
Banks have been talking about AI for years, but many deployments never make it past the pilot stage. What’s causing these projects to stall?
There are several reasons. First, there are numerous dependent systems around any single deployment. You might have powered one of the systems with AI, but without thinking about the ecosystem around it, that single transformed piece would not sustain itself. For that matter, it might never even reach the stage where it’s available to a wider internal audience.
Second is guardrails. Because banking is so heavily regulated, no off-the-shelf model is a perfect fit. There have to be guardrail layers sitting on top of the model to keep the data secure and the process compliant.
You’ve spent 15+ years working inside legacy banking environments. What makes integrating AI into these systems so much harder than vendors tend to suggest?
The challenge is not that AI integration is inherently more difficult. The real question is whether we have clearly identified the problem we are trying to solve with AI, especially within the constraints of legacy banking systems and processes.
I recall working with a financial institution where we were trying to streamline credit card onboarding and decision-making using machine learning. Interestingly, the biggest challenge was not the AI integration itself, but the complex internal processes surrounding credit card onboarding.
While AI models could automate much of the onboarding process, including OCR-based document verification and credit card eligibility decisions, the existing workflow still required a manual KYC review after the automated checks were completed. This created an unnecessary layer of manual intervention, potentially duplicating checks that the system had already performed.
The real challenge was figuring out how to eliminate this redundant step without disrupting the broader transactional flow. More importantly, we had to convince stakeholders that removing the manual intervention would not compromise security, compliance, or risk controls.
To address this, we mapped out both the existing and proposed workflows, highlighting where AI could replace manual activities, how the necessary controls would be maintained, and why the new process could be more efficient, secure, and less prone to human error.
The key lesson was that successful AI adoption is not simply about integrating intelligent models into existing systems. It is about rethinking the end-to-end business process, challenging legacy practices, and building stakeholder confidence in a new way of working.
The difficulty is not legacy banking itself but the fact that the entire ecosystem is so tightly coupled. Without a 360-degree view of the system and the tasks the AI will perform, you might end up creating more problems than you solve.
These systems have worked together reliably for years, and a slight change in one system, without considering its effect on another, can disrupt everything downstream.
Prompt engineering is often treated as an afterthought in enterprise AI. How does prompt design affect accuracy and compliance in customer-facing financial applications?
Prompts with context are the backbone of any AI integration, irrespective of the domain you are integrating with. While prompts let the model know the task and persona, the context provided to the prompt and the system on which the tasks are intended to be performed are just as important.
Consider a system where a customer asks the AI agent to “change my address, I have moved.”
If the agent lacks the context to ask for the old address, take the customer through the verification stage, ask for the new address, and follow the required KYC process under banking regulations, the prompt and the agent are of no use.
The agent should not simply interpret the request and start changing the customer’s address.
Fraud detection and risk assessment are two of the most cited AI use cases in banking. Where are these tools genuinely outperforming traditional approaches, and where are results falling short of expectations?
Fraud detection and risk assessment were among the first areas in banking where AI took hold at enterprise scale, and both have performed well.
Fraud detection has always been a pain point in the banking and finance industry, and many existing systems rely on rule-based detection. However, those rules had to be set up manually, and the systems were not necessarily learning from the anomalies identified or the patterns being established.
This is where an AI layer has helped, particularly in identifying patterns and anomalies that traditional rule-based systems might miss.
Where things fall short is when these systems generate too many false positives, or when banks cannot clearly explain why a particular transaction was flagged.
In my experience working on digital banking and risk-related implementations, this becomes particularly important when an AI model flags a legitimate customer activity as suspicious. For example, a customer making an unusually large transaction or accessing banking services from a new location might trigger a fraud alert, even when the activity is genuine. The challenge is not just improving the model’s accuracy, but ensuring that the bank understands why the transaction was flagged and can distinguish genuine risks from normal customer behavior. One practical approach is to combine model-generated risk scores with explainable decision factors, transaction history, and human review for borderline cases. This helps reduce unnecessary customer friction while maintaining appropriate fraud controls.
In a regulated industry, that explainability gap is a real problem, which is why these systems must be implemented within a controlled environment.
Many banks face a choice between modernizing core systems and layering new technology on top of what they have. How should institutions think about that tradeoff when live operations can’t be disrupted?
Modernizing is the right choice. .Layering new technology on existing systems only buys time before modernization becomes unavoidable.
Banks offer many products, some are customer-facing and exposed to the external world, while others are internal systems used by the bank for day-to-day banking operations.
Before exposing any new system to external users, the better path is to modernize an internal system first, test it thoroughly, and then gradually cascade the effects of those changes to other systems. Another benefit is, since the system is internal and with limited user load, you may keep running both the systems together for a while, which helps in data reconciliation and helps prevent business discontinuity.
What does responsible AI implementation look like for financial institutions, and where should the line sit between automation and human judgment in banking relationships?
Put simply: Responsible AI = Human-in-the-loop AI.
Even when AI can perform most of the tasks in a process, there is always a need for human oversight at different stages, and not just at the end of the process.
If you are modernizing a four-step process and put a human review only at the end of the fourth step, that human would have very little room to intervene or correct things, given the complexity with which AI agents or tools arrive at their results.
Guardrails, along with human review at the right stages, are the answer to responsible AI integration in the banking and finance industry.

Ricardo Tomaz is VP of Finance & Operations at Ztudium, the 4IR technology company behind Businessabc, Citiesabc, and IntelligentHQ. He owns the group's finance and operations function, spanning financial strategy, budgeting, reporting, and multi-entity operations, and works directly with leadership to align commercial goals with data-driven insight.
Working at the core of a technology group operating across AI and blockchain, Ricardo writes on the practical intersection of finance and emerging technology: how automation and digital transformation are reshaping business operations, what data-driven decision-making looks like in practice, and the questions of governance and ethics that come with AI-led systems.
