Why Most Bank AI Strategies Stall Before They Reach the Operations Floor

An AI business case promises hours back. A year later, two questions decide whether it worked: “Did cost per case fall?” and “Where did the released capacity go?” Answering either requires a measurement most AI programs never took.

nCino’s first AI in Banking Benchmark, fielded in late March and early April 2026 among 150 senior US banking technology and business decision-makers, found 21% tying AI investment to revenue growth and 26% tying it to cost reduction. Eight in ten said they are prioritizing adoption over return. Cornerstone Advisors, surveying 416 executives at institutions mostly in the $250 million to $50 billion asset range, found bank deployment of generative AI had tripled in a year, to 49%.

The industry has stopped debating whether to deploy. It has not started reporting what deployment returned.

The data banks are confident about is not the data the business case rests on

The nCino benchmark holds one more detail worth pausing on. Asked about access to good quality data, 87% of those executives were confident. Asked about governance, 93% reported at least one problem and 52% pointed to data sitting in silos.

Both numbers are true. The space between them is where the money goes missing.

A bank holds a great deal of data about its systems. Transaction records, workflow logs, case statuses, cycle times. That data is clean, governed and well plumbed, and the confidence in it is earned. Far fewer hold a measured record of where their people’s hours go. The checking. The rework. The chase for a missing document. The manual preparation that keeps a case moving when the system will not take it. Little of that reaches a log, because little of it touches the system.

An AI business case is written to leverage the second kind of data and targeted using the first. Aimed with one measurement, judged with another. The mismatch usually surfaces only when the savings fail to show up in anyone’s cost line.

Ron Shevlin at Cornerstone Advisors was blunt about the forecasts driving all this adoption. McKinsey’s projection that AI automates half of banking tasks, Accenture’s 20% to 30% cut in operating costs: he called them “wildly, overly optimistic projections,” while allowing the direction is right. His conclusion was that there is no AI strategy without a data strategy underneath it. That is often read as a data warehouse problem. A cleaner warehouse improves what the models see. It does not, by itself, account for hours that were never recorded anywhere.

For community and regional banks, the gap costs real money. Leaders are absorbing rising transaction complexity and tighter service expectations on flat or shrinking headcount, while the AI program meant to help sits somewhere upstream of the floor.

None of this reflects a shortage of ambition. Budgets are approved, policies written, vendors evaluated, and governance forums established. The breakdown happens after approval, when the strategy fails to change how a manager handles workload, staffing, and service levels on a Tuesday. Three patterns show up repeatedly.

The roadmap is built for the boardroom, not the work queue

AI roadmaps are organized around use cases. Fraud detection, credit decisioning, document processing. Each is technically sound. But a use case is not an operating model. It says nothing about how a team leader picks which of forty open items to work next, how spare capacity in one team reaches a backlog in another, or how a manager knows by 10 a.m. whether today’s staffing survives a volume spike.

Here is what that looks like on the floor. A US card disputes team was meeting its Regulation Z deadlines. What the Reg Z clock could not tell them was whether the work was spread evenly across the people doing it. It was not. The team covered two verticals with nobody skilled across both, so work backed up on one side while capacity sat on the other. Nothing in the case system showed it, because the case system counted cases.

Once the hours were measured, the team priced a cross-skilling plan, forecast a 5% monthly productivity gain, and wrote the investment into its weekly plans. Productivity rose 17% in three months and 33% over six. Volumes then climbed 15% to 20% during the shift to home working, and the flexibility they had bought held every service level and every regulatory deadline.

No model was involved. A measurement was, and it paid for a training investment that a spreadsheet of case counts would never have justified.

Managers know where the work is piling up and cannot move it

This is the pattern most often misdiagnosed. Banks answer it by buying dashboards, and dashboards are not the constraint. A manager who watches a team fall behind with no way to move capacity toward it has been handed a sharper account of a problem she still cannot solve.

Knowing where the hours go and being able to redirect them are two different capabilities. The second is the one that turns an hour released by automation into an hour redeployed somewhere it earns.

The ActiveOps Platform is built for the second. It sets capacity against demand, helps managers weigh their redeployment options, and flags cases drifting toward a breach while there is still time to recover them. The decision to move the work stays with the manager. What changes is that she has a basis for making it and a record that it happened.

