The next workforce problem for banks may have nothing to do with finding enough people. It may be figuring out how many artificial workers they already have. An AI assistant answering employee questions is relatively easy to identify, but an agent checking documents, another monitoring transactions, another preparing credit analysis and several more operating quietly inside technology supplied by external vendors are much harder to see as a single workforce. Individually, each system may have an owner, a purpose and an approval. Collectively, they are beginning to create something banks have never had to manage before: a digital workforce that can expand far faster than the governance structures designed to supervise it.
AI governance was built for a different generation of technology
Much of the banking industry’s existing AI governance architecture was developed around models with relatively narrow and predictable purposes. A fraud model scored transactions, a credit model assessed risk and a recommendation engine identified products that might be relevant to a customer. Banks could document the model, validate its performance, establish ownership and periodically review whether it continued to behave as expected.
Generative AI complicated that framework because its outputs became less deterministic. Agentic AI takes the challenge considerably further. An AI agent can potentially receive an objective, determine the steps required to achieve it, interact with several systems, retrieve information and make decisions along the way. Multiple agents can also collaborate, allowing one system to initiate an activity that another system completes.
The governance question therefore moves beyond whether a model produces an accurate answer. Banks increasingly need to understand what an AI system is allowed to do, which systems it can access, what information it can retrieve, which decisions it can influence and where human approval remains mandatory.
That distinction becomes increasingly important as banks move beyond pilots. As Banks Have Plenty of AI Ideas. Getting Them Into Production Is the Hard Part. examined, financial institutions are discovering that the difficult part of AI adoption is not necessarily finding promising use cases. It is integrating those capabilities safely into complicated banking environments. Once AI begins participating directly in operational processes, governance stops being primarily a model-management exercise and starts becoming part of the bank’s infrastructure.
Every bank may eventually need an AI inventory
Banks know how many employees they have and, at least in principle, what those employees are authorised to do. Departments have reporting structures, employees have job descriptions and identity-management systems determine which applications and information each person can access. When responsibilities change, permissions can be adjusted. When somebody leaves, access can be removed.
Something similar may eventually be necessary for AI.
A bank operating hundreds or potentially thousands of AI-powered processes will need more than a traditional register of algorithms. It may require a continuously updated inventory showing every AI system operating across the organisation, who owns it, what purpose it serves, what information it can access, which applications it can communicate with, what decisions it can influence and how much autonomy it has been given.
The challenge is that many of these capabilities may not originate inside the bank. AI is increasingly being embedded into cloud platforms, cybersecurity products, customer relationship management systems, payment infrastructure, compliance applications and everyday productivity software. A financial institution could therefore acquire additional AI capabilities simply because an existing technology provider updates its product.
This creates the possibility that a bank’s effective digital workforce could expand without the institution deliberately launching new AI projects. Keeping track of AI may consequently become as much a vendor-management problem as an internal technology-management problem.
The digital workforce will not come with an HR department
Human employees operate within organisational structures that have evolved over decades. Someone approves their recruitment, determines their responsibilities, grants access credentials and reviews their performance. There are established procedures for changing roles and removing access when someone leaves.
AI agents do not naturally fit into those structures, yet their capabilities may evolve just as significantly.
Consider an agent introduced to assist a commercial banking team. Initially, it might simply summarise customer information and prepare meeting notes. Over time, additional integrations could allow it to retrieve financial statements, analyse company accounts, prepare credit documentation, identify risk factors and recommend lending conditions.
Technically, the bank may still regard it as the same AI application. Operationally, however, its role has changed substantially.
Banks will therefore need mechanisms for recognising when an AI system has accumulated enough capability to require a new level of oversight. Approval cannot simply be something that happens when an application is first deployed. Governance will need to follow the evolution of the system throughout its operational life.
Permissions may matter more than intelligence
Much of the industry’s attention is currently focused on increasingly powerful AI models, but the sophistication of a model is only one part of the risk equation. A highly capable AI agent with almost no access to operational systems may pose relatively limited risk. A less sophisticated agent that can retrieve customer information, modify records or initiate payment instructions could create far greater consequences if its permissions are poorly designed.
This is why identity and access management could become one of the foundations of agentic banking. AI systems may increasingly require their own identities and credentials rather than relying on broad application-level permissions. Banks could then establish precisely what each agent can read, modify, recommend or initiate.
A customer-service agent, for example, might be able to retrieve account information but never change it. A payments agent could prepare an instruction but be prevented from authorising the transaction. A credit agent might analyse financial information and make a recommendation while final lending authority remains with a human.
