Transaction volumes in Thailand today are five to ten times what they were five years ago, according to Charamporn Jotikasthira, executive director of Bangkok Bank. Every coffee purchase, scan at a point-of-sale terminal and digital payment now leaves a behavioural trace. Banks know more about their customers’ financial lives than at any previous point, but many still run processes designed for a world in which that data did not exist.

Many banks have invested in data platforms, analytics teams and generative artificial intelligence (AI) tools, yet still route decisions through operating models built for sparse information, delayed intervention and specialist hand-offs. AI agents, or systems that perform defined tasks inside a business process, will not create full value if banks use them only to speed up legacy work.

Charamporn used collections to illustrate the point. "In the past, if you asked about the skills of the collections department, their skill was one-to-one calling," he observed. "They only went after the customer when the loan went bad." The data was thin, so the process was reactive. A customer stopped paying, the account moved into arrears and the bank responded.

That information environment has changed, but the collections process has not kept pace. Charamporn said Bangkok Bank’s teams can now begin to identify stress signals within 30 days rather than waiting until a loan is 90 days in arrears. That shift changes collections from a late-stage recovery function into an early-warning and customer-intervention capability. Banks capture the value of richer data when they redesign skills, tools and decision paths around earlier action.

AI adoption stalls when institutions apply new tools to old processes. Value depends on rebuilding work around agents, human judgement and governance from the start.

A loyalty workflow shows the operating model shift

Peng Chen, chief executive officer of Minden.ai, a Singapore-based coalition loyalty platform, said its programme with more than 200 partners generated campaign requests that could overwhelm sequential workflows. Under the original model, a partner request travelled from account manager to marketing team to external agency before a creative was produced and approved.

The cycle took weeks because each step depended on one specialist group handing work to another. A first generation of AI adoption shortened that cycle by replacing the agency with an internal marketing team using generative AI tools.

Peng framed that first phase as necessary but limited. "That is still the first step," he noted. "What I call using new technology to fit the old work process." The larger shift came when the marketing team stopped producing every creative output and started building agents instead. It built a copywriting agent and a creative agent, then gave account managers the capability to produce materials directly during client conversations.

The relevance for banks lies in the operating model rather than the loyalty use case. Relationship managers, collections officers, branch teams and product managers often depend on central analytics, marketing, risk or operations teams before they can act on customer information. The example illustrates a shift from central production to governed frontline capability. "Instead of being a production house, they are now an enablement house," Peng said.

Minden.ai applied the same principle to analytics. Rather than routing data requests through a lean central team, account managers now use AI-enabled query tools built by the analytics team to query the database directly. The data-to-insight cycle that previously took a week can now complete within the conversation. Banks face a similar challenge when hand-offs prevent customer-facing teams from acting on relevant data quickly and responsibly.

The new workflow changes the role of people

Once organisations redesign work around agents, they also change the meaning of roles. Peng said many employees try AI, expect complete automation, then disengage when it does not fully solve the problem. "The technology today doesn't solve 100% of your problem."

The more realistic target is human-AI collaboration rather than a machine that works alone. "If AI can solve 70% of your problem, the target operating model is to achieve human-AI collaboration." The remaining 30% of human judgement still matters, but the employee no longer starts from zero.

That distinction is important in banking, where judgement, accountability and customer context remain central. A collections officer cannot outsource a treatment decision to an agent, and a relationship manager cannot rely blindly on an AI-generated recommendation. Employees will need to manage, test and refine agents, which means banks must build skills in agent supervision, process design and risk-aware frontline decision-making.

Peng said the message he wanted to transmit was not enthusiasm but obligation. "Not trying is not an option."

At Bangkok Bank, the scale of the challenge is larger and the adoption model has to be more deliberate. With 18,000 employees, Charamporn described a top-down education programme targeting the N-minus-1 and N-minus-2 leadership layers, meaning executives one and two levels below the chief executive. The immediate objective is not value measurement, but ensuring that senior leaders understand the opportunity before they are asked to pursue it. "If they don't drive or set the goal for their organisation, it will not move."

Boards must govern AI as operating risk

Charamporn said the Bank of Thailand, the country’s central bank and banking regulator, issued AI governance guidelines placing accountability at board level. That shifts AI from a management implementation issue to a board-supervised operating-risk issue.

Charamporn’s framing was direct: "If your brakes are not as good as your engine, you're not going to reach the destination in one piece." The passenger in that vehicle is the customer. Hallucination errors, wrong credit decisions, poor customer treatment and compliance failures can create reputational, legal and conduct consequences. Boards therefore need to understand AI well enough to challenge management on risk appetite, model oversight, data use, audit trails, escalation rules and customer outcomes.

The governance question becomes more complex as AI agents are embedded across an institution. Customer-facing systems, credit decisioning support, collections processes, fraud operations and AI assistants for relationship managers all create different forms of risk. Every application must be checked against regulatory expectations, and every step in an AI-enabled workflow must be broken down and assessed. The balance between speed and control is not a technology preference; it is a board responsibility.

Saj Kumar, regional business leader for manufacturing at Microsoft Asia, noted that industries where AI errors can damage equipment, products or people still require human oversight. "We always have a human in the loop when it comes to manufacturing decisions," Kumar noted. Banks face a different version of that risk through customer harm, unfair treatment, financial loss and regulatory breach.

The next advantage will come from redesigning the bank

Charamporn closed with a warning about competitive risk. The processes that banks run today were built around the constraints of an earlier technology environment. Those constraints shaped what institutions assumed was possible, and those assumptions became embedded in operating models, skill sets and organisational structures.

When the constraints change, the assumptions do not automatically update. A competitor that recognises the change and rebuilds around the new environment does not merely improve incrementally. It operates from a different set of possibilities. Charamporn’s advice to chief executives was to identify which assumptions underlying their current business model are now obsolete and act before a competitor does.

For Bangkok Bank, the assumption under revision is that credit risk management begins when a customer stops paying. Transaction data has made earlier intervention possible. Rebuilding the collections process to act on that data within 30 days is not simply an AI project. It reflects a change in the information environment and in the skills, controls and customer treatment strategies needed to operate in it.

Across the bank, the next phase of AI adoption will require a systematic redesign of operating models. The starting point is not the tool, but the decision path, including where a signal is detected, who has authority to act, what evidence is required, which controls apply and when the board needs visibility. Banks that answer those questions function by function will be better placed to move agents from experimentation into production without creating unmanaged risk.

Treating AI agents as part of institutional design would change how banks organise central teams, frontline authority and governance. It would require clearer decision rights, guardrails for staff acting on AI-supported insights and controls that scale with adoption. Institutions that make those choices early will not simply automate existing processes. They will reset how the bank senses risk, serves customers and allocates accountability in a more data-rich operating environment.