Banks have always been complicated organisations, but the next generation of banking could introduce a fundamentally different management challenge. A growing share of work will no longer be performed exclusively by employees using software, but by a combination of people, AI agents, automated decision systems, third-party platforms and algorithms capable of completing increasingly complex tasks with limited human intervention. That changes more than productivity. It changes what managers are responsible for. A banking executive who once supervised teams, processes and technology may increasingly oversee an operating environment in which machines perform work, make recommendations, communicate with other systems and sometimes initiate actions themselves. As banks push further into agentic AI, the difficult question may therefore shift from how many jobs AI can automate to something more fundamental: how do you manage a bank when an increasing proportion of its workforce is no longer human?
The Traditional Management Model Was Built Around People
Most banking operating models were designed around a relatively straightforward division of responsibility. People performed work, technology helped them perform it, managers supervised the people and organisational structures determined who was accountable when something went wrong. Even highly automated systems generally operated according to predefined rules, making it possible to trace a process back to the business unit responsible for it.
Agentic AI begins to disturb that arrangement because the technology can take on more of the work itself. An AI agent might gather information from several systems, analyse a customer case, decide what information is missing, communicate with another system, prepare a recommendation and route the case to the next stage. As the technology becomes more capable, the human employee may move from performing each step to supervising the outcome or intervening only when an exception occurs.
Deloitte argues that this difference matters because earlier automation generally followed predetermined rules, while agentic systems can pursue objectives and adapt their behaviour according to context. It says banks therefore need to think beyond individual agents and design an “Agentic Operating Model” in which people, processes, controls, data, technology and evidence are considered together.
That sounds like a technology architecture problem, but it is equally a management problem. The more work that moves into autonomous systems, the less useful it becomes to think about management purely in terms of the number of employees reporting to an individual.
A Manager Could Eventually Supervise People and Machines
Consider a future financial crime operations team. Today, a manager might oversee analysts investigating alerts generated by transaction-monitoring systems. In a more agentic environment, one AI agent could assemble customer information, another could analyse transaction relationships, another could research counterparties and a further system could prepare the initial investigation narrative. Human investigators might concentrate on ambiguous cases, escalations and decisions carrying greater regulatory consequences.
The manager is still responsible for the operation, but the nature of that responsibility has changed. Performance can no longer be understood simply through employee productivity and case volumes. Management may need visibility into how accurately agents perform, how frequently humans override them, where they fail, whether their behaviour changes over time and whether automated decisions remain within approved boundaries.
This is no longer entirely theoretical. Deloitte reported earlier this year that around one-third of financial institutions were allocating budgets specifically to agentic AI and that many were creating roles to supervise its use. It cited Wells Fargo’s work with Google Agentspace on custom AI agents and PNC Financial’s use of orchestrated agents in developing its mobile banking experience as examples of how agent-based systems are moving into financial institutions.
As adoption increases, banks may consequently need to treat AI systems less like ordinary software applications and more like operational participants with defined responsibilities, permissions, performance expectations and escalation rules.
Organisational Charts May Stop Showing How Work Really Gets Done
There is another complication. AI agents do not necessarily respect the departmental boundaries around which banks have traditionally organised themselves.
A lending process, for example, can involve customer onboarding, identity verification, credit analysis, fraud detection, pricing, compliance, documentation and disbursement. Those responsibilities may sit across several departments, yet an agentic workflow could move between the underlying systems almost instantly.
That creates a potential gap between organisational ownership and operational reality. Technology might own the AI platform, the lending business might own the product, risk might define credit policies, compliance might establish regulatory controls and data teams might manage the information feeding the models. If an autonomous workflow produces an inappropriate outcome, identifying who actually owns the problem can become considerably more difficult.
EY has argued that banks moving towards agentic models will need an orchestration layer capable of enforcing guardrails and maintaining auditability across machine-executed decisions. EY Deloitte similarly argues that named individuals need practical mechanisms to oversee, challenge and constrain agents operating under their responsibility rather than allowing accountability to disappear into the technology.
For bank management, this suggests that deploying AI agents without redesigning responsibility around the process could eventually create an organisation in which work moves faster than accountability.
Management Could Become More About Orchestration
This may gradually change what it means to be a manager inside a bank. Traditional management involves allocating work, monitoring performance, coaching employees, resolving problems and escalating decisions. Some of those activities remain essential, but AI introduces another responsibility: deciding which work should be performed by people, which can be delegated to machines and where the two should interact.
A manager overseeing an AI-enabled operation may need to determine the autonomy permitted for different activities. One agent might be allowed to gather information but not make decisions. Another could make low-risk decisions within predefined thresholds. Higher-risk actions might require human approval, while unusual cases could automatically be escalated.
Management therefore becomes partly an exercise in designing the boundary between human and machine authority.
This is particularly important in banking because efficiency is not the only objective. Controls that appear to slow a process can exist precisely because the consequences of an error are significant. Deloitte has warned that banks should not automatically remove what it describes as “useful friction” when redesigning processes around agentic AI, because existing controls can preserve accountability, safety and trust rather than merely compensate for inefficient human workflows.
The challenge for managers will be knowing which friction AI should eliminate and which friction the bank still needs.
Performance Management Will Need New Measures
Banks also know how to measure human work relatively well. Operations teams can track cases completed, error rates, service levels, customer outcomes, revenue performance and compliance breaches. AI workers introduce different questions.
