The "Hidden Barrier" to AI in Financial Services - Data

It seems like every bank and financial institution (FI) is racing to implement AI. From enhancing fraud detection to personalizing customer experiences, the promised benefits are huge. But as someone who is involved in helping banks and FIs make the best use of technology, I'm seeing a disturbing trend: firms are jumping into the deep end without knowing how to swim.

A recent MIT report, "The GenAI Divide: State of AI in Business 2025," is an eye-opener. It found that a staggering 95% of enterprise Gen AI projects are failing to deliver a measurable return on investment. While the study points to a fundamental "learning gap" where generic AI models can't adapt to complex, brittle enterprise workflows, my own experience from several client conversations and projects tells me the primary culprit is often much simpler.

The result? Projects are stalling or outright failing, not because the AI isn't smart enough, but because the underlying data is a mess (or sort of). It's like trying to build a skyscraper on a foundation of sand. It just won't work.

The Hard Truth: Data Is Not a by-product, It's the Main Event

AI models are only as good as the data they're trained on. In the financial sector, where data is often fragmented across legacy systems, business units, and geographical silos, this is a massive problem. You've got inconsistent formats, poor quality, and a lack of clear governance.

Think of it this way: a fraud detection model trained on a small, biased dataset might get really good at flagging transactions from a specific zip code but miss sophisticated schemes from elsewhere. Or, a customer service chatbot that's fed incomplete client data will give generic, unhelpful responses, frustrating the very people it's supposed to help.

Real-World Examples of AI Gone Wrong

These aren't just theoretical problems; they've caused significant headaches for major players. 

• The Apple Card Debacle: In 2019, the Apple Card, issued by Goldman Sachs, faced accusations of gender bias. The AI model's credit limit decisions seemed to favor men over women, even with similar financial backgrounds. The bank maintained that gender was not a factor, but the model's outcomes suggested it had learned and amplified historical biases present in the training data. This led to a regulatory investigation and a major public relations nightmare.
  

• Chime's Overzealous Fraud Detection: Neobank Chime's AI-powered fraud detection system was so aggressive that it mistakenly flagged and froze thousands of legitimate accounts. Customers were locked out of their funds, and the bank lacked a quick human-review process to resolve the issues. The result was widespread customer frustration and reputational damage.
  

• Commonwealth Bank's AML Failure: The Commonwealth Bank of Australia was hit with a record $700 million fine after its AI-driven anti-money laundering (AML) system failed to flag over 53,000 suspicious cash deposits. A technical glitch prevented the system from reporting transactions, a failure that went undetected for years, allowing criminals to launder millions.
  

These aren't just isolated incidents. They are cautionary tales about what happens when you treat AI as a quick fix rather than a long-term strategic investment built on a foundation of clean, reliable data.

How to Get Started Right: A "Data-First" AI Strategy

So, what's the solution? You can't just throw more money at the problem. Banks and FIs need a fundamental shift in mindset, starting with a robust, "AI-ready" data infrastructure.

1. Conduct a Data Readiness Assessment

Before you even think about an AI project, you need to understand where you are. A Data Readiness Assessment is the first step. It's a structured review of your organization's data landscape to determine if it's suitable for AI initiatives. It helps you identify gaps and prioritize which foundational elements need to be addressed first. You can be at a "nascent" stage with siloed spreadsheets or a more "managed" stage with some centralized data warehouses. Understanding this is step one.

2. Clean House and Centralize

You can't train an AI on dirty data. Invest in a dedicated effort to cleanse, standardize, and centralize your data. This means breaking down those data silos and bringing everything into a unified platform - be it a data lake, data warehouse, or a modern data fabric. This process is time-consuming but non-negotiable.

3. Implement Strong Data Governance

Data governance isn't just a buzzword; it's the rulebook for your data. Establish clear policies for data ownership, stewardship, access control, quality, security, and lineage. This is especially critical in the highly regulated financial industry. Robust governance ensures your data is not only clean but also compliant and trustworthy.

4. Think Small, Scale Smart

Instead of a massive, organization-wide AI project, start with a pilot. Choose a high-impact, well-defined use case where you know you have high-quality, accessible data. This could be something like using AI to automate a specific back-office process or to enhance a single fraud-detection rule. Once you achieve a quick win and prove the value, you can gradually expand your initiatives.

The AI revolution in banking is real, and the potential is immense. But success won't come from a flashy new model; it will come from the hard work of building a solid, data-first foundation. By taking a strategic, incremental approach and focusing on data readiness, banks and FIs can move from simply joining the bandwagon to actually leading the race.

Want to know more? Let's discuss.

About the Author

Sameer Goyal

is a technology leader specializing in banking technology, cloud solutions, data engineering, and AI/ML. He currently serves as Senior Director and Global Head of Banking Technology at Acuity Knowledge Partners, leading engineering and technology initiatives for financial services clients globally.

He has previously worked with Moody's, Sapient, and Birlasoft, with a focus on AI, cloud transformation, and scalable technology solutions.