As artificial intelligence becomes a material capital allocation decision, banks need to separate the creation of AI value from the judgement of what that value is worth.
By Nibras Adambawa, Founder and Principal Advisor, Limiere
Artificial intelligence entered banking through technology budgets, innovation programmes, transformation agendas and individual business use cases, so it was entirely natural that the people closest to those programmes also became the people responsible for explaining their value. Technology vendors produced benefit cases around the capabilities they were selling, implementation partners quantified expected efficiencies, internal AI teams measured adoption and operational improvement, and strategy functions assembled those inputs into investment narratives for management and boards.
That arrangement was understandable while AI remained largely experimental. It becomes increasingly difficult to defend once the scale of investment becomes material and the question before management changes from whether the technology works to whether the economics justify further capital.
Banking has spent decades developing institutional separation around precisely this type of problem. Credit is originated by one part of the organisation and challenged by another because the party seeking growth cannot also be the sole authority on the risk attached to that growth. Models are developed and then independently validated because technical competence does not eliminate the possibility of optimism, error or incomplete assumptions. Acquisitions are subjected to valuation and due diligence beyond the enthusiasm of the transaction sponsor, while financial statements move through assurance processes because management assertion and independently tested evidence are not regarded as equivalent.
The principle is deeply embedded in how banks operate: when the financial consequence of a decision becomes significant, the judgement applied to that decision cannot rest entirely with those responsible for originating, advocating or delivering it.
Artificial intelligence has not yet developed an equivalent discipline.
In many institutions, the same broad ecosystem that proposes an AI initiative also helps construct its economic case, implements the programme, measures its success and provides much of the evidence later used to justify expansion. That does not make the evidence unreliable, nor does it imply poor conduct by the people involved. It does, however, create an institutional weakness when those inputs are treated as sufficient evidence of financial value without an independent valuation discipline between delivery and the next capital decision.
That is the gap banks now need to close.
The issue is not integrity. It is the architecture of incentives.
The debate about independence is often weakened by the assumption that it is fundamentally a question of conflicts of interest in the ethical sense. That framing is unnecessarily crude. The more important issue is that different participants in the AI ecosystem operate under different economic and institutional objectives, all of which can be legitimate while still producing different interpretations of the same evidence.
A technology vendor is commercially rewarded when its platform is adopted more widely. A systems integrator participates economically when implementation scope expands. A consulting firm with downstream transformation or managed-services capabilities may benefit when an initial strategy engagement develops into a broader programme. Internal AI and technology functions, quite properly, seek the resources, organisational mandate and executive sponsorship required to develop institutional capability. The original governance-era argument made this point clearly: the structural tension remains even when every person involved is acting professionally and in good faith.
The CEO and CFO, however, are solving a different problem. Their responsibility is not to maximise deployment, adoption or programme scale, but to determine whether the financial return produced by AI is sufficient to justify the capital, operating expenditure and organisational capacity being committed to it.
Those objectives may frequently align, but alignment should be demonstrated rather than assumed.
A vendor may be entirely correct that its platform has improved processing speed by 20 percent. An internal programme team may be entirely correct that employee adoption has exceeded expectations. A business unit may be entirely correct that an underwriting model has improved predictive accuracy. The difficulty begins when those operating achievements are converted into financial value without adequately testing the assumptions that connect one to the other.
A reduction in processing time becomes a financial benefit only if the released capacity can be monetised through lower cost, greater throughput, higher revenue or some other economic mechanism. Improved credit prediction creates financial value only to the extent that it changes approval quality, portfolio composition, default formation, pricing, provisioning or capital consumption in a way that can be separated from other changes occurring at the same time. Higher adoption creates value only if adoption changes behaviour or productivity sufficiently to affect the economics of the institution.
The operating evidence can therefore be entirely accurate while the financial conclusion attached to it remains overstated, understated or simply unproven.
This is why the party supplying evidence should not automatically become the party whose judgement defines the valuation.
