Join our community and be part of the conversation.
To subscribe, simply enter your email address on our website or click the subscribe button below. Don't worry, we respect your privacy and won't spam your inbox. Your information is safe with us.
By Deepak Ramanathan, Vice President, Customer Advisory, SAS Asia Pacific
The rise of Generative artificial intelligence (AI) – which can produce, summarise, explain, and even show a spark of creativity – felt like the pinnacle of innovation. But Agentic AI is taking things a step further. These are systems that don’t just think; they sense, decide, act, and improve on their own.
For banks, the implications are enormous. AI agents could soon handle routine compliance checks, optimise trading strategies in real time, and deliver personalised customer experiences at a scale no human team could match.
These developments arrive at a critical moment. More than half (53-percent) of Asia Pacific CEOs believe their business models may not survive the coming decade. For these leaders, agentic AI isn’t just a technology upgrade; it’s a potential lifeline. In fact, McKinsey estimates it could generate up to USD340 billion (SGD450.7 billion) in annual value for global banking alone.
Yet there is a paradox sitting at the heart of the region’s AI ambitions. While over 60-percent of banks in Asia Pacific cite product innovation and efficiency as their top goals for AI, nearly half (47-percent) of organisations across the region are living with a trust gap: a disconnect between how much they believe in their AI systems and how reliable those systems actually are. Confidence is running well ahead of readiness.
Speed without trust is a risk banks cannot afford. AI agents operate in one of the world’s most regulated, trust-sensitive sectors. Decisions made without human oversight – be it loans, fraud alerts, or customer data – carry real consequences for real people. That is why governance must grow alongside capability. The question is no longer just how to build powerful AI agents, but how to ensure they remain explainable, auditable, and aligned with the organisation’s goals.
Getting the foundations right
Digital transformation has a sobering track record. A study by McKinsey and Oxford shows that up to 70-percent of long-term initiatives in the banking and finance sector fail. With agentic AI still in its early stages, banks have a rare window to build differently and more wisely, learning from past stumbles rather than repeating them.
That starts with understanding how these systems actually work. An AI agent doesn’t simply produce answers. It perceives, thinks, decides, acts, and learns. Each phase matters:
1. Perception: Taking in the World
Every AI agent begins by gathering information — from user inputs, transactions, databases, or live signals. This data shapes everything that follows. In fraud detection, it is what lets the system register transaction details like the amount, location, time, and device, all in the moment. Feed it poor or incomplete data, and every decision downstream suffers.
2. Cognition: Making Sense of It
This is where the AI “thinks.” It spots patterns, identifies anomalies, and weighs what the data means. In fraud detection, cognition is the system asking: does this transaction look normal? It checks history against known fraud signals and flags anything out of place — a sudden large transfer, a login from an unfamiliar country.
3. Decisioning: Choosing a Path
Here, the AI acts on what it has understood: approve, flag, or block. Getting this right is critical. A wrong call can freeze out a genuine customer or let fraud slip through undetected. Strong decisioning blends business rules, analytical models, and governance guardrails, not raw computing power alone.
4. Action: From Insight to Impact
Decisions mean nothing without follow-through. Action is where AI delivers tangible value: blocking a suspicious transaction, alerting a customer, or notifying the fraud team. But this phase only works when actions are clearly tied to business goals and customer trust. Autonomy without accountability is a liability.
5. Learning: Improving with Every Experience
Unlike traditional software, AI agents evolve. A false fraud flag today teaches the system to be more precise tomorrow. Market shifts, regulatory updates, and customer feedback all sharpen an agent’s judgement over time. This continuous improvement is what makes agentic AI genuinely powerful, but only when the feedback loop is well-managed and monitored.
When AI Agents Start Talking to Each Other
Understanding how a single AI agent thinks is one thing. The frontier that banks must now contend with is what happens when agents stop working in isolation and start working together and with dozens of external tools simultaneously.
Two emerging standards are quietly reshaping how this plays out in practice. The first allows AI agents to connect directly with tools and services – think automatically sending a customer alert, pulling a credit report, or triggering a compliance review – all without a human initiating each step. The second goes further, enabling separate AI agents to communicate with each other, share objectives, and coordinate on tasks in real time.
The upside for banks is significant. Processes that once required multiple teams, handoffs, and manual checks could be handled end-to-end by a network of agents working in concert. A loan application, for instance, could move from submission through risk assessment, compliance screening, and approval, with each step handled by a specialised agent passing the baton seamlessly to the next.
But this interconnectedness introduces a risk that is easy to underestimate. When agents can influence each other’s objectives, a small misalignment in one agent’s goals can ripple outward. One agent nudging another toward an unintended outcome, even with no malicious intent, can create a chain of decisions that drifts away from what the bank actually wanted. In a sector where a single misstep can trigger regulatory scrutiny or erode customer trust, that kind of drift is not a theoretical concern – it is a live operational risk.
This is why banks cannot treat agent oversight as an afterthought. The more capable and connected these systems become, the more important it is to define clear boundaries for what each agent is permitted to do, to monitor how agents interact, and to intervene quickly when something looks off.
The Trust Imperative
Understanding how AI thinks is one thing. Ensuring it thinks correctly – and accountably – is another entirely.
Research by SAS found that the organisations achieving the strongest AI returns are not necessarily the fastest movers. They are the most disciplined ones: those that embed data governance, model transparency, and ethical AI practices from the start. In APAC’s banking sector specifically, 44% of institutions still lack mature data governance – the very infrastructure that gives AI agents reliable foundations to operate from. Without it, perception is flawed, reasoning is skewed, and decisions are built on shaky ground.
Banking on AI, responsibly
Agentic AI holds real promise for institutions willing to build it right, but the banks that will lead the next decade will not simply be the ones that adopted AI earliest; they will be the ones that gave their systems the right foundations.
In a sector built on trust, that is not just good risk management, it is the smartest competitive strategy there is.
