Credit Unions Think They’re Ready for AI. Their Data Says Otherwise
By Shyam Pradheep, Co-founder and COO at FinRank
Simple Subscribe
Subscribe Now!
Ask a credit union executive whether AI will matter to their institution and you will get a confident answer. Ask what data that AI would actually run on, and the confidence tends to thin.
In a faculty-sponsored independent study I conducted at Stanford Graduate School of Business, 78% of the 46 credit union executives I interviewed said AI will be a source of competitive advantage. The activity behind that belief varied enormously. Some institutions were using generative AI for internal drafting and summarization; others had moved into real automation or production machine-learning use cases. But one pattern in the responses should give banking leaders pause.
Reality check: Executives rated their institutions’ AI readiness an average of 3.3 out of 5. They rated frontline data visibility 3.0, and the degree to which their workflows actually support data-driven work only 2.8. More telling than the gap itself was how loosely those numbers held together: self-assessed AI readiness had only a weak relationship with the strength of the underlying data foundations, a correlation of r = 0.23. Twenty-four of the 46 executives, more than half, rated their AI readiness above the average of their own data-visibility and workflow scores.
I do not read that as executives misunderstanding AI. I read it as a definitional problem. Ask most institutions how AI-ready they are, and they answer a question about access: Do we have approved tools? Have we run a pilot? Is there a policy? Are our vendors shipping AI features? Has someone shown the executive team a demo that landed? Those are all evidence that an institution is experimenting with AI. None of them is evidence that it can use AI well.
AI Readiness Has a Measurement Problem
An institution can deploy an AI assistant, buy an AI-enabled vendor product, and stand up an internal working group without becoming meaningfully more capable of putting intelligence to work across the business. The distinction is easy to miss, because modern AI tools make experimentation unusually cheap. A department can start using a model in days. A vendor can bolt an AI feature onto its product without touching the institution’s underlying architecture. And an executive team can watch a polished demonstration long before the data required to reproduce that experience reliably exists anywhere in its own environment. The result is a dynamic that’s all too familiar with AI that looks mature at the interface while remaining immature at the operating layer.
Consider what a genuinely AI-ready institution should be able to do. It should be able to identify the relevant information about a member. That information should be clean and consistent enough to trust. Systems should be able to reach it without extensive manual stitching. A governed workflow should be able to act on it. And someone should be able to say what business outcome the AI is meant to improve and be accountable for measuring whether it did. That is a considerably higher bar than having access to a model.
Key insight: This matters more in financial services than in most industries, because the AI problem here is a different shape. The value is not primarily in asking a general-purpose model questions it already knows how to answer. The larger opportunity is connecting intelligence to proprietary institutional information and to real operating workflows and that is precisely where most institutions are weakest.
The Real Bottleneck Is Data That Can Move
The most consistent finding in my interviews was not that credit unions lack data. It was closer to the opposite: they have enormous quantities of it. What they lack is a unified, accessible foundation that lets people and systems act on that information while it still matters.
Forty-five of the 46 executives said control of member data was important. In the same conversations, they described fragmented cores, definitions that drift from one department to the next, vendor-controlled data environments, and real uncertainty about who owns the institution’s data strategy. One large institution told me it had spent roughly 18 months building a data lake and still could not answer basic questions about member behavior across product lines.
Key insight: That story captures the distinction that matters: having a data warehouse is not the same as having data readiness. A warehouse can centralize information without making it operational. The question is not only whether the data sits in one place, but whether the institution can see it, trust it, and use it inside the workflows where decisions actually get made.
For AI, that gap gets expensive quickly. A model will summarize a clean document remarkably well. It will not reconcile five departments’ contradictory definitions of “active member.” It can surface a predicted attrition risk, but the prediction is worth little if nothing downstream determines what happens next. It can personalize an interaction, but only if the institution can assemble enough of the member relationship to know what personalization is appropriate. AI amplifies whatever operating foundation sits underneath it: strong data and workflows make intelligence more useful, and weak ones make errors faster, harder to trace, and considerably more convincing.
Five Questions That Reveal Whether an Institution Is Actually Ready
Instead of asking “Do we have an AI strategy?”, leaders should work through a more operational set of questions.
1. Can we see a coherent view of the member?
Pick one valuable use case and test it. If a frontline employee needs to understand a member’s deposit relationship, loan history, digital activity and recent service interactions, can that person get to a coherent picture quickly? If the answer involves four systems, three logins and one tenured employee who knows where everything is buried, the AI problem starts well before AI enters the conversation. This does not require a perfect enterprise-wide member 360 on day one. It requires a reliable view of the information the specific workflow needs.
