In Pillar 1 of this series, I argued that community banks should not try to win the AI race the big banks are running.
If that argument made sense, your next question was probably about cost. What does this actually take? How does a community bank justify the investment to a board that has watched neighboring institutions burn money on AI projects that delivered nothing?
That is the right question. It has a specific answer.
The right way to fund AI in a community bank is not as a speculative bet with an 18-month payback. It is as a sequence of 90-day operating increments, each one designed to produce enough measurable value to fund the next. We call this the self-funding operating model. The companies that have actually scaled AI to material business impact are doing some version of it. Most institutions are not.
This article lays out why most AI investments are designed to fail, what the self-funding model actually looks like, how it compares to traditional pilot funding, and how to start your first increment.
The Economist Enterprise published a 2026 report titled Making AI Deliver based on a survey of 1,200 executives at firms already using advanced AI. The findings are uncomfortable.
About 60% of firms take between 7 and 12 months to move an AI project from idea to live production. Barely 1 in 25 manages it in under three months. About 1 in 14 takes more than a year. About 60% of firms lack a fully established AI development life cycle, which means every new project relitigates the same questions: who owns the data, how the system gets tested, who approves deployment. The report calls the result perpetual piloting. Experiments multiply. The discipline to scale what works or kill what does not is missing.
Consider what this means economically.
Every AI project that takes 9 months to deploy is consuming engineering capacity, IT capacity, executive attention, and operating budget for nine months before producing any business return. If 95% of those projects then fail to deliver value at scale, the institution is paying for nine months of activity to find out the project does not work.
This is the structural reason most AI investments fail. They are funded as long bets with no intermediate accountability. The board approves a number, the technology team disappears for a year, and the institution waits for the demo. That is not transformation. That is gambling with an extended fuse.
The same report found that only about 40% of firms formally require teams to track business impact, defined as cost savings, revenue gains, or efficiency improvements. The implication is direct. The majority of institutions investing in AI today have no formal mechanism to determine whether their investments are actually working. They are spending without measuring.
The fix is not more governance committees or a bigger budget. The fix is a different funding model entirely.
"Adoption is an activity metric."
Jose Manuel Silva, VP for Technology and Chief Digital Officer at Natura, quoted in Making AI Deliver, Economist Enterprise, 2026
This is the single most important reframe for any board considering AI investment.
Counting tools deployed, licenses purchased, or pilots launched tells you nothing about whether the technology is creating value. Those are activity metrics. They tell you something is happening. They do not tell you whether the bank is better off than it was last quarter.
The metric that matters is operational outcome. Cycle time. Hours per transaction. Error rate. Cost per loan processed. Revenue captured from faster decisions. These are the numbers that show up on the P&L. They are also the numbers most AI programs cannot produce, because the programs were never designed to.
The self-funding operating model exists to force operational outcomes into the structure of every investment. You cannot deploy an increment without defining what success looks like. You cannot move to the next increment without measuring the result. The model treats the question "is this working" as a requirement, not a hope.
This is the discipline that separates the 5% of firms capturing AI value at scale from the 95% that are not. It is also the discipline most community banks have not yet applied to their AI thinking.
The self-funding operating model has four properties. Each one is non-negotiable.
This is the opposite of how most large bank transformations work. Large banks try to redesign everything at once. They spin up enterprise programs with 18-month timelines and architectural diagrams. Twelve months in, the strategy team has moved on. Eighteen months in, the project gets quietly defunded. That is the mechanic of the 95% failure rate.
A community bank does not have to operate this way.
The structural difference is easier to see when you compare the two models directly.
|
Pilot Funding (Most AI Programs) |
Self-Funding Increment (The Discipline) |
|
|---|---|---|
|
Time to first value |
7 to 12 months |
90 days |
|
Accountability |
Quarterly check-ins, project updates |
Outcome defined and measured upfront |
|
Measurement |
Often informal or absent |
Required, verifiable, in writing |
|
If results disappoint |
Project quietly continues |
Project stops, next increment does not start |
|
Board governance |
Updates and status slides |
Measurable result inside 90 days |
|
P&L impact |
Speculative |
Funds the next increment |
|
Risk profile |
Long valley of declining productivity |
Contained J-curve inside a single quarter |
The table makes the choice visible. Pilot funding is the default model. It is also the model the BCG and MIT research show is producing the 95% failure rate. Self-funding is the discipline the 5% are using.
