rigorous governance and safety guardrails for AI deployments to mitigate risks associated with non-deterministic behavior. This ensures that AI systems operate within acceptable parameters and regulatory limits.

"testing um AI agents um becomes really challenging because these AI agents are non-deterministic. So very unlike your standard software testing, right?"

Actinver has successfully implemented AI agents to manage 'customer golden records,' going live with a system that identifies and resolves duplicate customer data. This initiative highlights a practical application of AI in banking operations.

The AI system significantly outperforms traditional regex-based methods for identifying duplicates, demonstrating enhanced accuracy and efficiency. This improvement directly contributes to data quality and operational streamlining.

The system's ability to identify duplicates improves efficiency by accelerating customer onboarding processes and ensuring a single, accurate view of each customer. This leads to better service and reduced operational costs.

"we recently went live with a system that is holding our customer golden record and we are using AI to identify duplicates"

A secondary AI-driven process performs daily checks on newly created records to identify and address any potential missed duplicates. This layer of verification adds an extra safeguard against data inaccuracies.

Errors can still occur even with human involvement, underscoring the necessity of automated double-checks. This combined approach minimizes human error and enhances data integrity.

Joint responsibility between AI agents and human operators is critical for effective error reduction. The AI acts as an intelligent assistant, flagging discrepancies for human review and ultimate decision-making.

"there's another process where it goes and checks everything that was created on that day and then double checks if it's if it's a unique client."

It is recommended to initially avoid deploying autonomous agents for tasks that directly involve money or customer interactions. These critical functions require a higher degree of control and auditability.

Reversibility is identified as the critical threshold for determining agent autonomy. Systems where actions can be easily undone are better candidates for initial agent deployment than those with irreversible consequences.

When possible, rule-based engines should be prioritized for tasks requiring high auditability. Their predictable nature allows for clear tracking and justification of decisions, which is essential in regulated environments.

"if something is going to touch money or touch a customer, I wouldn't do an agent, you know, the first the first part. I will what I would do is you know focus on a specific case that we would like to to to to solve"

Large Language Models (LLMs) are inherently unpredictable and can hallucinate rapidly, necessitating robust structural guardrails. These safeguards are crucial to prevent erroneous or unintended actions.

The focus should be on defining clear 'forbidden' boundaries for agents, outlining what they absolutely must not do. This approach helps control their behavior in complex banking environments.

Human supervision remains essential for at least the next six months as AI agents become more integrated into operations. This oversight period allows for continuous monitoring and refinement of agent behavior.

"I think one one of the main things you need to have in mind is that uh those engines can hallucinate and they can do it very quickly."

Testing AI agents requires moving beyond traditional pass/fail metrics. The unpredictable nature of AI necessitates a more nuanced approach to evaluation.

It is necessary to analyze the distribution of behavior across multiple runs to understand an agent's performance and identify potential anomalies. This statistical approach provides a more comprehensive view of its reliability.

"You can test you can test a system an agent as you test a system. You you have to pass from uh from a vision that does I pass the test pass that's usually as you test a system from a what's the distribution of the behavior around many runs"

Validation for AI agents is no longer a simple pass/fail mode; it must include comprehensive checks on tool-use accuracy and the agent's decision trajectory. Every action taken by an agent must be verifiable.

Every step an AI agent takes must be traceable to ensure auditability, particularly in a regulated industry like banking. This allows for detailed post-mortem analysis and compliance checks.

Agents require explicit stop behaviors to prevent uncontrolled operations and potential hallucinations. These built-in termination conditions act as a critical safety mechanism.

"It's not a pass fail mode now. We need to take the different approach. One of them is not only that we need to validate the output but it's also important for example, if some agent is doing some call to any other tool we need to be sure that the tool was called accurately."

A novel strategy involves 'agents testing other agents,' where more mature AI agents evaluate the performance of less mature ones. This peer-to-peer testing framework aims to improve reliability and robustness.

Mature agents can validate the performance and accuracy of newer or less developed agents, providing an automated and scalable method for quality assurance. This approach leverages AI capabilities for self-improvement.

Testing frameworks must account for loop detection and self-correction mechanisms within AI agents. This ensures that agents can identify and recover from iterative errors without external intervention.

"There's one strategy that we're looking into that is agents testing other agents. You have an agent that is more mature than the one you're putting in live and then testing the agent that is just not as mature as the other one."

The development of voice agents introduces complex variables, including understanding the tone of conversation, accents, and emotional nuances. This goes beyond simple text-based interactions.

Actinver is in an exploratory phase, focusing on how voice agents can understand customer frustration and find ways of diffusing tense situations. This involves advanced sentiment analysis and de-escalation strategies.

