Most AI initiatives in banking stall not because models fail, but because fragmented systems prevent AI from reasoning across data. Until institutions build a shared data foundation, AI will continue to amplify operational chaos rather than deliver value.
Every banking and mortgage conference I've attended in the past eighteen months has had the same energy: urgent conversations about AI, impressive vendor demos, and executives leaving with a mix of excitement and anxiety.
Then I watch what happens next.
The pilot gets approved. A team gets assembled. A use case gets selected—usually something safe like document classification or chatbot enhancement. Three to six months later, the initiative is either quietly shelved or limping along with results that don't justify the investment.
I've seen this pattern repeat at regional banks, credit unions, and independent mortgage banks across the country. And after serving as CTO at a digital bank and spending years as a fractional CTO helping financial institutions navigate technology decisions, I've come to a conclusion that might be uncomfortable:
The problem isn't AI. The problem is that AI is exposing a foundation problem we've been ignoring for years.
Let's acknowledge the reality. The pressure to "do AI" is intense and coming from every direction.
Boards are asking what the AI strategy is. Competitors are announcing AI-powered everything. Vendors are repackaging existing products with "AI-enabled" labels. And the headlines keep coming: banks that embrace AI will thrive; those that don't will be disrupted.
The research backs up the urgency. Surveys show that 70% of banking executives say their firms are already using agentic AI in some capacity. AI adoption in financial services has grown to 63% of companies reporting they use generative AI for at least one function. In mortgage lending, Fannie Mae projects AI adoption will rise from 38% to 55% by the end of 2025.
So the intent is there. The investment is there. The talent—while scarce—is being recruited.
Why aren't the results there?
Here's what I've learned from looking under the hood at dozens of financial institutions: 
The typical small to mid-sized bank or IMB runs somewhere between 10 and 30 distinct SaaS platforms. Let me list what I commonly see:
Each of these systems was selected to solve a specific problem, and most of them do their job reasonably well in isolation.
But here's what nobody talks about at the vendor demos: each system has its own data model, its own customer identifiers, its own logic for what constitutes a "loan" or an "account" or a "customer relationship."
There is no shared semantic model. No common language. No unified view.
The research confirms this isn't just my observation. Industry surveys show that 74% of lenders report fragmented data systems as their biggest obstacle to improving borrower experience. More than half of banking executives say legacy infrastructure is their single biggest hurdle to transformation. And over 87% of organizations across industries struggle with disconnected data sources.
In financial services, we've been living with this fragmentation for years. We've built workarounds. We've hired people whose entire job is to reconcile data between systems. We've accepted that getting a true 360-degree view of a customer requires pulling reports from five different platforms and assembling them manually.
It was inefficient, but it worked well enough.
AI changes that equation entirely.
Here's what makes AI different from previous technology waves:
Traditional software follows explicit rules. If X, then Y. The logic is deterministic, and the software only needs to understand its own domain. Your LOS doesn't need to understand your CRM's data model to process a loan application—it just needs to receive the right inputs in the right format.
AI—particularly generative AI and agentic AI—works differently. It reasons. It infers. It connects patterns across contexts. And to do any of that effectively, it needs to understand the relationships between things.
When you ask an AI system to help with customer retention, it needs to understand:
When you ask an AI system to accelerate loan processing, it needs to understand:
None of this is possible when your systems are isolated islands with no shared understanding of entities, events, or relationships.
You can't bolt AI onto a fragmented foundation and expect it to magically create coherence. AI amplifies what's already there. If what's already there is chaos, you get faster chaos.
At this point, many institutions think: "We need better integration."
So they embark on integration projects. They connect System A to System B with an API. Then System B to System C. Then System A to System D. Before long, they have a spider web of point-to-point integrations—each one custom, each one brittle, each one requiring maintenance when either endpoint changes.
I've seen institutions with 50, 80, even 100+ point-to-point integrations. The team responsible for keeping them running is perpetually underwater. When something breaks—and something always breaks—it takes days to trace the failure through the dependency chain.
This isn't integration. It's spaghetti.
And here's the critical point: even well-executed point-to-point integration doesn't solve the semantic problem. You can move data between systems without those systems actually understanding each other. You're just synchronizing confusion.
The integrations don't know that "Customer ID 12345" in your core banking system is the same person as "Contact ID 67890" in your CRM and "Borrower ID 11111" in your LOS. They just pass data back and forth, leaving it to humans (or fragile matching logic) to reconcile the mess.
Let me be direct about what I see as the actual impediment to AI success in banking and mortgage:
Most financial institutions have no enterprise data layer.
No data warehouse. No data lakehouse. No unified semantic model. No single source of truth for the entities that matter to their business.
They have lots of data—more than they've ever had. But it's scattered across systems, stored in incompatible formats, governed (if at all) by different standards, and accessible only through the narrow interfaces each vendor provides.
In this environment, AI initiatives face an impossible task. Before they can deliver value, they have to solve the data unification problem for their specific use case. They build custom pipelines to pull data from three or four systems. They create bespoke logic to match entities. They stand up isolated data stores just to support a single application.
And then the next AI initiative comes along and does the same thing from scratch, because there's no reusable foundation.
This is why AI pilots succeed in narrow, controlled conditions but fail to scale. The pilot team heroically solves the data problem for their scope—and then discovers that solving it again for broader deployment would require a level of effort that makes the business case collapse.
If you're a banking or mortgage executive reading this, here's my challenge:
Stop asking "What AI use cases should we pursue?"
Start asking "Do our systems understand each other well enough for AI to reason across them?"
If the honest answer is no—and for most institutions it is—then your AI strategy needs to begin with data architecture, not model selection.
This isn't a technology problem you can delegate to IT. It's a strategic problem that determines whether your institution will be able to compete in an AI-enabled market or watch from the sidelines while more prepared competitors pull ahead.
The good news: this is a solvable problem. Financial institutions of all sizes are finding pragmatic paths to data unification that don't require ripping out existing systems or betting the business on multi-year transformation programs.
But it starts with acknowledging the real blocker—and having honest conversations about the foundation before chasing the headlines.
In my next post, I'll explore a different way to think about your existing technology stack—not as a mess to be replaced, but as a set of assets that can be connected through modern architectural patterns.
The systems you already have are likely more capable than you realize. The problem is how they're wired together (or not).
And in the final post of this series, I'll lay out what an "AI-ready" architecture actually looks like for a mid-sized financial institution—and why it's more achievable than you might think.