2. Can the workflow act on that information without manual stitching?
Visibility is necessary and not sufficient. Say a model flags a member whose deposit behavior suggests a high probability of attrition. What happens next? Does the insight trigger a workflow, does a specific person receive it, is there a defined response, and can anyone measure whether the intervention worked? A prediction sitting on a dashboard is not operational intelligence; readiness means information, decision and action are connected to each other. It may be no accident that workflow support for data-driven work was the lowest-rated capability in the study, at 2.8 out of 5.
3. Is one person accountable for the data foundation?
Plenty of institutions describe data as a strategic priority while spreading responsibility for it across technology, operations, marketing, finance and the individual business lines. Shared involvement is healthy; shared accountability usually is not. The institutions moving fastest in my research were more likely to have a dedicated data leader. In other words, someone whose job is the foundation itself. That role needs more than a title. It needs enough authority to set common definitions, set priorities, push back on vendor limitations and move resources. If everyone owns the data strategy, there is a real chance that no one does.
4. Can we access and move our own data?
Vendor dependence is unavoidable for most community institutions, which makes portability, rather than independence, the useful objective. For every major technology relationship, a leader should be able to answer a short list of questions. What data does this vendor hold? Can we export it, and is the export usable? How often can we get it? Can another system act on it? Do we need the vendor’s permission every time we want to build a new workflow? An institution does not control its AI future if it cannot reliably reach the information required to train, ground or operate intelligent systems.
5. Can leadership name one measurable use case?
“Use AI” is not a strategy, and neither is “become an AI-first institution.” A strong first use case has four properties: a defined workflow, accessible data, a measurable economic or service outcome, and an accountable owner. It might be reducing manual document processing time, improving fraud detection, identifying likely deposit attrition, lifting loan conversion, improving call routing, or helping service staff find answers faster. Which one matters far less than the discipline of proving the whole chain works. Data reaches the model, the model improves the decision, the workflow acts on that decision, and someone measures the result. Until that chain closes, expanding AI mostly expands the number of experiments.
Sequence AI Around One Workflow
The temptation is to begin with enterprise transformation. The more practical path runs roughly the other direction: start with a single workflow important enough to matter and narrow enough to understand.
Map the process as it exists today. Understand what information each step requires, where that information lives, who owns it, and where a person currently intervenes to hold it together. Then fix the minimum data foundation that one workflow needs. This is where inconsistent definitions, access limitations and vendor dependencies stop being abstract, and they are far easier to solve in service of a concrete outcome than as a standalone modernization program.
Only then introduce AI, and only where it produces a specific improvement: automating document handling, improving routing, surfacing a prediction, helping an employee make a better call. Do not ask the technology to reinvent a workflow the institution has not yet taken apart.
Key insight: Build governance alongside adoption rather than after it. One risk executive in my research framed the concern perfectly: the worry is not AI itself, but adopting it badly and creating a compliance problem that is difficult to unwind. That is much easier to prevent than to reverse once dozens of disconnected tools have already spread through the organization.
Then measure. Did processing time fall? Did conversion improve? Did losses decline? Did employees resolve issues faster? Did a predicted attrition signal actually help retain deposits? If the workflow improved, expand it; if it did not, find out why before scaling.
Sequencing this way also builds a more realistic path toward the higher-value uses of AI. The research surfaced a useful three-layer framework: operational efficiency, which automates manual work and removes friction; member insight, which uses intelligence to understand needs and personalize interactions; and strategic intelligence, which improves portfolio decisions and competitive positioning. Most institutions are still working through the first layer, and that is not a failure. Operational use cases are where an institution builds confidence, governance and data discipline while generating a tangible return. The mistake is not starting with efficiency. The mistake is assuming that buying tools for layer one automatically creates the foundation required for layers two and three.
Bottom line: The institutions most likely to win with AI will therefore look less impressive in the short run. They will spend more time defining data ownership than announcing pilots. They will fix an integration before adding another model. They will kill a promising experiment because the underlying information cannot be trusted, and they will stay on one unglamorous workflow until it works exceptionally well. That discipline looks slow, and it is what lets an institution move quickly later.
The AI race in banking will not be won by whoever assembles the longest list of AI tools. It will be won by the institutions that can connect intelligence to reliable proprietary data, embed it in governed workflows and turn it into measurable action. Which makes the strongest opening question one that has almost nothing to do with AI: can we see, trust and use our own data well enough to act on it?
