- 95% of large enterprise AI initiatives are failing to deliver business value (MIT, BCG)
- Big banks lead the failure rate because they have the budget to keep funding failed programs longer
- Four specific structural mistakes drive most of the failure
- Community banks can win this race by not copying any of the four mistakes
The biggest banks in the world have spent more on AI in the last three years than community banks will collectively spend on technology in the next decade. And 95% of what they are building is failing.
Curious where your institution actually sits relative to these patterns? A 30-minute conversation is enough to find out.
In Pillar 1 of this series, The 90-Second Bank, I argued that community banks should not try to win the AI race the big banks are running. This piece goes deeper on the why. Specifically, on what big banks are getting wrong and what community banks should do differently.
The failure is not random. It is structural. Once you see the pattern, you can avoid it.
The 95% Pattern Is Not Bad Luck
Two pieces of research from 2025 tell the same story. MIT found that only 5% of integrated AI pilots have delivered significant value at scale. BCG, surveying more than 1,250 executives, found that only 5% of firms actually realize AI value at scale, and those firms achieve roughly five times the revenue growth of their peers.
Two studies. Different methods. Same number.
The temptation is to assume the technology is not ready or the talent is insufficient. Neither is the answer. The largest banks have the deepest data science teams, the most enterprise vendor relationships, and the right tools. They are still failing at the 95% rate.
The failure is structural. Big banks fail at AI in predictable ways for predictable reasons. Here are the four specific mistakes.
Four Mistakes Big Banks Are Making with AI
Mistake 1: Treating AI as a Technology Procurement Problem
Big banks treat AI as something to buy, then figure out how to use. They sign enterprise license agreements with the major AI platforms and stand up a center of excellence to drive adoption. Technology arrives before the operational problem has been defined.
The result is technology in search of a use case. Drew Cukor, who led AI transformation at JPMorgan and previously at the Pentagon, calls this the distinction between AI-augmented and AI-native organizations. AI-augmented means you bolted intelligent systems onto workflows that were never redesigned. AI-native means the work itself was rebuilt around what AI can do.
The community bank alternative: Start with the operational problem, not the technology. Pick the highest-friction process. Redesign it. Deploy AI into the new design.
Mistake 2: Funding AI as a Long Bet with No Intermediate Accountability
The Economist Enterprise published research in 2026 showing that about 60% of firms take between 7 and 12 months to ship an AI project, and about 60% lack a fully established AI development life cycle. The report calls this perpetual piloting.
At a big bank, this looks like an 18-month transformation program approved by the board. Twelve months in, the strategy team has moved on. Eighteen months in, the project gets quietly defunded. That is gambling with an extended fuse, not transformation.
The community bank alternative: The self-funding operating model. Pick one process. Deploy in 90 days. Measure the result. Use the operational value to fund the next increment.
Mistake 3: Investing in Human Review Instead of Workflow Redesign
The same Economist research found something striking. About 50% of firms cite human review as a top ongoing cost of running AI systems. Only 4% cite employee upskilling as a significant cost.
Institutions are paying heavily for human beings reviewing AI outputs, because the AI is producing too many bad outputs to trust. The expense line grows. The returns do not. This is the natural consequence of Mistake 1: when AI is deployed onto legacy processes with sub-optimal data flows, architecture, and governance, the outputs are unreliable enough that humans have to check every one.
The community bank alternative: Redesign the workflow first. Co-create it with the users who will work alongside the AI so they can provide meaningful input, influence how their roles will evolve, and mitigate adoption risk.. The human review burden drops because the AI is no longer being asked to operate inside a process that was never built for it, and is replaced by higher-value contributions from those team members.
Mistake 4: Deploying AI Across the Bank Instead of Process by Process
Big banks set enterprise AI strategies that span every department: lending, fraud, compliance, customer service, marketing, treasury. They mandate AI adoption across all of them simultaneously.
Nothing gets done well because nothing is the focus. By trying to deploy AI everywhere, the institution makes no single deployment good enough to scale. Every pilot becomes a half-built system competing for engineering and operations attention with twenty other half-built systems.
The community bank alternative: One process per increment. Ninety days. Measurable outcome. Then the next process. Depth over breadth.
The Two Approaches, Side by Side
|
Big Bank Approach |
Community Bank Alternative |
|
|---|---|---|
|
Starting point |
Buy enterprise AI platform first |
Identify highest-friction process first |
|
Funding model |
18-month transformation, quarterly check-ins |
90-day self-funding increments |
|
Investment focus |
AI tools plus human review of outputs |
Workflow redesign plus targeted AI deployment |
|
Scope |
Enterprise-wide AI strategy across all departments |
One process per increment, sequenced |
|
Accountability |
Status updates and adoption metrics |
Measurable operational outcome defined upfront |
|
Failure mode |
Quiet defunding after 12 to 18 months |
Disciplined stop after one increment |
Each row is a strategic choice. The choices on the left produce the 95% failure rate. The choices on the right are the discipline of the 5% that succeed.
Why Community Banks Are Structurally Better Positioned
The optimistic read on the 95% failure rate is that community banks have structural advantages in not falling into it.
You cannot afford 18-month bets without accountability, so you will not make them. Big banks can absorb a write-off on a failed AI program. Community banks cannot. That constraint forces discipline, which is exactly what the 5% have.
Your operational structure is simpler. Fewer departments. Fewer process variations. Fewer political fiefdoms. Single-process focus is dramatically easier to maintain when the institution does not have to navigate fifty competing stakeholders. Your operations leaders are also closer to your customers, which makes workflow redesign vastly more achievable.
The 95% pattern is real, but it is not destiny. Community institutions that recognize the pattern and choose differently in the next twelve months will produce a structural advantage over the big banks stuck in the failure pattern and the community banks copying them.
What to Do Instead
If you are a CEO, President, or Board member at a community bank, the move is to honestly evaluate where your institution sits relative to the four mistakes.
-
Are you treating AI as a technology procurement problem?
-
Funding it as a long bet with no intermediate accountability?
-
Investing in human review instead of workflow redesign?
-
Trying to deploy AI across the bank instead of process by process?
If two or more answers are yes, the program needs to be restructured. There is a supporting piece in this series specifically on the questions every community bank board should be asking about AI that gives a framework for the evaluation. The self-funding 90-day increment model is the practical path that bypasses every one of the four mistakes by design.
If the patterns hit close to home, the next step is a 30-minute conversation.
No pitch. No deck. Just listening.
