Most banking AI projects will fail. Yours doesn't have to.

The technology is easy. Most banks will struggle to understand the work it needs to do.

Many community banks are currently evaluating AI solutions using a familiar process: There is board-level pressure to have an AI strategy in place. An AI Steering Committee is formed. AI vendors provide polished demos that look like magic. Now the bank must answer two key questions: Which vendors do we trust, and how do we maximize our chances of success? The answer has little to do with the vendor and almost everything to do with the bank.

Over the past twenty-five years working in commercial lending operations at community banks, we’ve seen wave after wave of technology arrive, from core replacements and loan origination system migrations to digital imaging and workflow automation. Some of these projects were successful, but many fell short, and rarely for the reasons anyone anticipated when the contract was signed.

We believe AI is going to be the same story, but more so. Most banking AI implementations over the next several years will fail to deliver on the promises made in the demos. The problem will not be the AI itself. Most banking AI projects fail because no one inside the bank can clearly define the requirements the AI needs to work its magic. In an age when answers to questions are essentially free, it’s the questions that matter.

What banks really want AI to do

Modern AI is genuinely impressive. It can read a document, extract information, summarize a conversation, draft a memo, classify a transaction, perform complex data analysis, and much more. For the easy 80% of most tasks, it is faster than a human and increasingly more consistent.

The 80% is the portion you see in the demos.

The part you do not see in the demo is everything that must be true for the AI to perform that task the way your bank actually needs it done. What is the exact policy on this loan type at this bank? Why does this particular borrower get treated differently than the general rule would suggest? Why does the credit administrator always check the prior year’s statements before signing off on this kind of file? What does it mean when the relationship manager writes “see attached” in the comments? Is that referencing something in imaging, in email, in a shared drive, or in a folder on someone’s desktop?

This is the work. The model is not the work.

The model can read the document. But somebody has to tell it which document, which fields, in what context, with what exceptions, with what tolerance for ambiguity, and with what definition of “done.” That somebody has to know the bank well enough to answer those questions correctly. And in most community banks, the answers to those questions are not written down.

Humans are good at papering over cracks. AI is not.

There is an analogy from anthropology that translates almost perfectly to community banking. We came across it recently in a post on Instagram (@sierra_irl), who has spent three years implementing AI within large organizations. She points out that ethnography, one of the core methods in anthropology, is about understanding how things actually work versus how people say they work. Those two things, she observes, are almost never the same. Most of how any organization actually operates is never fully written down. It lives in people’s heads as tribal knowledge.

The core insight is that humans are extraordinarily good at compensating for operational inefficiencies. We adapt, improvise, and fill gaps. We work around broken processes without even realizing we are doing it. A long-tenured credit analyst does not consciously think “the loan policy is silent on this scenario, so I will apply the unwritten convention I learned from my predecessor in 2014.” He just does it. The convention is part of how he thinks.

AI systems do not adapt this way. They follow the process they are given. If the process is incomplete, the AI will improvise the missing pieces, but not in the way a human would. It guesses what belongs in the gaps based on patterns and produces output that is confident, plausible, and occasionally wrong. The wrong outputs are the dangerous ones, because it takes time for someone to notice that the cracks have been papered over incorrectly.

When you put AI on top of an undocumented process, it exposes every crack humans have learned to paper over.

This is not a problem AI vendors talk about, because the demo is built on a clean example. It’s the messy 20% of the processes that will take 80% of the time to get right.

The Linda Test

When we ask a community bank how a particular process is defined, the answer almost always falls into one of three categories:

Defined. “Here is the procedure document, here is the policy, here is the workflow in the system, and here is the audit trail.” This usually includes things like the general ledger, regulatory reporting, and anything else examiners look at. These processes are clean because they have to be.

Depends. “It is generally done this way, but it depends on the situation, and we have some exceptions documented somewhere.” This is where many lending processes live. The documentation exists, but it is partial. Real-world execution involves judgment calls that are not documented.