If an agent completes 50,000 tasks but humans override 15% of its recommendations, is it performing well? What if its productivity improves while the quality of its reasoning deteriorates? What if an agent remains accurate in normal conditions but behaves unpredictably during unusual market events? And how should management compare the cost of an automated process with the operational risk created by greater autonomy?
These questions suggest that banks may eventually develop something resembling performance management for autonomous systems. Measures could include accuracy, escalation rates, human overrides, control exceptions, decision consistency and the ability to reconstruct how an outcome was reached.
The scale of this challenge could itself drive further automation. Deloitte argues that if organisations eventually deploy thousands or tens of thousands of agents, traditional governance and risk-assessment processes could become a bottleneck, requiring parts of AI governance itself to be automated while exceptional cases remain subject to human review.
The paradox is significant: banks may eventually need AI to help manage the growing population of AI operating inside the bank.
Middle Management Could Change Rather Than Simply Disappear
Much of the workforce debate around AI assumes that automation will reduce management layers because machines can coordinate information and monitor processes more efficiently. Some of that may happen, particularly where management roles depend heavily on reporting, work allocation and information aggregation.
But greater autonomy can simultaneously create new supervisory responsibilities. A smaller human team operating a highly automated banking process may require fewer conventional supervisors while placing much greater responsibility on the managers who remain.
PwC’s 2026 Financial Services Workforce AI Survey illustrates the scale of the workforce transition financial institutions are contemplating. Nearly eight in ten financial-services leaders surveyed expected their workforce to shrink by at least 20% over the following five years, yet only 42% said they had conducted high-level modelling across their organisation to understand how AI would change labour capacity. PwC argues that reducing headcount and designing an AI-enabled workforce are fundamentally different exercises. PwC
That distinction could prove important for banks. Removing people from an existing operating model does not automatically create an AI-native operating model. Someone still has to redesign the work, determine machine authority, establish escalation mechanisms and understand where human judgement remains necessary.
Banks May Need Managers Who Understand More Than Their Own Department
The traditional banking career path often produces deep functional specialists. Credit executives understand lending, operations leaders understand processes, technologists understand systems and risk professionals understand controls. AI increasingly connects those disciplines.
A manager responsible for an AI-enabled lending process does not necessarily need to become a machine-learning engineer, but they may need enough understanding of models, data, technology and automation to challenge the systems operating within their business. At the same time, technology leaders building those systems need sufficient understanding of banking regulation and risk to appreciate why certain decisions cannot simply be optimised for speed.
That could increase the value of managers capable of operating across organisational boundaries. The important skill may not be knowing how to build an AI model, but knowing enough about the business, technology, data and controls to recognise when the system is behaving in a way the bank should not accept.
Deloitte’s recent research into AI governance across global banks points in a similar direction. It argues that AI should be treated as a transversal risk driver rather than an isolated technology issue and describes a movement from having a “human in the loop” for individual decisions towards having humans “on the loop”, supervising increasingly automated systems at a broader level.
That is a considerably different management responsibility from approving individual transactions one by one.
Accountability May Become the Hardest Part to Automate
Ultimately, the most difficult management question may arise when something goes wrong. If an employee makes an inappropriate lending decision, banks generally know where responsibility sits. If an AI agent produces the analysis, another agent retrieves the data, an automated workflow applies the policy and a human manager supervises the overall process, responsibility becomes less obvious.
Regulators are unlikely to accept an AI system itself as the accountable party. Deloitte notes that existing obligations around customer outcomes, resilience and accountability continue to apply when agentic systems are introduced, meaning banks must be able to demonstrate how autonomous decisions were made, what constrained them and who remained responsible.
This could make accountability one of the defining human responsibilities in an increasingly automated bank. AI may perform more analysis, execute more processes and recommend more decisions, but institutions will still require people with sufficient authority and understanding to take responsibility for the systems operating underneath them.
The result may be a bank with fewer people performing routine work but a more demanding management environment for those who remain. Executives will be responsible not simply for employees and technology, but for an operating system in which humans and autonomous machines continuously interact.
Banks have spent decades learning how to manage people and technology separately. Agentic AI is beginning to merge the two. The institutions that benefit most may not necessarily be those that deploy the largest number of AI agents, but those that learn how to manage a workforce in which the definition of a worker itself has changed.
What this means for the industry
- Bank management will increasingly include machine supervision. Managers may become responsible for the performance, permissions and escalation behaviour of AI agents alongside human employees.
- Organisational ownership will need to follow workflows, not just departments. Agentic processes can cross technology, operations, risk and compliance boundaries, making clear end-to-end accountability increasingly important.
- Performance metrics will have to evolve. Banks may need to measure AI accuracy, overrides, exceptions, escalation behaviour and decision quality alongside traditional employee and operational metrics.
- Middle management may be redesigned rather than simply removed. AI can automate coordination work while simultaneously creating new responsibilities around orchestration, exceptions and autonomous systems.
- Cross-functional banking knowledge could become more valuable. Managers will increasingly need enough understanding of technology, data, risk and the underlying banking process to challenge automated systems effectively.
- Accountability remains fundamentally human. As machines perform more of the work, banks will still need identifiable people capable of explaining, constraining and taking responsibility for automated outcomes.