The issue becomes even more complicated when agents begin interacting with other agents. If one system delegates a task to another, banks will need to determine whether permissions travel with the request or whether every agent operates strictly within its own authority. Without clearly defined boundaries, chains of automated actions could gradually move beyond the authority originally intended when the first agent was deployed.
Banks will need to know what their AI did yesterday
Governance is not only about controlling what an AI system can do. Banks must also be able to reconstruct what it actually did.
Financial institutions already maintain extensive audit trails for human and system activity. When an employee accesses an account, changes customer information or approves a transaction, those activities can generally be traced. Agentic systems will require comparable, and potentially more sophisticated, records because a single outcome could involve several autonomous decisions across multiple systems.
Imagine a customer-service agent identifying unusual activity and referring it to a fraud agent. That system requests additional verification from another service, which subsequently triggers a temporary restriction on the customer’s account. If the customer challenges the decision, the bank needs to understand the entire sequence: what initiated the process, which information each agent considered, what decisions were made and why the final action occurred.
This is where the governance of AI begins to overlap with cybersecurity. The Next Cybersecurity Arms Race Will Move at Machine Speed explored how automated systems could increasingly operate on both sides of the security equation. Banks will therefore need monitoring capabilities capable of observing machine activity at the same speed at which that activity occurs.
AI may eventually need managers of its own
One of the more significant consequences of this transition could be organisational rather than technological. Banks may eventually develop roles responsible for supervising groups of AI systems in much the same way managers currently oversee teams of employees.
These managers would not necessarily build models or write software. Their responsibility could be ensuring that AI systems continue operating within defined objectives, permissions and risk limits. They might review exceptions, investigate unexpected behaviour, approve changes in authority and determine when a process requires greater human involvement.
A lending operation, for example, could eventually contain specialised agents responsible for documentation, verification, analysis, monitoring and customer communication. Instead of employees manually completing every stage of those processes, a smaller number of people could supervise the systems performing them and intervene where judgement or accountability requires human involvement.
That represents a fundamentally different interpretation of automation. AI would no longer simply be software used by employees to work faster. It would become an operational participant that itself requires supervision.
Third-party AI makes the problem considerably harder
Banks already depend heavily on external technology providers, and AI could deepen those relationships. Many financial institutions will use models, platforms and agents developed by vendors rather than building everything themselves. The bank may control what information an external service can access while having much less visibility into how the underlying technology evolves.
A model upgrade could change behaviour. A vendor could introduce additional agentic capabilities into an existing product. Another technology provider further down the supply chain could become part of the process without being immediately visible to the end user. The result is an AI governance problem that extends well beyond the organisational perimeter of the bank.
This makes AI governance increasingly inseparable from third-party technology risk and operational resilience. Banks will need to understand not only the artificial intelligence they intentionally deploy but also the intelligence embedded within the platforms and infrastructure on which they depend. The inventory of a bank’s digital workforce may ultimately extend across much of its technology supply chain.
The winning banks may not be those with the most AI
There is an understandable temptation to measure AI progress by deployment numbers. Banks can count use cases, pilots, employees with access to generative AI tools and processes that have been automated. Those metrics can demonstrate activity, but they may reveal relatively little about whether an institution has actually developed the ability to operate AI safely at scale.
A more meaningful measure may eventually be the amount of autonomous activity a bank can confidently manage. An institution operating hundreds of poorly governed agents could be less advanced than one running a smaller number with clearly defined identities, permissions, monitoring and accountability.
This could change the competitive dynamics of banking AI. Institutions that invest early in the control architecture surrounding autonomous systems may ultimately be able to deploy them faster because every new agent does not require an entirely new governance process. Identity, permissioning, monitoring and auditability become reusable infrastructure.
The race, therefore, may not be about which bank can put AI into the most places first. It could be about which institutions can build the governance machinery that allows them to keep adding AI without losing visibility or control.
What it means for the industry
- AI inventories could become core banking infrastructure. Banks will increasingly need continuously updated records of AI systems, agents, owners, permissions and dependencies across their organisations.
- Identity management will expand beyond human employees. AI agents are likely to require distinct digital identities and carefully controlled access privileges as they gain the ability to interact directly with banking systems.
- Permissions may become as important as model performance. Governance will increasingly focus on what an AI system can actually do rather than simply whether its outputs are accurate.
- Audit trails will need to follow machine-to-machine activity. Banks must be capable of reconstructing decisions that move through several interacting agents and systems.
- Human management will evolve rather than disappear. Some banking roles could increasingly involve supervising digital workers, reviewing exceptions and determining where autonomous systems require human intervention.
- AI maturity will be measured by control as well as deployment. Banks capable of governing large populations of autonomous systems may ultimately scale AI faster than institutions focused primarily on launching more use cases.
