AI agents in banking operations: building an AI-first operating model

Banking has spent a decade increasing technology budgets, yet productivity in operations has not improved in line with that investment.

To support an AI-first retail bank, you now need AI-first banking operations with AI agents embedded in the core workflows, not just better interfaces on top of the old processes.

Why operations must change

Retail and universal banks face a structural squeeze.

• Revenue pools are still growing, but costs and regulatory overhead are growing faster.
• Many traditional banks run with cost-to-income ratios well above those of digital leaders.
• Customers expect near-instant decisions and transparent processes, even in complex processes like onboarding, lending, and financial crime investigations.

Past waves of digitisation focused on channels and partial automation. They improved individual steps but left the underlying operating model largely unchanged. That is why tech spend has risen without a corresponding step-change in operational productivity.

From responding to solving: chatbots vs AI agents

Most banks' first experience with AI in operations has been chatbots and simple automation.

• Chatbots answer FAQs and deflect basic queries.
• Robotic process automation automates isolated back-office tasks.

They do not, however, take real ownership of outcomes. They were designed to respond, not to solve.

AI agents are different. In the emerging industry language, an agent is a software entity that can:

• Observe its environment: data, events, and user input.
• Reason and plan: break down a goal into steps.
• Act with tools: call systems, trigger workflows, and update records.
• Check and adjust: check each result and change the plan when the case changes.

In banking operations, that means AI agents can complete end-to-end tasks and processes, such as:

• Handling a credit-limit increase from request to decision to confirmation.
• Running a customer onboarding process from data capture through checks and account opening.
• Managing a collections interaction from outreach through negotiation and payment plan setup.

The basic unit of work shifts from individual tickets to outcomes.

A concrete illustration: one request, no queue

Consider a routine request: a customer wants a credit limit increase on their card.

In a traditional operation:

• The customer waits in a queue.
• An agent opens multiple systems (core banking, CRM, credit engine, KYC tool).
• They manually check criteria, make a decision, generate a document, send an email, and update records.

This approach is slow, repetitive, and expensive, and it ties up skilled staff on work that largely follows standard policies.

In an agentic operating model, the same request is handled differently:

1. An AI agent engages the customer, authenticates them, and understands the request.
2. A supervisor capability breaks the goal into steps: verify identity, review account behaviour, assess affordability and risk, decide, communicate, and log.
3. Specialist capabilities collect the necessary information from existing systems and apply the bank's rules.
4. If the criteria are met, the agent prepares the decision and a person approves it. The agent then confirms it to the customer in the same interaction, generates any required documentation, and updates the records.
5. If something is unusual or high-risk, the AI routes the case to a human with full context and a recommended action.

From the customer's perspective, there is one interaction and a clear outcome. From the bank's perspective, the work that used to traverse multiple queues is resolved by an intelligent layer that follows existing policies.

What an AI-first operating model looks like

We describe an AI-first operating model in operations in terms of how work is managed, not in terms of specific systems.

How work is orchestrated

In the target state, work is organised around outcomes:

• A process-level orchestrator is responsible for seeing each request through end-to-end, for example, "complete this onboarding" or "resolve this card request within policy."
• Under that, specialist capabilities apply risk rules, perform checks, or assemble documentation at specific decision points.
• Beneath those, micro-tasks carry out focused actions such as extracting information, checking a single condition, or logging an outcome.

This mirrors how strong operations teams already function today: clear ownership of the end-to-end outcome, experts for complex parts of the workflow, and well-defined tasks for routine steps. AI agents take on these roles in software; the management logic stays recognizable to an operations leader.

Where this starts in banking operations

While the pattern can eventually touch most of operations, we see banks typically start in a small number of high-impact areas:

• Customer onboarding and servicing, where fragmented processes are highly visible to customers.
• Lending and credit operations, where decision speed and consistency directly affect growth and risk.
• Collections and financial crime operations, where intelligent workflows can materially improve productivity and case quality.

In these areas, the aim is to:

• Move from fragmented, ticket-based work to end-to-end, outcome-driven flows.
• Free human capacity from routine processing to higher-value decisions and interactions.
• Build a reusable intelligent layer that can extend to adjacent processes over time.

EY estimates that generative AI could raise productivity in Indian banking operations by up to 46 percent by 2030. (Source: EY, March 2025)

Implications for operations leaders

For a COO or head of operations, the shift to AI-first operations is primarily about redesigning work, not about chasing specific tools.

In an AI-first model:

• Frontline teams spend less time on routine, rule-based processing and more time on exceptions, complex cases, and customer relationships.
• Operations leaders decide which processes should be supported end-to-end by AI, which should remain predominantly human-led, and where human checkpoints add the most value.
• New capabilities emerge around supervising the intelligent layer, refining workflows, and collaborating with risk and product teams on where to increase or limit autonomy.

The aim is to handle more volume with the same team and respond faster, while people stay responsible for the decisions that matter most.

The decision in front of you

Over the next few years, the competitive line in banking operations will not fall between "AI" and "no AI". It will fall between:

• Banks that build an AI-first operating model with AI agents orchestrating key operations processes, and
• Banks that remain in a world of tickets, queues, and fragmented automations.

The technology and patterns for agentic AI in banking operations are now mature enough to deploy. The practical questions for operations leaders are:

• Which processes you transform first.
• How much of each process you are comfortable handing to AI agents.
• How you organise your people and processes around this new way of working.

At ISTINA AI, our focus is helping banking operations leaders answer those questions and move from vision to execution, without requiring them to become AI engineers.

FAQ: AI agents in banking operations