AI value is unusually vulnerable to mismeasurement
The case for independence becomes stronger when the nature of AI value itself is examined. Unlike many conventional investments, the financial return from AI rarely appears as a direct consequence of the technology. It emerges through a chain of operational and financial effects, each of which introduces assumptions that have to be tested.
An AI intervention first changes something technical: prediction quality, processing speed, decision accuracy, automation or the quality of an interaction. That technical improvement must then alter an operating outcome inside the bank. The operating outcome must affect a financial lever, such as revenue, operating expense, credit loss, pricing, RWA or capital consumption. Only after that effect reaches the income statement, balance sheet or capital position can management begin to determine what the intervention was economically worth.
Every transition introduces the possibility of leakage, delay or misattribution.
A contact-centre system may release thousands of employee hours without reducing payroll because the capacity is redeployed rather than removed. That may still be valuable, but its value cannot honestly be represented as a corresponding reduction in operating expense. An AI underwriting model may improve default prediction while the observed improvement in portfolio losses is also being influenced by changes in the macroeconomic environment, customer mix, risk appetite or collections strategy. A productivity tool may appear to produce substantial efficiency gains during a pilot but require additional infrastructure, controls, licences and support costs when deployed across the enterprise.
The central valuation problem is therefore not merely to identify benefit. It is to establish what portion of the benefit can reasonably be attributed to AI, how much of it is financially realisable, how persistent it is likely to be, and what remains after the continuing economics of operating the technology are recognised.
That requires a counterfactual, not merely a before-and-after comparison. It requires Finance to distinguish gross benefit from realised value. It requires assumptions to be made explicit rather than buried inside a business case. It requires uncertainty to be reflected in the valuation rather than disguised by a precise number that the evidence has not earned.
For a bank Executive Committee, this distinction is critical because false precision can be more dangerous than acknowledged uncertainty. A return estimate expressed to two decimal places can look financially rigorous while resting on assumptions that would not survive even moderate challenge. The purpose of valuation is not to make an investment appear measurable; it is to determine how much financial confidence the institution can reasonably place in the number.
Banking already knows why independent challenge matters
There is nothing novel about the principle Limiere is arguing for. What is new is its application to artificial intelligence.
Banks already separate the creation of financial exposure from the judgement of that exposure because the industry has learned, often through painful experience, that institutional optimism must be counterbalanced by independent challenge. The original article made the same argument in the context of governance, observing that the quality of oversight is inseparable from the independence of the input informing it. The same logic applies even more directly when the issue becomes valuation.
A business unit may understand a credit opportunity better than anyone else in the organisation, but that does not make it the sole authority on risk. A model developer may understand the architecture and performance of a model better than anyone else, but that does not eliminate the need for independent validation. A transaction sponsor may have the strongest strategic understanding of an acquisition, but that does not make its valuation immune from challenge.
The institutional distinction is important because closeness to an investment produces information advantage and potential bias at the same time. Banking does not resolve that tension by excluding the people closest to the decision. It resolves it by using their knowledge as an input while creating a separate point of judgement capable of challenging the conclusion.
AI valuation should develop in the same way.
The delivery ecosystem should supply the evidence. Vendors should explain what the technology is capable of and what has been observed. Internal teams should provide operating data, cost information and model performance. Business leaders should explain the commercial mechanisms through which value is expected to arise. None of those perspectives should be discarded.
What is missing is a distinct financial discipline that asks whether the combined evidence supports the valuation being claimed.
That discipline should be capable of challenging the counterfactual, testing attribution, distinguishing capacity creation from cost removal, examining whether benefits survive at scale, incorporating the full economics of ownership and determining whether the result remains attractive after uncertainty is recognised.
Most importantly, it must be capable of arriving at any conclusion without its own commercial economics changing materially as a consequence.
