Monday, 14 September 2026
Banks are beginning to move artificial intelligence beyond systems that answer questions, generate documents or predict outcomes towards technology capable of taking actions inside financial processes. At Ai4 2026 in Las Vegas, banking and technology executives described an emerging model in which AI agents could analyse information, determine the next step in a workflow and carry out routine actions without waiting for an employee at every stage. The discussion, however, exposed an important distinction between the industry’s ambitions and its immediate reality. The agentic bank is beginning to emerge, but financial institutions remain reluctant to give AI unrestricted authority over consequential decisions involving customers, money and regulatory obligations.
Ash Kaduskar, Head of AI and Advanced Analytics at First Citizens Bank, described the objective as making the organisation faster and more intelligent while redesigning how work is performed. The distinction between conventional automation and agentic AI is important. Traditional systems generally follow predetermined rules, while agentic systems potentially have greater freedom to interpret information, determine what needs to happen next and execute a sequence of tasks. That makes them particularly interesting for banking processes containing large quantities of unstructured information that previously required employees to read documents, transfer information between systems and make repeated assessments.
Third-party risk management provides a useful example. Before a bank introduces a new supplier, extensive information may need to be collected about security, compliance, financial stability and operational risk. Suppliers can face hundreds of questions, while specialists across technology, legal, cybersecurity and risk subsequently review their answers. Generative AI could retrieve information from existing documentation, prepare responses and carry out much of the preliminary assessment, allowing employees to concentrate on exceptions rather than reviewing every routine item manually. The significance is not simply that individual tasks become faster. Processes that previously moved between multiple departments over several months could potentially be compressed substantially if AI handles information gathering, classification and initial assessment continuously.
The same principle can be applied across controls, governance, compliance and other internal processes. These areas may become some of the earliest candidates for agentic AI precisely because the desired outcomes can often be defined relatively clearly. Information must satisfy specific requirements, controls must be present and exceptions must be escalated. First Citizens’ approach is therefore initially concentrated heavily on internal processes rather than placing autonomous agents directly in front of customers. This is an important counterweight to the popular idea of an autonomous bank where AI independently manages customers’ financial lives. The near-term transformation is likely to be much less visible, with customers experiencing faster onboarding, quicker responses and fewer administrative delays while much of the change happens behind the interface.
The central question is how much decision-making authority should be transferred. The panel broadly favoured automating preparation and routine assessments while preserving human intervention for exceptions and high-impact decisions. An agent could collect information, determine whether requirements have been satisfied and recommend that a process continue, but a person might still approve the final action when financial, regulatory or customer consequences become significant. Autonomy therefore becomes a spectrum rather than a simple choice between humans and machines. The value comes from removing unnecessary human involvement from thousands of routine steps without removing human accountability from decisions where judgement remains important.
That distinction also exposes one of the risks of agentic systems. An agent capable of adapting to previous decisions could potentially begin reproducing patterns that were never intended to become formal policy. Systems therefore need monitoring, defined permissions and audit trails so that changes in behaviour can be detected. One anecdote discussed during the panel involved an AI system that allegedly became increasingly willing to approve certain loan cases after observing that human loan officers usually accepted similar recommendations. The example should not be treated as a documented banking incident without independent evidence, but the underlying governance problem is credible: systems that learn from historical human decisions can reproduce existing patterns or develop unintended shortcuts unless their permitted actions remain constrained.
For regulated institutions, governance therefore cannot be added after an AI application has been built. Kaduskar argued that there is effectively no AI use case valuable enough to justify creating a serious regulatory or internal audit problem. Risk, compliance and legal teams need to become partners in development rather than departments asked to approve a completed system at the end. Banks also need standardised methods for assessing generative AI so that applications can pass through repeatable controls rather than requiring the organisation to reinvent governance for every project. Once an institution has clearly defined what data an agent can access, which decisions it can make, what evidence must be retained and when a human must intervene, additional applications can potentially be deployed more quickly. Governance effectively becomes infrastructure.
This also changes the economics of AI investment. Banks initially experimented with large numbers of small applications because generative AI was new and organisations wanted employees to explore what it could do. Attention is now shifting towards larger processes capable of generating measurable financial value. Kaduskar described these as organisational highways: workflows passing through multiple departments where redesign can produce substantially greater returns than automating isolated individual tasks. A manual activity performed once a month may not justify AI, while another performed hundreds of times every day could represent an obvious candidate.
