Why banks will fail to capture the financial value of AI unless they redesign the work around it
By Nibras Adambawa, Founder and Principal Advisor, Limiere
Artificial intelligence is beginning to produce impressive productivity numbers across banking. Processing times are falling, analytical work is accelerating, employees are completing tasks with less manual effort, and functions that once required large teams are discovering that the same volume of work can be handled with considerably less human intervention.
The temptation is to translate those gains directly into financial value.
If an AI-enabled process reduces the time required for a task by 30 percent, the business case often assumes that 30 percent of the associated labour cost has effectively been saved. If a copilot releases several hours of employee capacity each week, that capacity is multiplied by fully loaded compensation and presented as productivity value. If an operations function can theoretically perform the same work with 200 fewer full-time-equivalent employees, the corresponding payroll number begins to appear in the return calculation long before 200 salaries have actually disappeared from the income statement.
The arithmetic is easy. The economics are not.
A bank can become significantly more productive without becoming proportionately more profitable because productivity measures how efficiently work can be performed, while financial value depends on what the institution subsequently does with the capacity that has been released. Unless management changes the operating model, reduces or avoids cost, redirects capacity into higher-value activity, improves risk outcomes, or creates additional revenue, a large proportion of the apparent productivity gain may never reach the financial statements.
That distinction is becoming one of the most important questions in AI valuation.
The problem facing bank CEOs and CFOs is therefore not simply whether AI is making employees more productive. It is whether the institution has built the organisational and financial mechanisms required to convert that productivity into an economic outcome that can be measured, attributed and defended.
The productivity gain is only the beginning of the valuation
The financial pathway from AI productivity to realised value contains several stages, and each one can absorb part of the benefit before it reaches the P&L.
An AI intervention may reduce the time required to complete a task, increase the number of cases an employee can handle, or eliminate a portion of manual activity altogether. At that point the bank has created productivity, but what it actually possesses is released capacity. The economic value of that capacity remains unresolved until management determines whether it can be captured and what the institution intends to do with it.
This distinction is particularly important in banking because labour economics are not infinitely divisible. An employee whose workload falls by 15 percent does not cost the institution 15 percent less. A twenty-person team that becomes 20 percent more productive does not automatically become a sixteen-person team. Regulatory minimums, service-level requirements, organisational structure, peak-volume capacity, control responsibilities and role indivisibility can all prevent theoretical productivity from being converted directly into reduced expenditure.
The problem becomes even more pronounced when the productivity improvement is distributed thinly across a large workforce. Saving forty minutes per week for ten thousand employees sounds economically substantial when the hours are multiplied by loaded labour cost, but unless those fragments of time can be consolidated, redirected or incorporated into a redesigned operating model, the value may remain largely theoretical.
This is where many AI business cases confuse capacity value with financial value.
The capacity has value because the institution can potentially do something economically useful with it. What cannot be assumed is that the value has already been realised merely because the capacity exists.
Productivity creates potential value. Conversion creates financial value.
The executive implication is significant. A bank should not accept an AI productivity claim as an economic return until management can identify the mechanism through which the released capacity will become financially consequential.
The most common productivity equation is incomplete
The shortcut is familiar:
Hours saved × fully loaded labour cost = AI value.
The calculation can be useful, but only if it is described accurately. What it estimates is the theoretical value of capacity released, not necessarily the financial benefit realised by the institution.
Suppose an employee with a fully loaded annual cost of $100,000 becomes 20 percent more productive because an AI tool removes repetitive analytical work. It would be convenient to declare that the bank has created $20,000 of annual value. In reality, the salary remains $100,000 unless something changes in the economic use of that employee.
If the released capacity allows the institution to remove or not replace a role, there may be a genuine cost benefit. If the employee uses the capacity to manage more customers or originate additional business, the value may appear through revenue rather than expense. If the time is redirected into better collections activity, the value may appear through lower credit losses. If the employee simply absorbs additional internal activity that was previously deferred, the bank may become operationally healthier without producing an immediate financial benefit at all.
The same productivity number can therefore support very different valuations depending on what happens after the time is released.
A rigorous cost-side valuation would have to consider not only the amount of capacity created, but the proportion that can realistically be captured, the financial mechanism through which that capacity will be converted, the timing of that conversion, and the incremental operating costs required to sustain the AI intervention.
Revenue, risk and quality benefits require their own financial pathways. They should not be treated as labour savings simply because the originating intervention improved productivity.
