Who Takes the KPI Hit When AI Creates Value for the Bank?

An agent accelerates loan application preparation. Underwriting inherits more cases. Operations absorbs the exceptions. Technology carries the integration cost. The business expects faster growth.

The business case may look attractive at bank level while making the transition difficult for individual teams.

That is a leadership problem worth addressing before scaling agents.

The value gap is real. Its cause is contested.

Economist Impact's 2025 survey of 1,700 senior banking executives found that 50% reported limited returns from generative AI and another 11% reported none. Those results establish a value gap. They do not, by themselves, explain its cause. Model reliability, integration cost, data, use-case selection and adoption can all play a part.

Delivery is harder than the demo suggests. FDE-Gym, a recent research benchmark, evaluates agents on 95 delivery tasks adapted from 20 customer scenarios. Leading models lose substantial ground when they must discover unstated constraints. The benchmark has limits: simulated stakeholders, one minimal agent framework and no measured human baseline. But it raises a useful question for banking leaders: how much essential context still lives in people's heads, departmental practices and undocumented exceptions?

My argument is narrower than "it's all politics." Where the bottleneck is conflicting incentives or unclear decision rights, a better model alone won't remove it.

Who owns the trade-off?

Banks have not ignored accountability. The Bank of England and FCA's 2024 survey of UK financial services found that 72% of firms using or planning to use AI assigned accountability for AI use cases to executive leadership. Most reported three or more accountable people or bodies.

Multiple accountable bodies can reflect appropriate oversight. But naming accountable people is different from deciding who can resolve conflicting priorities. When the benefit lands in lending and the extra work lands in underwriting, someone has to own that trade-off.

The model is not the only constraint

The same survey identified safety, security and robustness, access to skills, and transparency and explainability as the leading non-regulatory constraints on AI adoption. That supports treating adoption as a broader engineering and organizational challenge. It does not prove that internal politics dominates. It does suggest that improving the model, on its own, is an incomplete strategy.

This is established banking governance, applied to agents

None of this requires inventing new governance. The Basel Committee's operational-risk principles already connect senior-management responsibility with adequately resourced change management and clear responsibilities across the lines of defence. They call for change assessments that consider the human resources and technology investment required, with monitoring during and after implementation.

Agents raise the stakes because they can change the volume and speed of work faster than the surrounding organization adjusts.

Follow the queue

Consider a lending agent that reduces application-preparation work but increases the number of cases reaching underwriting. Before calling that a success, the business case should account for underwriting capacity, exception handling and customer waiting time. Otherwise, a local productivity gain may simply move the queue.

The same applies to payment exceptions. An agent identifies the issue and prepares a resolution in seconds. If nobody has agreed who can authorize the action, which evidence is sufficient, or who owns the resulting risk, the case still waits.

Some approvals are necessary safeguards. Others duplicate work. Modernization requires understanding the difference.

Board, executives or steering committee?

Basel distinguishes board oversight of the operational-risk framework and risk appetite from senior management's responsibility for implementation. It also calls for independent control functions and clear authority.

That gives a precise answer to a common question. The board oversees the framework and risk appetite. Management resolves operating trade-offs within its mandate. A steering committee can make those decisions if it has delegated authority. The committee is not necessarily the problem; a committee without authority is.

The affected team needs a credible agreement

Performance management is a recognized lever for aligning individual accountability with organizational results. The US Government Accountability Office's work on organizational transformation makes that point, though it concerns public-sector organizations rather than banking AI.

Applied here: a team asked to absorb more work and a worse scorecard for someone else's benefit needs more than encouragement to embrace AI. It needs funded transition capacity, revised responsibilities, and time-limited performance adjustments with a baseline, milestones, an expiry date and stop criteria. It should not be promised protection from every adverse consequence, and risk standards should not be weakened to buy agreement. Some trade-offs remain hard even after incentives are aligned.

What to agree before scaling an agent

• Define the end-to-end outcome. Measure cost per resolved case, completion time, errors and customer impact.
• Make the economics visible. Identify where benefits accrue, where new work appears, and who funds integration, controls and transition capacity.
• Align performance expectations. Give teams credit for the enterprise outcome and agree any temporary KPI adjustments against milestones and an expiry date.
• Establish decision authority. Let teams act within approved boundaries, preserve independent risk challenge, and escalate trade-offs beyond that authority to the executive sponsor or executive committee.
• Build a repeatable path. Basel already supports documented governance, committee mandates, thresholds and a common inventory of controls. Reusable decision rules should mean that changes within approved boundaries follow an established path, while material changes receive fresh review.

That last point is an intended benefit, not a guaranteed one. But it is the difference between scaling a portfolio of agents and renegotiating the organization for every new one.

There is also evidence that governance can support speed. Economist Impact's report cites Standard Chartered's FRIDAY framework as reducing AI model review times by 60% and development turnaround by 30%. That is one example of governance supporting delivery, not proof that every control framework will produce the same result.

The strongest objection

A fair critic would say: weak AI returns are mostly a technology and data problem, and organizational redesign is a consultant's answer to an engineering question. Sometimes that's right. If the model can't do the task reliably, no incentive scheme will fix it. The point is not that the operating model is always the binding constraint. It's that when it is, no model upgrade will remove it, and most business cases don't test for it.

The bottom line

Before scaling agents, leadership must agree how value, transition costs, decision authority and performance accountability will be distributed across the workflow. Otherwise, technical success can coexist with organizational stalemate.

Before approving the next agent, ask: who benefits, who absorbs the disruption, and have we changed the incentives accordingly?

Sources

• Economist Impact, Intelligent banking: The future ahead (2025), supported by SAS, pages 8 and 15.
• FDE-Gym: The Last Mile of AI Delivery, arXiv 2610.22350: https://arxiv.org/pdf/2610.22350
• Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services 2024: https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024
• Basel Committee on Banking Supervision, Operational risk consolidated guidelines (ORR10): https://www.bis.org/committees/bcbs/basel-consolidated-guidelines/module/orr/10
• US GAO, Results-Oriented Cultures: Creating a Clear Linkage between Individual Performance and Organizational Success (GAO-03-488, 2003): https://www.gao.gov/products/gao-03-488