The commercial test remains traditional despite the new technology. An AI investment ultimately needs to increase revenue, reduce costs or prepare the organisation for future competitive requirements. Deploying an agent simply because agentic AI has become fashionable does not create value. This also changes the debate around AI operating costs. Token consumption and computing expenses become a problem when companies cannot demonstrate what they receive in return. If an agent materially reduces processing costs, accelerates revenue or enables substantially more business to be handled without equivalent increases in staffing, its computing expense becomes easier to justify. The important metric is therefore not how much AI an organisation consumes, but the economic output created by that expenditure.
Banco Azteca’s perspective added the revenue side of the equation. Beyond reducing processing times, AI agents could eventually use customer and transaction data to identify financial needs that traditional segmentation misses. An institution could potentially determine that a customer needs a product it does not currently offer, identify a more appropriate credit facility or recognise an opportunity to retain a customer before that person moves elsewhere. Agentic systems could therefore evolve from executing existing banking processes towards helping institutions identify new products and commercial opportunities. The objective would shift from simply making the bank more efficient towards using AI to increase sales, deepen customer relationships and identify previously invisible demand.
Customer interaction could also change substantially. Instead of navigating through an application, selecting a product and completing a series of screens, customers could increasingly interact with a conversational financial assistant. The agent might understand an instruction, identify the relevant banking process and coordinate the necessary systems in the background. Banking would become less dependent on navigating individual applications and more focused on expressing an intended outcome. The customer may not even know that multiple agents are handling identity checks, risk assessments, product selection and administrative processes behind the conversation.
Yet the workforce consequences could be considerably larger than the change in customer interfaces. If agents take responsibility for collecting information, making routine assessments and coordinating processes, many administrative roles will change. Panel participants generally framed this as increasing employee productivity rather than simply eliminating positions, with people moving towards exceptions, supervision and higher-value decisions. Banks frequently need to process more business without increasing staffing at the same rate, making AI potentially valuable as a way of increasing organisational capacity rather than purely reducing headcount.
Kaduskar nevertheless raised a more difficult question: what happens to jobs whose value has historically depended heavily on possessing information that AI can now retrieve instantly? Knowledge itself becomes less scarce when employees throughout an institution can access specialised information through AI. Expertise will increasingly need to involve judgement, experience, accountability and the ability to determine what should be done with information rather than simply knowing where to find it. Banks will consequently need workforce strategies alongside their technology strategies, including retraining employees whose existing responsibilities become increasingly automated.
There is also a competitive dimension. Large banks can invest heavily in proprietary platforms, specialist AI teams, cybersecurity, model testing and governance infrastructure. Smaller institutions may have considerably less capital available for experimentation. If agentic AI materially reduces operating costs, improves customer acquisition or enables faster product development, the technology could widen the gap between institutions able to build sophisticated internal capabilities and those dependent largely on third-party platforms. At the same time, external AI services could eventually give smaller banks access to capabilities that previously required substantial internal technology teams, meaning the competitive outcome is not yet predetermined.
For the banking industry, the immediate future therefore appears more controlled than the term autonomous AI might suggest. Agents are likely to receive progressively greater authority inside narrowly defined processes while people retain control of exceptions, final approvals and decisions carrying substantial financial or regulatory consequences. As systems prove reliable, that boundary may gradually move. The larger transformation is that software is beginning to move from supporting decisions towards participating in them, changing the governance question fundamentally. Banks no longer need only to ask whether an AI system produced the correct information. They increasingly need to know what the system did, why it was permitted to do it, which rules governed the action and who remains accountable when something goes wrong.
The agentic bank is therefore unlikely to arrive through a single autonomous system suddenly taking control of an institution. It will emerge gradually as thousands of individual decisions and administrative steps move from employees towards governed AI agents. The competitive advantage will not necessarily belong to the bank that gives machines the most independence. It may belong to the institution that determines most precisely which decisions should be automated, which should remain human and how the two can operate together at scale.
Source: CIJ.World Research & Analysis Team