"We are in exploratory stage right now. The voice agent needs to understand the tone of the conversation. If the customer is frustrated, how to find ways of diffusing the situation."

Integrating voice agents with legacy Interactive Voice Response (IVR) and traditional banking systems presents a significant hurdle. This complex integration is far from a simple upgrade.

Unlike text-based interactions, voice interaction lacks the immediate context of a web page, introducing unique environmental challenges. Agents must infer context from speech and conversational flow.

Thorough testing for voice agents requires handling a diverse array of languages and regional dialect variations. This ensures broad applicability and accurate interpretation across different user bases.

The evolution from chatbots to voice agents is not as straightforward as it initially appears, demanding extensive development and integration efforts to overcome these inherent complexities.

"The voice agents are very complex. It sounds easy. It sounds like a well the evolution of a decision tree but it's not as easy because first of all you have uh legacy system integration the the IBR and the and the boy and the the traditional systems that what works current around uh an agent"

Industry regulations mandate strict risk control for financial data, making observability a non-negotiable requirement rather than an optional feature. It must be designed into the system from the outset.

Observability fundamentally consists of three core ingredients: robust detection mechanisms, comprehensive tracing capabilities, and reliable fallback protocols. These ensure that systems can identify issues, track their origins, and recover gracefully.

AI systems handling sensitive financial data must be designed with risk control from the very start. This proactive approach minimizes potential financial risks and ensures compliance with regulatory standards.

"It's not a nice to have. It's not optional. Observability and risk control has to be designed as long as you design the agent because you have to have risk control."

Development teams must define potential failure states before establishing success metrics for AI agents. This inverted design approach ensures that safeguards are prioritized from the initial stages.

Implementing robust fallback ladders is crucial to prevent agents from entering infinite loops or incurring uncontrolled expenditures. These mechanisms provide structured recovery paths.

Human-in-the-loop systems require structured data to intervene effectively. Clear, actionable data points allow human operators to understand agent decisions and step in when necessary.

Graceful decline is a critical capability for AI systems to protect corporate reputation. Systems must be designed to fail in a controlled and non-disruptive manner.

"Instead of you know in instead of wondering how the success look like with agents we need to start the opposite. What's the worst thing that the an agent can do?"

The quality of data directly determines the depth of hallucinations an AI agent may experience. Poor data hygiene leads to more frequent and severe errors.

Legacy systems often serve distinct purposes and are not inherently designed to feed AI directly, necessitating careful integration strategies. Directly feeding raw legacy data can introduce inconsistencies.

Dedicated integration layers are critical for managing data flow between legacy systems and AI agents. These layers ensure data is properly curated and formatted for AI consumption.

Proper data placement is essential for system reliability, as data context and accessibility significantly impact agent performance and accuracy. Data needs to be where the AI expects it, in the format it expects it.

"I think one of the main you know elephant in the room is data when coming to agents... if your data is not properly curated taken care of, your hallucinations are going to get like deeper and deeper."

Traditional pass/fail quality assurance (QA) methods are insufficient for evaluating AI agents due to their non-deterministic nature. A more dynamic and context-aware approach is required.

AI agents should be evaluated by other agents using specific, predefined criteria, shifting the focus from binary outcomes to nuanced performance metrics. This allows for automated and continuous assessment.

Reliability must be measured across multiple iterations and scenarios to accurately gauge an agent's robustness and consistency. Single-instance testing fails to capture the full spectrum of behavior.

Golden sets and shadow mode deployments are key strategies for rollouts. Golden sets provide benchmarks for performance, while shadow mode allows for real-world testing without direct customer impact.

"You cannot test agents in a traditional QA way, right? So you have to be able to test agents number one with another agent, number two creating not measure them in pass or fail but based on valuations."

Prompt injection and jailbreaking remain critical threats to AI agents, underscoring the need for advanced security measures beyond typical functional testing. Attackers can manipulate agents through malicious inputs.

Security testing for AI agents must extend beyond basic functionality to cover resilience against various attack vectors. This ensures the agent can withstand attempts to compromise its integrity or purpose.

Standardized tools like API managers and MCP servers should be utilized to handle protection layers across all agents. This provides a consistent and robust security infrastructure.

Security layers should be applied globally rather than on a per-agent basis to ensure comprehensive protection. This prevents gaps in security that could be exploited by sophisticated attacks.

"You need to make sure that you're not only testing functional matters in the agent but also how this agent can be cracked and how efficient the trajectory is, how robust the agent is, how aligned and how resilient it is to attacks."