An independent valuer should be just as comfortable concluding that an initiative deserves significantly more capital as concluding that it should be redesigned, delayed or discontinued. Independence becomes meaningful only when none of those outcomes is commercially preferable to the party producing the judgement.
This is where externality and independence must also be distinguished. An advisor can sit outside the bank and still have a material economic interest in the programme continuing if valuation leads into implementation, technology selection, transformation or managed services. The relevant question is not whether the advisor is external, but whether the advisor’s economic model is neutral to the conclusion.
That is the standard Executive Committees should apply.
The capital case should remain alive after approval
There is a second institutional change that follows from this argument. Valuation should not appear for the first time after a programme has already been designed, funded and implemented.
By then, much of the capital allocation decision has already occurred, significant sunk costs may have accumulated and internal sponsorship has had time to deepen. The risk is that the post-implementation exercise becomes an attempt to justify a decision already made rather than a genuine reassessment of whether the original economic thesis has been realised.
A stronger approach treats the AI valuation as a living capital case.
Before investment is approved, management should be explicit about the financial variable the initiative is intended to affect, the causal pathway through which that effect is expected to occur, the evidence required to demonstrate attribution and the level of return necessary for the programme to earn further capital.
After deployment, those same assumptions should be revisited against observed evidence.
If the original investment case depended on reducing operating expense, then post-deployment success should eventually be assessed against the cost base rather than replaced by a new metric such as hours saved. If the case depended on reducing credit losses, then management should examine whether the expected loss improvement materialised and whether the result can reasonably be separated from changes in portfolio mix or the external environment.
The capital case should not disappear once the cheque has been written.
This continuity would represent an important step in the financial maturation of AI inside banks because it would connect initial expectation, realised evidence and subsequent capital allocation through one coherent economic record. It would also make it considerably harder for organisations to substitute operational progress for financial performance when the original investment thesis begins to weaken.
The next stage of AI maturity will be financial
The first phase of AI adoption in banking was necessarily concerned with capability. Institutions needed to experiment, understand the technology, build infrastructure, develop skills and identify applications that could operate safely at institutional scale.
The next phase will be defined less by whether banks can deploy AI and more by whether they can distinguish economically consequential deployment from activity that merely looks advanced.
That distinction will matter increasingly to CEOs because the opportunity cost of capital committed to AI will become more visible as portfolios grow. An institution with hundreds of initiatives but no defensible understanding of their financial contribution may be technologically active while remaining economically uncertain. Another institution with fewer initiatives, but a much stronger view of where AI produces superior risk-adjusted returns, may ultimately capture significantly more value because its capital is being directed with greater discipline.
The executive question therefore needs to become more demanding.
Management should be able to explain what an AI investment was expected to produce when capital was approved, what the evidence now demonstrates that it has produced, how much of that result can reasonably be attributed to the intervention, how confident the institution is in the valuation, and whether the resulting economics justify allocating the next unit of capital in the same direction.
If the answer depends primarily on the same parties whose institutional or commercial interests favour continued deployment, the bank has not yet completed the valuation process. It has developed an investment case, supported by programme evidence, but it has not subjected that case to the independent financial judgement routinely demanded elsewhere in banking.
Artificial intelligence is now important enough for that distinction to matter.
Banks should continue to rely on vendors, technology teams, implementation partners and internal sponsors for the expertise required to make AI work. Those parties will remain essential to capability and execution. What should change is the assumption that the same delivery chain should also be the final authority on what the resulting investment is financially worth.
Banking already understands the principle. The institution that originates an exposure should not be the sole judge of its risk, the party developing a model should not be its only validator, and the sponsor of a transaction should not determine its value without challenge.
AI capital should now be held to the same standard.
Limiere is a pure-play AI valuation advisory built exclusively for banking, with no vendor relationships, no implementation revenue and no downstream commercial interest in whether an AI investment is expanded, redesigned, deferred or stopped. Its role is to provide the independent financial judgement between AI delivery and the next capital decision.
Explore the AI Valuation Method