Let me make this concrete with one process: small business lending.
Most community banks have a structurally inefficient small business lending workflow. The economics are tight, the underwriting work is real, and the customer experience often loses to fintech competitors who can produce a decision in hours rather than days. Here is the difference between how that process runs today at most community banks and how it runs after a single 90-day increment.
The math, even at a small community bank, is substantial. If your bank processes 200 small business loan applications a month, the staff time savings alone are roughly 1,800 to 2,400 hours annually. The faster decision speed also captures applications that would have been lost to abandonment. Both effects show up on the P&L.
This is what a single self-funded increment looks like. One process. Ninety days. Measurable outcomes that materially exceed the cost of the engagement.
The loan officer does not get replaced in this picture. The opposite happens. She is no longer buried in document collection and data entry. She is having more substantive conversations with more applicants, making faster decisions, and producing better outcomes for the bank. That is the AI-native version of her job. It is also the version her customers want her to have.
This is the part of the conversation most institutions are not having. It matters because it explains why most AI spending does not produce returns.
The Economist Enterprise report found something striking about where AI budgets actually flow. About 50% of firms cite human review as a top ongoing cost of running AI systems. Only 4% cite employee upskilling as a significant ongoing cost.
Read that again.
Institutions are spending heavily on human beings reviewing the outputs of AI systems, because the AI is producing too many bad outputs to trust without verification. Meanwhile they are investing almost nothing in training the workforce to work with the AI more effectively. This is a misallocation of resources that guarantees the AI investment will underperform. You cannot make an AI-augmented workflow productive if every output requires a human safety check.
This is the same trap Drew Cukor described in Fortune as the difference between AI-augmented and AI-native organizations. AI-augmented means you bolted intelligent systems onto existing workflows and now you have a human reviewing every output to catch the errors that come from a process that was not redesigned. AI-native means the workflow itself was rebuilt around what AI can actually do, and the human role was redefined to focus on the judgment AI cannot yet replicate.
The self-funding operating model is built to avoid this trap. Every 90-day increment includes workflow redesign, not just tool deployment. Every increment includes the training and role redefinition that makes the new workflow actually work. The institution does not end up paying for armies of human reviewers checking the outputs of AI systems that should never have been deployed in their current form.
This is the difference between an AI investment that pays back and one that becomes an expense line that grows every quarter.
If you are a CEO, a President, or a Board member at a community bank and the model makes sense, here is the practical move.
Pick the process. Look at your operation honestly and identify where friction is highest. Not the most glamorous process. Not the most visible one. The one where your staff spends the most hours doing work that AI can now do faster, and where customers feel the friction in their experience. For most community banks, this is small business lending, mortgage processing, fraud monitoring, or compliance documentation.
Define what success looks like. Before any work begins, write down the metric you are going to measure. Cycle time, hours per transaction, error rate, abandonment rate, or some specific combination. Get the operations leader who owns the process to agree to the measurement method, and provide a current performance baseline. Decide upfront what number will count as success and what number will trigger a stop.
Assign the three roles. Every successful increment requires an executive sponsor (usually CEO or COO), an operations leader who owns the process, and an implementation partner who handles the workflow redesign and AI deployment. The implementation partner cannot succeed without the first two roles being fully present and accountable.
Run the increment. Ninety days. Measurable outcome. Honest evaluation at the end. If it worked, the savings fund the next increment. If it did not, you stop and learn.
This is what we do at WNDYR. We work with community banks and credit unions in 90-day increments against specific high-value processes. We do not sell strategy decks. We do not sign on for 18-month programs. We work with a limited number of institutions per market because the work requires depth and because if we partner with you, we will not be helping your competitor across town.
The institutions that figure out the self-funding model are going to capture significant operating advantage over the next 24 months. The institutions that do not will continue to fund pilots that produce no measurable result.