Models work. Use cases deliver value. And yet, according to MIT NANDA, as many as 95% of organizations are still seeing little or no return on their AI investments.*
The reason is rarely the one banks expect.
This is not a problem of ideas or technical capability. In most cases, it comes down to something more basic: the order in which organisations try to scale.
That shift, from experimentation to industrialization, is where many banks are now getting stuck.
As we explored in AI industrialization vs transformation: How banks move from pilots to scale, the challenge facing banks is no longer whether AI works. It’s whether it can be run, repeatedly and reliably, at scale.
Most AI programs fail in the same place
A model works in a controlled environment. A use case delivers value. The business responds in a predictable way: scale it, replicate it, build more.
Thats where things begin to change.
The conditions that made the initial success possible—limited integration, curated data, contained scope—no longer hold. The same model is now expected to operate across multiple systems, processes, and decision points. Complexity increases faster than capability.
What looked like acceleration starts to create friction. Delivery slows. Costs become harder to explain. Governance effort expands.
At that point, many banks assume they have a scaling problem. In reality, they have a sequencing problem.
The question is how to scale without creating drag.
Banks scale AI in the wrong order
Most banks scaling AI follow a fairly similar evolution consisting of four stages. Yet whether these banks recognize and plan for this evolution or submit to it unwillingly—after some degree of trial and error—is what separates the best in class from the rest.
1. Prove value, but contain the complexity
Early AI work should focus on demonstrating value. That typically means working in controlled environments, limiting dependencies, and prioritizing speed over scalability.
This is where confidence is built.
The risk at this stage is assuming that success in isolation will translate directly into success at scale.
2. Stabilize the foundation before expanding
This is the step most organizations delay, and where many programs quietly fail.
Not because organizations disagree with the undertaking, but because it’s difficult to justify. Their approach to scaling AI does not produce immediate revenue or visible outcomes. It looks like cost, and it can feel like lost momentum.
It’s the equivalent of reinforcing the foundations after the first few floors are already built. From the outside, progress appears to pause. But this step is what allows the building to continue. It can mean the difference between building a skyscraper and being forced to stop at three stories.
When this step is skipped, the need doesn’t disappear. It shows up later, when systems are already under pressure and much harder to change.
3. Scale selectively, not broadly
Once the underlying environment is stable enough, expansion becomes viable, but it needs to be controlled.
At this stage, the goal isn’t more use cases. It’s easier use cases.
That only happens if teams stop rebuilding from scratch, leveraging shared environments, consistent processes, and governance that can be reused rather than recreated.
Progress here is measured in repeatability, not volume.
4. Industrialize through repetition
At this stage, AI becomes part of how the organization operates.
New use cases are delivered into an environment that’s already designed to support them. Outcomes become more predictable. The cost and effort of scaling both stabilize.
Scale is no longer something that must be forced. It becomes a property of the system.
Where your programs are probably going wrong
Without deliberate sequencing, a familiar pattern emerges.
- Pilot success drives expansion.
- Multiple initiatives are launched in parallel.
- The underlying environment is expected to catch up.
- Instead, complexity builds, and delivery slows.
As a result, AI efforts shift from building new capabilities to managing what already exists.
By the time this pattern becomes visible, it’s already difficult to unwind.
The trade-off most banks try to avoid
Scaling AI is not just a technical challenge, but a sequencing decision.
It requires a different kind of discipline.
Leaders need to be deliberate about where they accelerate, and where they invest ahead of visible demand. They need to recognize that some of the most important work may not show immediate returns but will in fact determine whether returns are possible at all.
It also requires organizations to measure progress differently—not just in what has been delivered, but in how easily it can be repeated.
Avoiding that trade-off doesn’t remove it. It simply delays it to a point where it becomes more expensive to resolve.
The order determines the outcome
Most banks will move through these stages in some form. The question is whether they do it early, when it’s controlled, or later, when it’s disruptive.
AI rarely fails because the models don’t work. It fails when organizations try to scale before the environment is ready to support what they have already built.
And by the time that disconnect becomes clear, the cost of correcting it is significantly higher than getting the sequence right in the first place.