Andy Rangel, chief executive officer for Malaysia and Thailand at TP, believes the financial services industry is approaching a meaningful operational inflection point. His perspective is drawn from working with over 350 financial services clients globally, spanning banks, fintechs, cryptocurrency exchanges and digital asset players. Daniel Dua, vice president & head of Southeast Asia business development and BFSI SME at TP, brings a complementary regional view across banking, financial services, insurance and emerging fintech segments.

Their shared observation is that many firms are running artificial intelligence (AI) in some form across customer operations, but few have moved to fully autonomous, end-to-end agentic execution. AI-assisted servicing, workflow automation and decision support remain materially more common than full agent deployment.

Where is AI genuinely deployed in customer operations today?

Rangel's starting point is that AI delivers the strongest results where processes are well-defined, measurable and scalable. Collections, payment reminders and next-best-action recommendations are common examples. They are environments where intelligent systems can deliver customer context, surface history and recommend a response with greater speed and consistency than traditional workflows allow.

Dua noted that banks across the region remain at markedly different stages of maturity. Some are deploying internal robo-advisory tools to help relationship managers process data and generate recommendations more quickly. Others are introducing conversational interfaces to replace traditional interactive voice response systems, allowing chatbots to handle routine enquiries with greater fluency. "They need to find a sort of right fit," Dua said. That variation reflects not just technology readiness but differences in process maturity, governance capability and workforce preparedness, factors that shape outcomes as much as the AI tools themselves.

Dua also pointed to reputational risk as a constraint that banks cannot afford to underestimate. Misinformation or service inconsistency carries direct brand cost, which means agentic solutions need to operate within tightly defined guardrails.

Process design is the constraint technology cannot solve

Rangel's central argument is that banks cannot simply layer agentic AI onto existing processes and expect meaningful results. If the underlying process is broken or poorly understood, the AI will reproduce the failure at scale. "If you layer on agentic solutions to a process that's broken or maybe mishandled, you're just going to do that at a larger scale," he said.

Banks that achieve measurable improvements share foundational characteristics. Their underlying processes are well-defined. Compliance and regulatory parameters are embedded rather than assumed. Governance is built in from the outset rather than retrofitted. Critically, pilots are designed as components of a wider operating model rather than as isolated tests. "Pilots are initiatives that are built in silos and not necessarily built to scale," Rangel said.

The strategic implication is direct. AI readiness is less about technology procurement than process redesign. Institutions that succeed are those willing to do that foundational work before deploying.

What measurable improvements are banks seeing?

Rangel pointed to collections as a concrete example. TP's agentic collections solution, TP.ai FAB Collect, has delivered improvements of up to 25% in collections performance, with comparable gains in recovery and connection rates, he said.

In some deployments, delinquencies have moved from double-digit to single-digit levels. These figures reflect claims made in the interview rather than independently validated outcomes.

The mechanism, in Rangel's account, is intelligent differentiation. An agentic system can identify whether a customer needs a reminder because they are travelling, or whether a pattern of delinquency requires an altogether different approach. "How you prompt an agentic solution to reach out to a customer can be different," he said.

Dua framed the broader efficiency picture in similar terms. Across TP's client base of more than 350 banks, fintechs, cryptocurrency exchanges and digital asset firms globally, institutions typically target efficiency gains of 15% from AI deployment in customer operations. Outcomes depend on how well existing systems integrate with AI tools and how clearly the institution has defined the problem it is trying to solve. Dua argued that problem definition is where credible engagements begin, with TP working with the client on a roadmap from problem statement through to target efficiency outcomes.

Where human judgement remains non-negotiable

Rangel argued that agentic deployment does not remove the need for human judgement. It changes where that judgement is most valuable. In banking, the distinction matters because the customer relationship is fundamentally about trust.

Customers use banks to pay bills, support families and manage obligations that may be deeply personal. These are interactions that require discretion, situational judgement and an understanding of what is actually at stake for the customer.

"We're talking about someone's money, their livelihood," Rangel said. Agentic systems are well-suited to routine, replicable queries such as balance enquiries, transaction lookups and standard servicing requests, and handling those at scale frees human agents for interactions where judgement is the point. Collections illustrate that division clearly. A customer in genuine financial difficulty needs a different conversation from one who has simply forgotten a payment. An agentic system can identify the scenario. Resolution often requires human handling.

The design question, in Rangel's view, is not just whether to include human oversight, but where to position the handover and how to make it transparent to the customer. "You want to be able to merge the AI solution with a person so that at the end of the day, your brand is still protected," he said. Building those escalation paths clearly into the process architecture is what separates a credible agentic deployment from one that creates reputational risk.

For Dua, the commercial dimension reinforces the same point. AI deployment can improve customer engagement, outreach and conversion rates, he said, but it does not on its own grow or retain a customer base. In collections, the presence of human handling alongside agentic outreach is what supports retention and longer-term customer loyalty. The same logic applies in adjacent areas such as relationship management and insurance claims, where empathy is integral to the interaction and is something AI cannot replicate at this stage.

Governance, model drift and the limits of set-and-deploy

A governance dimension Rangel raised, less commonly addressed in industry discussion, is the risk that model behaviour evolves over time even when an institution has not changed its own processes. Underlying language models continue to develop, and their responses can shift in ways that are not immediately visible to the institution deploying them. This creates a compliance and operational risk distinct from questions of accuracy. The system's behaviour can change without any deliberate action by the institution.

Continuous monitoring, validation against defined benchmarks and clear accountability for model behaviour are, in Rangel's framing, not optional features of agentic deployment. They are part of the governance architecture required to sustain it. "It's more of a clinical trial of continued evolution, continued checks and balances, continued process enhancement," he said.

Building the workforce for AI-augmented operations

Rangel sees structural changes ahead in how banks organise customer operations teams. Effective agentic solutions require banking expertise and AI expertise working together within a single governance structure, not in parallel. "An AI SME alone, a banking SME alone, isn't going to build the proper process for you," he said. That integration needs to happen at the point of process design, not after deployment.

Dua argued that this reorganisation carries direct implications for existing staff. Banks need to ensure operational teams are AI-trained and AI-enabled, because processes and everything downstream of them will change. Maximising the use of AI requires a different way of working from traditional operations, and Dua sees this as a change element that has to be driven top down, into every level of the organisational structure. Rangel added that employee anxiety about AI deployment is real and often underestimated. "Transparency, communication, helping people understand how it impacts me, I think are keys that are sometimes not talked about enough," he said.

Banks that navigate this well are those that demonstrate concretely how AI creates capacity for higher-value work, such as judgement calls, better analysis and more difficult conversations, rather than simply displacing it.

The divide that will define the next phase

The meaningful divide in banking customer operations, Rangel and Dua argued, is unlikely to fall between institutions using AI and those that are not. Most banks in the region are running AI in some form. The more consequential divide is between institutions that have done the foundational work of process redesign, governance architecture, human integration and workforce transition, and those still layering pilots onto legacy processes without that foundation in place.

Dua returned to the importance of realistic sequencing. A 30% efficiency gain is a credible target, but rarely an immediate one. "We have to take steps in between," he said. The work is collaborative, moving along what Dua described as a "glad path" towards agreed targets, rather than promising an immediate end state.
The combined argument is that agentic AI in customer operations is a continuous process of governance, adjustment and evolution that has to be sequenced realistically. Rangel emphasised the architectural side. Institutions that close the gap between pilot and production will be those prepared to redesign their operating models around AI rather than waiting for the technology to make legacy processes viable. Dua emphasised execution, realistic targets and a deployment path travelled with the client rather than promised to them. The institutions most likely to succeed are those that hold both at once.