A duplicate customer record. A missing zip code. A workaround someone did years ago. On their own, these don’t seem like big issues. But together, they create inconsistencies that have a ripple effect – often with major consequences.

Here’s a truth that we can all agree on: If your data isn’t clean, your AI won’t be reliable.

In the spirit of the “of course we are” trend, I used some humor to shed light on this serious financial fraud issue in a recent episode of Brewing Curiosity. The response was telling: This isn’t an isolated problem. It’s something compliance and risk teams deal with every day.

When small data issues create real blind spots

The skit was about how a single customer’s name appeared in three different ways in three different reports. It was funny in the video, but not so funny in real life because it reflects a real challenge.

When a name is entered slightly differently, or a field gets truncated, you don’t just get a typo. You create what is called a phantom identity – a duplicate version of the same person that splits their history and changes how risk is calculated. Phantom identity, often called a synthetic identity in financial crimes, is anticipated to grow exponentially with AI, according to the Institute for Financial Integrity.

One record might suggest low risk. Another might trigger a Suspicious Activity Report (SAR). Same person, different outcome depending on which version of the data your systems pick up. It’s a common issue across financial institutions of all sizes.

The hidden cost of cleaning data just to operate

A lot of time goes to fixing data issues just to keep things moving. Teams spend hours reconciling records, correcting fields and rectifying inconsistencies. That work is necessary – but it doesn’t move the organization forward. It limits the time available for higher-value analysis.

Compounding the data troubles is legacy data foundations. But what if you use modern AI on top of that same data foundation? Instead of slowly exposing issues, AI accelerates them. Models don’t quietly degrade – they make the wrong call at scale, and often without obvious signals that something is off.

More data doesn’t solve the problem

There’s a common assumption that more data leads to better AI. In practice, that only holds if the data is consistent and trustworthy.

From a data standpoint, garbage in, garbage out – and potentially fraud out. If your data includes duplicates, inconsistent formats and missing fields, adding more of it won’t improve outcomes. Add AI to bad data and the chances that errors show up more often – and in more places – just got bigger. It can affect multiple systems, decisions and customer interactions. This can cause your AI to hallucinate.

And from a regulatory perspective, “dirty data” isn’t just a technical issue or operational headache; it’s what regulators call a “finding”.

The need for data access, data quality and data governance will continue to intensify. In highly regulated industries like banking, with strict compliance requirements, AI models require trusted data to help organizations avoid risk, costs and lost productivity.

Preparing for AI agent-driven environments

As if the stakes weren’t already high, they’re about to get even higher with agentic AI. Autonomous decision-making, fraud detection and customer service improvements offer significant advantages, but organizations must implement them carefully. If that data is inconsistent, their decisions will be too. Basically, you’ll be automating a problem, creating a much worse one.

Another consideration is what happens when AI agents talk to each other or bad actors create synthetic data to power autonomous fraud agents. The only defense against fraud generated by agents and fraudsters is clean, governed and trusted data.

Clean data is the foundation – not a nice-to-have

At its core, this comes down to alignment. If your systems can’t consistently recognize a single customer, it becomes much harder to detect risk, prevent fraud or trust model outputs.

Capabilities like entity resolution, data quality and governance aren’t back-office concerns. They directly impact how well your models perform and how confidently your teams can act on the results. Sound data management practices are now a mission-critical requirement as organizations expand AI workflows.

Agentic AI will be a strategic necessity for financial institutions. Organizations that clean up their data first are the ones actually getting value out of AI.

What’s up next?

Episode 3 will focus on AI governance. Experts will discuss the challenges and risks that banks face and the direct implications for how governance must be designed, executed and managed. Piqued your interest? Get primed with the white paper, AI governance: A banking leader’s guide.

Brewing Curiosity: Banking Unfiltered is always serving up hot topics for financial leaders. Don’t miss an episode! Add the series to your YouTube playlist.