This distinction will become increasingly important as AI enters professional and analytical work. Earlier banking technologies often automated highly standardised processing activities where the path to efficiency could eventually be expressed through centralisation, lower staffing intensity or changes in distribution. AI is now reaching credit analysis, risk, finance, compliance, customer management and other roles whose economic contribution is less easily reduced to headcount. The original historical pattern remains relevant: banking repeatedly captured the full economics of technology only after redefining the work around it, rather than simply placing new technology on top of old roles.
Released capacity has more than one economic destination
The next analytical mistake is to treat all productivity as though it should eventually reduce operating expenditure.
In practice, AI-generated capacity can create value through at least four economically distinct pathways, and each requires a different valuation logic.
The first is cost removal, which is the most visible and easiest to reconcile. If the same output can be produced with structurally fewer resources, and those resources genuinely leave the cost base through attrition, role consolidation, outsourcing reduction or workforce redesign, the value eventually appears as lower operating expense and potentially improved cost-to-income ratio.
The second is cost avoidance, which is often economically substantial but easily misunderstood. A growing bank may deploy AI and retain the same number of people while handling significantly greater transaction or customer volume. Current expenditure has not fallen, but the institution has avoided the additional headcount that would otherwise have been required. That is genuine economic value because the cost trajectory has changed, although it should not be reported as current-period cost removal.
The third is capacity redeployment. AI may release time from administrative or analytical work and allow employees to devote more attention to activities that carry greater economic value. Relationship managers may spend more time with customers, collections teams may increase contact intensity, credit officers may focus on complex cases, and analysts may perform deeper work without increasing headcount. The resulting value may appear through revenue, retention, recovery rates, customer economics or better decisions rather than through payroll.
The fourth is quality and risk improvement. A bank may deliberately retain the existing workforce while using AI to increase consistency, reduce errors, improve detection, strengthen decision quality or accelerate intervention. In such cases the economic benefit could appear through lower operational losses, better credit outcomes, reduced remediation, improved pricing or stronger control performance.
These four outcomes should not be collapsed into one productivity figure because their financial pathways, evidentiary requirements and timing are different.
The same hour released can have very different economic values depending on what the institution does with it.
For the CFO, this taxonomy matters because the confidence attached to each pathway will differ. Cost removed from the general ledger carries a different evidentiary weight from future cost avoidance. Revenue attributed to redeployed capacity requires a different counterfactual from payroll reduction. Risk benefits need to be separated from cyclical or portfolio effects before they can be credited to AI.
A useful AI valuation therefore does not ask merely how much capacity has been created. It asks where that capacity is going and what evidence demonstrates that the chosen destination is economically real.
The operating model is the conversion mechanism
The historical development of banking technology provides an important lesson. ATMs did not create their full economic value merely because machines could dispense cash. The value emerged as banks changed branch roles, servicing patterns, staffing structures and customer behaviour around the new capability. Digital banking followed the same pattern. Moving transactions online created potential efficiency, but institutions captured the deeper economics only when they altered distribution, servicing, process architecture and the role of physical infrastructure.
AI will require another operating-model transition, although this one is likely to reach further into analytical and professional work. The old article correctly recognised that AI is beginning to alter roles closer to the core of how banks create and protect value, including credit, risk, treasury and other decision-support functions. The economic implication is that value realisation can no longer be treated simply as a workforce-reduction exercise.
A bank may install highly capable AI into an existing organisational structure and achieve meaningful productivity gains while capturing only a fraction of the available economics. Managers may continue to staff teams according to historical workload assumptions. Role descriptions may remain unchanged even though the underlying tasks have shifted substantially. Approval layers, service models, outsourcing arrangements and performance measures may continue to reflect the pre-AI operating environment. In such circumstances, productivity becomes trapped inside the organisation.
This is why the operating model should be regarded as the conversion mechanism between AI capability and financial value.
Capturing that value may require changes in staffing ratios, spans of control, process ownership, workflow sequencing, service design, outsourcing, role architecture and the allocation of work between humans and machines. It may also require management to distinguish carefully between capacity that should be removed and capacity that should be preserved because it can generate more attractive economics elsewhere.
The objective should not be maximum workforce reduction. It should be maximum economic conversion.
AI can release capacity faster than the institution can economically absorb it.
That observation has an important strategic consequence. The bank that deploys AI most aggressively is not necessarily the bank that captures the most value. The advantage may belong to the institution that is most effective at redesigning its operating model around the productivity AI creates.
An illustrative banking case
Consider a banking operation employing 1,000 people at an average fully loaded annual cost of $80,000 per employee. The current annual workforce cost is therefore $80 million.
Assume that an AI programme reduces the human effort required across the relevant workflows by an average of 20 percent. A simplistic productivity calculation would identify 200 FTE-equivalent units of released capacity and attach $16 million of annual value to the programme.
