Somewhere right now, a founder is sitting in front of a model that works. 

It predicts fraud better than legacy systems. It identifies patterns no human analyst could see. It might even anticipate consumer harm before it happens. The demo is clean. The metrics are strong. The story is compelling. 

And yet, in banking — and in most large enterprises — that’s not enough. 

Because the question is never, “Does it work?” 

The question is, “Can we live with it?” 

That distinction is where most AI companies fail, and where a small number quietly succeed. 

The idea is the easy part. 

The path from idea to bank does not start with a purchase order. It starts with translation. 

A developer might describe their product as a machine learning model that uses unstructured data to predict behavioral anomalies. 

A bank hears a potential source of regulatory, operational and reputational risk. 

Before anything else happens, that idea has to be reframed, not as innovation, but as controlled value inside a regulated system. 

That reframing doesn’t just happen inside banks. It often happens earlier, at the venture level. 

Jeff Cherry, founder and managing partner of Baltimore-based Conscious Venture Lab, constantly sees that divide between compelling technology and deployable enterprise solutions. 

“It’s the same thing with any company, not just AI. Which is first, what is the unique insight that the founder has that will change the way we live and operate in the world?” Cherry said. “Technology without insight isn’t going to create a sustainable company. Second is always the ability to execute. Ideas are a dime a dozen, execution is really all that matters.” 

That distinction, between what sounds impressive and what can actually survive deployment, is where the real filtering begins. 

Gap between product, adoption

Most founders believe they are selling a product. 

In reality, they are asking an institution to absorb a new form of risk. 

That gap, between product and adoption, is where many companies stall, and where venture investors increasingly play an active role. 

Cherry describes the challenge this way: 

“We help companies explore pivots but frankly that’s not our value add. We’re not the experts at the specific topics. We are experts at helping founders see what they don’t see, but too often founders are unable to get out of their own way. We look at solutions and understand whether or not they are pointed at a real problem. We can help founders see that. Then it is up to them to develop the insight to find what’s next.” 

This translation layer is about reshaping the product, narrowing the use case, structuring outputs, aligning with how decisions are made internally and anticipating regulatory and operational scrutiny. 

Without that translation, even strong models struggle to get traction. 

From the outside, it might look like banks are slow to adopt AI. Inside, the reality is different. Adoption is happening, but it is being filtered through layers designed to ensure control. 

An idea enters through a business problem such as fraud, complaints, operational breakdowns, but no one is asking for AI, they’re asking for a solution.

If the solution shows promise, it moves into testing.  

Risk assessment

Following testing, model risk teams evaluate whether the system can be understood, validated and challenged. Legal and compliance assess potential consumer impact. Risk teams evaluate the vendor. 

This is where most founders hit a wall, because the system rejects it. 

“They don’t come to the table with a direct solution to a problem that is keeping the enterprise awake at night,” Cherry said. “Or more importantly, keeping someone specific in the enterprise awake at night. Generic ‘I can solve any problem’ solutions aren’t well received at the enterprise level. Get specific about the value you bring for a specific problem that no one can solve except you and your team.” 

It’s not a technology problem, but rather a systems problem. 

Over the past decade, the regulatory environment, particularly under the Consumer Financial Protection Bureau, has reshaped how banks evaluate decisions. 

Through complaint data, enforcement actions and a focus on consumer harm, the expectation is clear: If a bank cannot explain a decision in terms of its impact on a consumer, it cannot defend that decision. 

This has direct implications for AI. 

Models must do more than predict. They must explain outcomes, demonstrate fairness, withstand audit and align with consumer protection standards. 

This is not a constraint, it is a design requirement. 

And increasingly, it is shaping what projects get funded in the first place. 

Institutions remain accountable for the decisions their systems make. Regulatory priorities may shift. Enterprise accountability does not. 

This is why firms like AI-driven compliance providing company Safir have found traction. 

It delivers tightly defined solutions built for specific financial use cases, structured for existing workflows and designed for explainability and audit. 

When deployed, it looks less like innovation and more like infrastructure. 

That is not an accident. 

That is the requirement. 

Emerging challenges

For venture-backed companies, this creates a different kind of challenge. 

In most industries, speed is an advantage. In financial services, endurance matters more. 

Sales cycles stretch, validation takes time and procurement layers add friction. 

The question for investors is no longer just whether the product works, but whether the company can survive long enough, and adapt enough, to make it through the system. 

That is why the final filter is so simple, and so difficult. 

As Cherry summarizes: 

“A unique insight that changes the way we live, a clear and identifiable problem, and a solution that is clearly executable by the team.” 

The AI conversation is still dominated by capability, what models can do. 

But inside banks, the defining question is different: 

Not “How powerful is the model?” but, “how safely can it operate inside a regulated system?” 

The companies that succeed will not be the ones building the most advanced models. They will be the ones building models that fit within regulatory frameworks, align with institutional processes and can be defended when it matters.  

Because in the end, AI does not win in banking by being smarter. It wins by being acceptable. 

Somewhere, that founder with the perfect model is still trying to get a meeting. 

The ones who succeed will not be the ones with the best demo. They will be the ones who understand the system they are entering, and build for it from the start. 

Jim McCarthy is chairman for McCarthy Hatch, which provides data-driven insights for risk management. A founding member of the Consumer Financial Protection Bureau, he is a keynote speaker and fractional CRO/CCO in the financial services industry with more than three decades of experience.