Tribal. “Linda has been doing that since 1998 and she just kind of knows.”

We call this The Linda Test, and it is the single most useful diagnostic we have found for predicting where AI will work in a bank and where it will fail. The third category is the one that breaks AI implementations. Not because Linda is bad at her job. She’s excellent, which is why she’s been doing it since 1998. The problem is that her process exists only in her head. Automate it before documenting that knowledge, and the AI will produce results that look right, are often right, and are sometimes wrong in ways that may take months to discover.

Linda exists at every community bank. Sometimes there are several Lindas. The institution runs on the accumulated tribal knowledge of people who have been doing the work for decades. That knowledge is a major part of the bank’s operational system. The software is just the surface.

Run your bank’s most critical processes through The Linda Test before you sign any contract with an AI vendor. The Defined processes are candidates for AI today. The Depends processes need work first. The Tribal ones are not AI problems. They are documentation problems, and the documentation must come first.

What the AI labs themselves are telling us

If this sounds like a marginal concern, consider what the companies building the AI have done about it.

In May 2026, OpenAI and Anthropic, the two leading AI labs, committed a combined $5.5 billion to building consulting and deployment organizations staffed with what the industry calls Forward Deployed Engineers, or FDEs. OpenAI launched a deployment company with $4 billion in initial funding and a $10 billion valuation. Anthropic launched a competing services firm with $1.5 billion from Blackstone, Goldman Sachs, and Hellman & Friedman. Both ventures exist to embed FDEs directly within customer organizations to ensure the AI actually works. The model itself, both firms have made clear, is not enough.

These are the companies building the technology. They concluded, at the cost of billions of dollars, that the technology alone is not enough. The implementation work is the hard part, and the deployment work is the moat.

The major consulting firms have reached the same conclusion. OpenAI has announced multi-year partnerships with Boston Consulting Group, McKinsey, Accenture, and Capgemini. Anthropic has done similar deals with Accenture and Deloitte. Billions of dollars and hundreds of thousands of consultants are being mobilized for one purpose: to actually get these models into production inside real organizations.

If the AI labs themselves believe the implementation is harder than the model, every community bank evaluating an AI vendor should take that seriously.

What this means for community banks

A community bank evaluating AI solutions should ask questions different from those vendors would like. Instead of “How well does your AI solution work?”, the better question is, “Do we understand our own processes well enough to change them?”

If the answer is yes (processes are documented, data is clean, exceptions are defined, and institutional knowledge is captured), then many bounded AI use cases can be implemented successfully.

If the answer is no (and for most community banks, the answer is no), implementing AI is the second step. The first step is to prepare the bank: document workflows, clean and standardize data, and turn unwritten judgment calls into clear policies.

This work is unglamorous and hard to demo. It is also what determines whether AI creates value or quietly fails behind a dashboard no one uses.

At BankPoint, we help community banks do the foundational work that determines whether AI succeeds or fails: documenting workflows, standardizing data, and capturing institutional knowledge.

We call this the BankPoint Advantage Program. It prepares banks for successful technology initiatives, whether implementing a complex loan origination system, automating processes with AI, or pursuing broader digital transformation. As trusted advisors for banks, we’ve provided similar foundational consulting services for two decades.

The platform you choose matters less than the preparation. Whether you use our BankPoint platform, another vendor’s platform, or build a solution yourself, the readiness work comes first.

The unfashionable conclusion

The community banks that will get the most out of AI over the next few years will not be the ones that move first. They will be the ones that move seriously, treating AI as a forcing function for the operational discipline they have been deferring. The ones that get there first will not just have AI that works. They will have an institution whose operations are finally documented, which is worth doing on its own merit, with or without AI.

Ready to prepare your bank for AI and other technology initiatives? Contact BankPoint today to learn more about the BankPoint Advantage Program or schedule a demo.