That is not yet a valuation. It is an estimate of theoretical capacity.
Suppose management then examines how much of the capacity can actually be converted. Some of the released capacity is required to accommodate expected business growth without increasing headcount. Some remains trapped because activities cannot be consolidated cleanly across roles and locations. A portion can be removed over time through natural attrition and process consolidation, while another portion can be deliberately redeployed into activities expected to improve collections performance and customer retention.
The resulting economics might look very different from the initial $16 million headline.
For illustration, assume the bank concludes that 50 FTE-equivalent units can be removed from the current cost base, another 60 represent future hiring that can now be avoided, 40 can be redeployed into activities with an independently assessed revenue or risk benefit, and the remaining 50 cannot yet be monetised because of operating-model constraints. The AI platform, infrastructure, controls and support also carry a continuing annual cost that must be recognised.
The theoretical capacity remains 200 FTE-equivalent units, but the value is now distributed across realised cost reduction, avoided future expenditure, redeployed economic capacity and unrealised potential. Each component should be measured differently, recognised at the appropriate time and assigned an evidentiary confidence consistent with what management can actually demonstrate.
The example is intentionally illustrative, but it reveals why the phrase “AI saved 200 FTEs” is financially ambiguous. It could mean the bank removed 200 salaries, avoided hiring 200 people in the future, released work equivalent to 200 people without changing the cost base, or produced some combination of all three. Those are economically different outcomes and should never carry the same valuation.
Illustrative only. Released capacity and financial savings are not the same measure.
The discipline matters because valuation should make uncertainty visible rather than manufacture certainty. Current cost removed may be directly observable. Hiring avoidance depends on a credible baseline for future staffing requirements. Redeployment value requires evidence linking the changed allocation of effort to an economic outcome. Capacity that cannot yet be monetised should remain exactly that, potential value rather than realised value.
The four numbers every Executive Committee should demand
The most practical change a bank can make is to stop allowing AI productivity to be expressed as a single number.
Every material productivity case should present four separate measures.
The first is gross productivity created, which measures how much time, effort or processing capacity the technology has technically released. This is normally the number most easily observed and the one most likely to appear in programme reporting.
The second is releasable capacity, which asks how much of that productivity can actually be freed by the operating model after role indivisibility, workload peaks, service requirements, control obligations and workflow constraints are considered.
The third is converted value, which identifies the economic use management has assigned to that releasable capacity, whether through cost removal, cost avoidance, redeployment or improved risk and quality.
The fourth is realised value, which measures what has actually appeared in the financial performance of the institution and can be supported by evidence.
The distance between those four numbers is where much of the current ambiguity surrounding AI productivity resides.
A productivity claim is incomplete until all four numbers are visible.
For CEOs and CFOs, this creates a much more useful conversation with technology, operations and business leadership. Instead of debating whether a reported productivity percentage is impressive, management can determine how much of that productivity is economically accessible, what action is required to convert it, when the value should become visible and how confidently it can be attributed to AI.
The same framework also improves capital allocation because future AI proposals can be assessed against the institution’s demonstrated ability to convert productivity rather than against theoretical benefit alone. A bank that repeatedly generates large capacity estimates but converts little of them should adjust the value assumptions used in subsequent investment cases. Conversely, an institution with a strong operating record of capturing and redeploying AI-generated capacity may justifiably attach greater confidence to future productivity claims.
Over time, the conversion rate itself becomes a management capability.
The competitive divide will be in conversion, not adoption
The banking industry is likely to produce increasingly impressive AI productivity statistics over the coming years. Models will become faster, agents will undertake more work, employees will complete tasks with less effort and institutions will publish ever larger estimates of hours saved.
Those numbers will matter, but they will not tell us which banks are actually capturing the greatest economic value.
The deeper competitive distinction will be between institutions that create productivity and institutions that convert productivity.
The first group may deploy sophisticated technology and generate substantial amounts of theoretical capacity while allowing much of the value to remain trapped inside legacy roles, processes and organisational structures. The second will treat productivity as the beginning of a financial process, redesigning the operating model deliberately so that released capacity becomes lower cost, avoided future cost, additional revenue, better risk outcomes or some other measurable improvement in institutional economics.
For a bank CEO, that distinction should materially change how AI productivity cases are evaluated.
Before approving an investment whose value proposition depends on workforce productivity, management should be able to explain not only how much time the technology is expected to save, but what happens to that time once it has been saved, which management actions will convert it into economic value, where that value should appear in the institution’s financial performance, how long the conversion will take, and what evidence will eventually determine whether the original investment thesis was correct.
Until those questions are answered, the institution has estimated productivity.
It has not yet valued it.
