The Evolution of AI in Banking: Key Insights from Industry Experts

In a rapidly evolving regulatory and technological landscape, AI is no longer just a buzzword in banking—it’s a business imperative. But what does real adoption look like in an industry known for risk sensitivity and regulatory oversight?

That was the focus of H2O.ai’s recent webinar, "Turning AI Strategy into Results", where two industry leaders - Chris Sivalingam (ex-Scotiabank, ex-CBA) and Mark Landry (Kaggle Grandmaster, Director of Data Science at H2O.ai) shared their experiences from the front lines of AI adoption in banking.. Together, they explored the biggest shifts in enterprise AI, offered practical use cases in banking, and even tackled what it takes to operationalize GenAI and Agentic AI in a secure, compliant way.

Here are the key takeaways:

The AI Journey in Banking: From Skepticism to Adoption

The banking industry has undergone a significant transformation in its approach to AI over the past several years. As Sivalingam noted, back in 2016-2017 when AI was gaining momentum, financial institutions were initially skeptical about implementing artificial intelligence components for making recommendations, decisions, and processing data.

“It's a very risk-averse environment. But I think most of the banks have evolved over time, the MRM (Model Risk Management) function has evolved as well. Regulators have become more able to get the transparency from AI that they are more comfortable with banks moving into the AI space.”

Chris Sivalingam

The journey has progressed from traditional predictive modeling to generative AI, and now to agentic AI, where autonomous components can perform work with minimal human intervention. The next evolution will bring "agent builders" or "driverless agents" where agents can create other agents tailored to specific use cases.

The Importance of Benchmarks: H2O's Achievement on the GAIA Leaderboard

With numerous Gen AI platforms flooding the market, benchmarks have become crucial for evaluating performance. Landry explained that the General AI Assistants Benchmark (GAIA) is designed to mimic complex, multi-step tasks that people perform in workplace environments.

H2O.ai achieved the top global score on the GAIA benchmark in March, a milestone that reflected not just performance, but real alignment with enterprise needs.

“The GAIA benchmark gave us a lot of good data to go after, seeing these multi-step processes handled that are similar to some of the complex tasks that we get at work.”

Mark Landry

H2O.ai

Some of the Top AI Use Cases in Banking

Sivalingam outlined three key use cases emerging across financial institutions:

1. Customer Onboarding and KYC

The Know Your Customer (KYC) process involves numerous steps: receiving documentation, identifying regulatory requirements, verifying documents, extracting information, updating systems, and conducting ongoing due diligence.

"This whole process, from onboarding and then the ongoing lifecycle of customer verification, is a very laborious, human-intensive process," Sivalingam explained.

Banks are investing in Agentic AI to automate document processing, verify customer information, and update internal systems. These agents can also initiate due diligence workflows and assess customer risk levels.

2. Regulatory Compliance and Policy Management

For banks, staying on top of shifting regulations and updating internal policies to match is a constant, resource-heavy task.

"This whole coverage against the regulatory requirements is an agent's typical area where the bank's human staff can monitor how the agents are performing and provide necessary inputs," Sivalingam noted.

3. Collections and Call Center Operations

The collections process requires significant energy to follow up with customers, request information, gather evidence, and schedule payments.

"There's a lot of banks, especially in Australia and North America, looking at or even using agents to get this whole process automated," Sivalingam shared.

H2O's Approach to Document AI and KYC

Landry demonstrated how H2O's Gen AI and predictive models work together in KYC processes, with a platform that can classify documents, extract relevant information, and validate it against existing customer data.

"What's exciting is that as we've seen generative AI take off, it's getting strong enough to enable these agentic workflows," Landry explained. "This autonomous nature can be possible, whereas previously you often had a heavy use of humans in the loop."

The system can process documents from start to finish, identifying what's on each page, running necessary extractions, and self-checking the results before returning data in a format ready for database integration.

How Banks Are Implementing Agentic AI

Financial institutions are being approached by numerous technology companies claiming to have agentic platforms. Sivalingam emphasized that banks are focused on leveraging existing investments rather than purchasing new tools.

"Leveraging your existing investments is a key driver or requirement for most of the banks," he shared.

Banks are also working on bringing together various data sources:

• Internal structured and unstructured data
• External third-party data for fraud prevention and AML
• Existing AI investments

The goal is to build an agentic AI framework that can automate human-intensive use cases, allowing staff to focus on more productive work.

Sovereign AI: Meeting Regulatory Requirements

Regulatory requirements vary by geography, making data sovereignty a critical concern for financial institutions. Landry highlighted how banks must navigate different regulations like SR 11-7 in the US, which may not apply in other regions like Australia.

"Every bank is working this out in their own way," Landry observed. "or every project, it's not a one-size-fits-all answer."

H2O.ai defines Sovereign AI, as the ability for an enterprise, government, or nation to build, deploy, and operate AI using its own data, infrastructure, and governance frameworks. This approach helps banks meet regulatory requirements by keeping data within geographic boundaries and giving them full control over the entire stack—not just the models or data.

Sivalingam added that regulators themselves have undergone a significant shift in understanding AI: "The last two years, [regulators] have really gone through a major shift in understanding the power of AI, agentic AI, and generative AI."

Advice for Banking Executives

Sivalingam noted a major shift in how AI is being embraced across the organization. While CDOs and CDAOs initially led the charge, operational leaders are now showing strong enthusiasm for the technology.

"The guys who run day-to-day banking, credit cards, fraud, and AML—these non-techie folks—are very enthusiastic about AI and Gen AI and predictive AI and agentic AI," Sivalingam observed.

His advice for executives:

• Have a plan and strategy rather than allowing different departments to adopt various tools independently.
• Work with data offices to build a proper AI platform for the entire organization.
• Leverage assets effectively and cohesively across the organization.
• Ensure predictive models (like those for fraud detection) are integrated with agentic components

Future Trends in AI for Financial Institutions

Landry emphasized that AI models continue to improve at a surprising rate. While progress may seem incremental, the pace of innovation remains steady.

"A task you may have not been able to do last time you tried it, especially if you tried a couple of years ago—now we've got AI working, but it still didn't work well. As these get better, they're going to start to unlock more and more use cases," Landry explained.

Additional Insights for Financial Institutions

In closing, both experts emphasized practical approaches to AI implementation:

• Start small and tangible: Focus on low-hanging use cases rather than ambitious science projects
• Measure results: Choose use cases where you can clearly measure outcomes and ROI
• Be cautious with data: Production data often differs from test data, so be prepared for edge cases
• Take a platform approach: Build a comprehensive AI platform rather than adopting piecemeal solutions
• Focus on explainability: Ensure agents can explain their decisions and demonstrate value

As Sivalingam succinctly put it: "Keep it small, keep it tangible, and stay away from science projects. Try to make it work and put it into production to see whether you can drive business value."

Final Thought: AI Success Is a Journey

The financial services industry is rapidly learning how to combine the promise of GenAI with the discipline of data science, compliance, and measurable ROI. The future? Agentic workflows that scale. Secure models that perform. And an enterprise-wide strategy that doesn’t just experiment—but delivers.

This blog post summarizes key insights from a webinar featuring AI and banking experts from H2O.ai. For more information on implementing AI solutions in your financial institution, contact our team. You can watch the full recording here.