Can Your Bank Trust Its Data Enough for AI? Five Questions to Ask First

By Sam Palmer, Co-Founder and CEO of Libero Labs

Financial institutions of every size are asking the same difficult question: Is our data reliable enough to support artificial intelligence?

That concern is shared by small credit unions that still depend on spreadsheets and large banks with substantial cloud data budgets. Regardless of an institution’s size or technology investment, the challenge remains the same: AI is only as dependable as the information it receives.

People can often recognize when a number appears outdated or inconsistent. An experienced loan officer may question an unexpected figure or investigate its origin. An AI system, however, may use inaccurate information without hesitation and deliver an answer that appears just as polished as a correct one.

The Data Quality Assembly Line Problem

Many financial institutions still manage data quality after information has already moved through the system. A report is reviewed, an error is discovered, and employees work backward to identify where the issue began.

This approach resembles a factory that checks products only after they leave the production line. By the time a problem is found, the information may already have been included in dashboards, reports, forecasts, or business decisions.

Toyota addressed a similar challenge in manufacturing by giving workers the authority to stop an assembly line as soon as they identified a defect. The value was not in the cord used to stop production. It was in giving responsibility to the person closest to the problem.

Financial data pipelines need a similar capability. Errors should be identified and contained where they are created, rather than discovered much later by someone reviewing a report.

Key takeaway: Improving data quality is not simply about running more checks. Institutions need controls that detect issues at the source and prevent bad information from spreading through the organization.

Why Artificial Intelligence Increases the Risk

AI makes data governance more important because of the way these systems interpret and query information.

When an AI tool receives a question about institutional data, it may write a query against raw tables and attempt to interpret the available fields. Alternatively, it can use a governed semantic layer that defines important terms consistently across the organization.

For example, a semantic layer can establish exactly what the institution means by “active member,” “delinquent loan,” or “at-risk customer.” Without those definitions, an AI system may infer meaning from column names or surrounding data and produce an answer that sounds credible but is incorrect.

A 2026 benchmark from dbt Labs found that questions within a system’s modeled scope achieved approximately 84% to 90% accuracy when answered through raw, unmodeled data. The same questions reached roughly 98% to 100% accuracy when processed through a properly governed semantic layer.

The difference becomes even more important when a question falls outside the data that has been modeled. A raw query may still generate an answer, even when the system lacks sufficient information to answer accurately. A semantic layer is more likely to identify the limitation and decline to respond.

Key takeaway: An AI system that acknowledges what it does not know is safer than one that guesses confidently. In financial services, a cautious response is often more valuable than a fast but unreliable answer.

AI Developers Face the Same Data Challenges

This issue is not limited to banks and credit unions. AI companies are also discovering that reliable answers require carefully governed data environments.

Anthropic, the company behind Claude, has described challenges with its own internal analytics. The organization reported that query accuracy dropped as low as 21% before it developed a more controlled data layer. Even after making improvements, Anthropic acknowledged that silent failures remain a concern. These occur when an answer appears correct but is based on inaccurate or incomplete information.

That experience offers an important lesson for financial institutions. If companies building advanced AI systems still need extensive work to make their own data dependable, banks and credit unions should not expect accurate results simply by connecting an AI tool to a data warehouse.

Regulators Are Increasing Their Focus on AI Governance

Regulatory expectations are also moving toward stronger controls over models and the data used to support them.

In April 2026, the Federal Reserve, the Office of the Comptroller of the Currency, and the Federal Deposit Insurance Corporation issued revised model risk management guidance known as SR 26-2. The guidance replaced a framework that had been in place since 2011 and applies most directly to banking organizations with more than $30 billion in assets.

The updated framework identifies generative and agentic AI as novel and rapidly evolving technologies that are not fully covered by the guidance at this stage. However, its central principle applies broadly: financial institutions must be able to demonstrate how their models, data, controls, and validation processes work.

Smaller institutions and credit unions are also receiving guidance as they evaluate AI adoption. The National Credit Union Administration maintains AI governance resources designed to help credit unions manage the risks associated with these technologies.

Five Questions to Ask Before Deploying AI

Institutions do not need to choose a specific vendor or technology platform before addressing the fundamentals. They should first answer five practical questions about their data environment.

1. Can you trace every important number to its source? If a board member asks why a figure changed, can your team show the data’s full path, or would employees need to reconstruct it manually?
2. Do key business terms have consistent definitions? Terms such as “active member,” “delinquent,” and “at risk” should have one approved meaning across reports, departments, and systems.
3. Where are data problems detected? If incorrect information enters a pipeline, will an automated control stop it before it reaches a dashboard, or must an employee notice the issue first?
4. Can you predict the impact of a data change? Before a core migration, system update, or new data connection, can you identify which reports, models, and processes may be affected?
5. Is sensitive information protected as it moves? Security should cover data while it is being copied, combined, transformed, or exported, not only while it remains in its original system.

Reliable AI Starts With Governed Data

These five questions are connected. Each highlights the same underlying weakness: a gap between where data is stored and where decisions are made from it.

When there is no clear governance between those points, errors can move through multiple systems before anyone notices. Definitions can change from one report to another. Sensitive information can be duplicated without adequate controls. AI tools can then use that inconsistent data to produce answers that appear authoritative.

Closing the gap requires more than confidence in a data warehouse or an AI platform. Financial institutions should be able to demonstrate where information came from, how it was transformed, who controls it, and what safeguards are in place when something changes.

AI can improve decision-making, efficiency, and customer service, but only when the information behind it is accurate, governed, and traceable. Before deploying advanced AI tools, banks and credit unions should make sure their data is ready to support the responsibility that comes with automated answers.

Sam Palmer is Co-Founder and CEO of Libero Labs. He has more than 20 years of experience in capital, strategy, and execution and previously co-founded and led Appomattox, a $2 billion assets-under-management outsourced chief investment officer platform for ultra-high-net-worth families and endowments.

Source: thefinancialbrand.com