If you can't state a reason for introducing that data, don't know its purpose, and it adds risk, then it likely shouldn't be used for that AI use case, at least not yet.
Chad PrimakowAi & Data Leader, EverpureAs banks move AI from experimentation into production, understanding the data behind their systems becomes increasingly important. The challenge goes beyond data quality. Banks need to know what information they’re using, where it came from, why it matters, and whether it’s appropriate for the intended purpose.
That challenge becomes more consequential as AI evolves from generating insights to agents capable of taking action. Without a clear understanding of the underlying data, banks risk introducing unreliable results, exposing sensitive information, and making it harder to maintain appropriate oversight.
More data doesn't mean better AI
Financial organizations accumulate enormous volumes of information over time, often retaining records long after their original purpose has disappeared. Making more of that information available to AI applications without assessing its relevance can compromise the accuracy of their results.
Chad Primakow, AI and Data Leader at Everpure, who previously spent nearly two decades at American Express, most recently serving as Director of Cybersecurity and Enterprise Data Management, identifies two underlying problems: data rot, which encompasses redundant, obsolete, and trivial information, and dark data, information whose business context or value remains unknown. "Dark data is when you throw data into the ecosystem, and you don't know it's there," he explains. "You don't understand the meaning behind it. It just exists. Until you're able to define the context, it's perfectly meaningless."
Consider historical financial ledgers that have been incorporated into a bank's data environment. The records may be accurate, yet the people responsible for managing them may no longer understand why they were retained or what business purpose they serve. "It becomes a liability when the data context is unknown, or when the data is redundant, obsolete, or trivial," Primakow warns. "That leads to false outcomes. It skews results and makes you question the accuracy and validity of the data."
Banking systems contain sensitive customer information, including account numbers and national identification numbers, alongside confidential business information with intellectual property and trade secrets, such as proprietary fraud models, credit decision processes, unpublished marketing campaigns, and pre-release earnings. Identifying and classifying information before it enters an AI pipeline can help banks prevent sensitive or unsuitable records from reaching applications that have no legitimate reason to consume them.
AI is only as good as the data behind it
Data readiness should begin with a business objective. Without a measurable outcome, banks can spend considerable time assembling information without knowing how its relevance or quality should be assessed. As Primakow says, "It starts with defining a few measurable outcomes, not trying to boil the ocean. You need well-defined, high-quality data that builds trust in your data ecosystem.”
Every dataset introduced into an AI application should have a defined business purpose, identifiable ownership, and a reliable source. Banks also need to understand who will consume the information, how it will be delivered, and how current it needs to be to support the intended outcome. A dataset may contain accurate historical records that are no longer relevant to a lending decision, for example. Another may be timely and reliable but lack clear ownership or contain sensitive information that makes it unsuitable for a particular application.
Banks don't need to resolve every enterprise-wide data problem before putting AI to work. They can focus on the information relevant to specific use cases, assess its quality and sensitivity, and filter unsuitable records as data enters the pipeline. "If you can't state a reason for introducing that data, don't know its purpose, and it adds risk, then it likely shouldn't be used for that AI use case, at least not yet," Primakow says.
Scaling AI responsibly requires continuous oversight
The transition from AI experimentation to production introduces a different set of responsibilities for financial institutions, particularly as agents gain greater autonomy. An inaccurate AI-generated response may require correction, while an agent acting on the same information could execute a task with financial or operational consequences. "Remember that an agent won't question its objective by default," Primakow warns. "That's not something agents generally do unless they're told to."
Banks need clearly defined workflows that determine which information agents can access, what actions they can perform, and when human intervention is necessary. Data governance, security, and risk management should inform these decisions from the outset and evolve alongside AI capabilities. "All three need to move at the speed of business," Primakow says. "If they don't, the value will be lost, or the data may have aged by the time you revisit what was originally stated."
Once agents enter production, continuous monitoring can help banks identify deviations from established parameters and intervene before errors escalate. Human oversight is especially valuable during early deployments, when institutions are still establishing confidence in the quality of their data and the reliability of automated decisions. As applications demonstrate that they can operate within established quality and risk parameters, banks can gradually expand their autonomy while maintaining oversight.
Want to go deeper? Listen to Chad Primakow discuss dark data, data quality, and AI governance in banking on The Money Pot podcast, recorded at Money20/20.





