Reinvent or Disappear: Inside the AI Reckoning Reshaping Asia's Banks

In the demo room, everything works.

The agent reads the documents, drafts the memo, answers the follow-up question with unnerving poise. Executives nod. Somebody photographs the screen for the board pack. Somewhere in an annual report, a slide is updated: a confident chart, pointing up and to the right, announcing that the bank has an AI strategy.

Naren Gunawardana has sat through more of those demos than he can count. He has also seen what happens next, when the same system leaves the demo room and meets the bank itself: the core platform designed in the nineties, the customer data scattered across eleven systems that quietly contradict each other, the risk framework written on the assumption that a human being made every decision. The demo dazzles. Production stalls. And another promising project joins what he calls the graveyard of pilots.

That distance - between the announcement and the reality - is right now the most expensive gap in banking. The stakes are not rhetorical: McKinsey estimates generative AI could add $200 billion to $340 billion of value to global banking every year, up to 4.7 percent of the industry's total revenues, yet its survey of credit executives found just 12 percent of North American institutions had fully deployed even a single use case - independent confirmation of the pattern Gunawardana sees from the inside. And the window is short. The same firm's research shows generative AI reached 45 percent of the US working-age population in roughly two years, an adoption feat that took digital banking fifteen. In under three years, generative and agentic AI have travelled from science-fair demo to serious production conversations inside regulated institutions, and the shift is happening first and fastest in Asia. Yet most of what gets reported is ambition, not outcome. The headline says transformation. The hallway says blockage.

That was the subject of a recent Good to Great Podcast episode, which set out to answer the harder question: not whether AI will change banking - that question is over - but why it is so hard to make it real, and who is figuring out how.

Leading the conversation are two hosts who belong to the very generation now deciding where it fits in what comes next. Mindu Hapangama is a student at the University of Colombo in Sri Lanka and the University of London in England, and third runner-up of the HSBC/HKU Global Business Case Challenge 2026. Jacob Callaway is a commerce and law student at the University of Sydney, a member of the champion team at the HSBC/HKU Global Business Case Challenge 2026, and an International Case Squad member of the Sydney Consulting Club. Their questions carry the urgency of people who will spend their entire careers inside the world this episode describes.

Their guide is Naren Gunawardana, Managing Director at Synpulse in Singapore, where he steers financial institutions across Southeast Asia through the part everyone finds hardest: not the strategy deck, but the transformation itself - the operating models, the governance, the use cases that finally escape the pilot cycle and scale. He comes to it from a rare vantage point. Fifteen years in private banking and wealth management, beginning as a business analyst at UBS, where he won the bank's APAC hackathon in 2014, before rising through Synpulse from consultant to managing director. Today he works with private banks, family offices, and COOs on agentic AI for onboarding, advisory, and the middle office. And he is a builder in his own right, shipping his own AI applications - an FX trading tool, a daily prompt newsletter - precisely to learn what genuinely works and what quietly doesn't. When he talks about why AI stalls inside a bank, he is not theorizing. He has stood at the lending desk, in the name-screening queue, inside the unglamorous machinery where transformation actually lives or dies.

What follows is his view from inside the building.

01. The Line Between Real and Rumor

Ask Gunawardana where the truth sits in 2026, and he starts by puncturing the loudest number in the industry. Nearly every bank has announced an AI agent in production, he notes - but in reality, perhaps a tenth of them have actually done it. The gap between announcement and reality is not a rounding error. It is the defining fact of the moment.

What is genuinely real is older and quieter than the hype suggests. Banks have run machine learning and deep learning in production for years - across fraud detection and compliance - and switching those systems off would hurt immediately. The newer wave, sparked when bankers began using tools like ChatGPT and Claude in their personal lives, is now being applied to agentic workflows inside the institution. And the best return on investment so far is almost comically unglamorous: agentic coding inside the software development lifecycle. The middle and back office, not the customer-facing chatbot, is where AI is quietly earning its keep.

What remains a press release, in his blunt assessment, is anything with the words "fully autonomous" next to it - above all, autonomous advisory, which collides head-on with one of the most stringent regulatory frameworks in commerce.

The line, drawn plainly: AI is load-bearing in the back and middle office and still early at the front. The flashier the claim, the more likely it is still a slide.

02. What an Agent Actually Does All Day

Agentic AI is the phrase on every slide, and almost nobody defines it. Gunawardana's definition is disarmingly practical. A chatbot answers your question. An agent has the tools, the knowledge, and the autonomy to finish the task you set it. The difference, he says, is the difference between asking Google Maps for directions and having a self-driving car take you there.

His favorite live example is the source-of-wealth investigation - the obligation, when a client joins a private bank, to verify where the money came from. Gunawardana knows this work personally; it belongs to the unglamorous machinery he spent fifteen years inside, from the lending desk to the name-screening queue. Picture the old version: a single analyst or relationship manager, stitching a story together by hand from bank statements, company filings, and notes from client interviews, hunting for the thread that either holds the narrative together or unravels it. Today, agents assemble that case.