AI governance and operations are working from different numbers

The two functions often report through different lines, meet on different cadences, and run on different data. That produces more than duplication. It produces inconsistent decisions, because two teams looking at the same backlog are working from incompatible counts of it.

The consequence is concrete. A governance forum counting cases closed and an operations team counting hours worked will estimate the same automation opportunity differently, and the initiative gets scoped against whichever number reached the paper first. An AI program layered on top of that inherits the inconsistency rather than resolving it.

The pilots that go quiet had no starting point

Put the three together and the sequence is familiar. Promising pilot, clean proof of concept, executive sponsorship, then a long silence. MIT’s Project NANDA reported in 2025 that 95% of enterprise generative AI pilots produced no measurable P&L impact. The figure has been contested on methodology, and it is an industry report rather than peer-reviewed work, but the mechanism has held up better than the precision of the number. In many of those programs no starting point had been measured, so there was nothing for the result to be measured against.

In a bank there is a second reason baselines matter. A capacity gain that is not accepted by internal audit accept is not a gain, it is an assertion. A measurement taken before the change, from production data, held to the same definition afterward, gives risk, audit, control and finance a common basis for validating the benefit. Programs that skip it end up defending their results in the committee that approved them.

Baselines also move the target. CIBC Mellon baselined its fund accounting operation before choosing reconciliation work for automation, using activity-level effort data captured where the work was performed. That data put 40% of fund accountants’ time into reconciliation, more than 10,000 hours a month. Roughly 75% of that reconciliation effort sat in three types, Cash, Asset and Capstock, together approximately 30% of total daily fund accounting work. None of the three on its own would have been considered a good candidate for automation. The concentration was not what the process documentation implied, and it was the targeting decision rather than the technology decision that set the return.

What effort data will not tell you

It does not judge the quality of a decision. A judgment call that took four minutes and one that took forty are both recorded as effort, and neither is marked right or wrong.

It does not trace a transaction across systems. That is process mining, which reconstructs where the record went. Effort data measures where the people went, including work that never touched a system. Most automation programs need both, and many banks hold only the first.

It measures work, not people. Volumes, effort, and available capacity with role-based access. Individual-level reporting exists for the managers who need it, and each bank sets its own policy on how that is used. The operations that get the most from it say what is being measured and why before it goes live.

Six questions to ask before the next AI business case is signed

The nCino number to hold on to is the 87% who are confident about their data. These test what that confidence rests on, and they are worth asking before a vendor arrives to ask something easier.

Are you confident that… 

1. …you know where the hours actually go in the operation you are about to automate, rather than where the process documentation says they go? 

2. …the work you have selected is where the effort concentrates, and not simply where the automation case was easiest to write? 

3. …your baseline was measured before the program started, from production data, rather than reconstructed afterward from what the systems happened to log? 

4. …you would know what a 20% capacity gain is worth to you in FTE or in cost per case, and would name the teams it came from? 

5. …your managers have a way to move that capacity to the work that needs it inside the day, and not just a report telling them it exists? 

6. …the gains from last year’s program are still in the operation, and that you would evidence that to a board, an auditor or a regulator? 

If the honest answer to any of them is no, the constraint on the next AI dollar was never the AI. It is everything standing between the strategy and the floor.

Where to start

None of this requires a finished AI roadmap. It requires knowing where the hours go before the money is committed and holding that measurement steady while the program runs and after the program team has gone. Capacity released in month three and reabsorbed unnoticed in month nine delivered nothing, and without a baseline there is no way to tell which happened.

That measurement is what ActiveOps does, across more than 100 enterprise operations in regulated industries.

Call to action, linked to activeops.com/platform: See what the ActiveOps Platform measures, and why system data does not contain it.

Sources

nCino, AI in Banking Benchmark, 2026. Fielded late March to early April 2026, 150 senior US banking technology and business decision-makers.

Cornerstone Advisors, What’s Going On in Banking 2026. 416 executives, institutions mostly in the $250 million to $50 billion asset range. Ron Shevlin quotation from the same report.

MIT Project NANDA, 2025. Industry report, not peer reviewed, contested on methodology.

CIBC Mellon, ActiveOps and Duco, Intelligence and Reconciliation Optimization in Investment Operations, March 2026.

Card disputes figures from a published ActiveOps customer case study, available at activeops.com/more-stories.

Hear from your Peers…