But the detail people get wrong is the one that matters most. In a heavily regulated industry, the AI does not make the final decision - and in any well-designed system, it should not. The agent does roughly eighty percent of the legwork; a human in the loop approves the twenty percent that actually matters. The analyst is not replaced. The analyst's job changes shape - from searching and assembling to judging the output.

The reframe: Agentic AI is not replacing the banker. It is replacing the busy work, and turning the banker from a gatherer of information into a judge of it.

03. Why It Stalls

Banks would seem to have everything required to move fast: money, talent, urgency, a board mandate. And yet the efforts stall - reliably enough that calling Gunawardana in has become a pattern. His diagnosis of the real reason, not the one that goes in the post-mortem deck, is the center of gravity of the whole conversation.

The fundamental mistake, he argues, is treating AI as a technology project. It is a complete organizational redesign - of the people, the governance, and above all the data. It is never that the AI isn't smart enough. It is three problems underneath it.

The first is plumbing. Core banking systems designed in the nineties were never built to serve data to AI agents, and the perfect demo collapses in production when it turns out the bank's data lives in eleven different places, with contradictions among them.

The second is governance built for the wrong species. Decades of regulation and risk frameworks assume a person made the decision. Nobody has fully rewritten the accountability model for agents - which produces the strange spectacle, Gunawardana notes, of regulators moving faster than the banks they supervise.

The third is trust. If people believe AI is auditioning for their job, they simply refuse to adopt it - and no solution, however good, survives an analyst with twenty years of experience who declines to use it. Banks are already inventing new roles specifically to supervise AI agents, because the workforce can feel its jobs changing shape in real time.

The real post-mortem: Dirty data, governance written for humans, and broken trust. The technology is almost never the constraint.

04. The Line That Never Gets Crossed

The moment AI starts making decisions about people's money - who gets a loan, what advice they receive, whether a transaction looks suspicious - trust stops being a nice-to-have and becomes the entire business. So where does the line sit?

Gunawardana's answer is absolute. An agent can build the case and support the case, but it can never be the last check. On any AI-generated advice, investment or regulatory, there must always be a human who is accountable - and that constraint is not something to monitor into existence after launch. It belongs in the design from day one.

The second non-negotiable is explainability. Regulators have made it plain: a bank must be able to show, to the supervisor and to the client, exactly why an AI agent reached a decision, no matter how accurate the system is. AI cannot be a black box that swallows inputs and emits outcomes.

And there is a third principle that boards underestimate. Banks are eager to work with the frontier-model vendors - the Anthropics and OpenAIs of the world - and Gunawardana's caution is precise: you can outsource the technology and the research, but you cannot outsource the risk. Whatever model sits underneath, the bank remains accountable.

The design rule: Human accountability and explainability are not features to add later. They are the foundation, and no acceleration of the technology changes that.

05. The Innovation Nobody Is Watching

There is a revealing contradiction in how bankers live and how their banks operate. In their personal lives, Gunawardana observes, they reach for the most cutting-edge tools available. Inside the institution, the developments that are genuinely moving the industry forward are the ones nobody photographs for the keynote.

The one that animates him most is happening in the org chart, not the tech stack. Banks are beginning to hire for a role that did not exist three years ago: the human manager of a hybrid team, part people and part agents - someone whose job is to supervise a workforce that is partly synthetic. How do you recruit for that? How does work get checked when half your reports are software? The role is being redesigned in real time, and Gunawardana admits a personal stake in the outcome: he is the father of a young son, and the working world that boy inherits is being drafted in these hiring decisions now.

The second development is stranger still: the regulator as accelerant. In Singapore, the Monetary Authority is not merely issuing rules to comply with; it is publishing practical operating manuals. Its MindForge initiative convenes banks and the regulator in a single consortium, building a shared handbook for rolling out AI effectively and safely - a model Gunawardana expects other jurisdictions to copy, and one that helps explain the strange spectacle of supervisors outpacing the supervised.

The third is almost defiantly uncool. While the industry debates frontier models, the systems actually reaching production often run on small, cheap, local language models doing narrow tasks inside a workflow - unglamorous machinery that delivers economic returns and keeps sensitive data in-house. Which, in a regulated business, is precisely what separates a launched product from a permanent pilot.

The tell: The leaders of tomorrow will not be the banks with the flashiest avatar demo, but the ones quietly rebuilding their org charts and guardrails.

06. When the Whole Market Runs on AI

Widen the lens from one bank to the market it lives inside - the funds, the traders, the risk desks, the platforms on retail investors' phones, all increasingly running on models - and a structural question appears. What happens when a thousand models, trained on similar data, react to the same signals at the same moment?

Gunawardana's answer begins with an observation that should unsettle every technology executive: the moat is moving. For decades, an institution's edge could be its engineering - the best builders, shipping fastest. But it has never been easier to build, which means building is no longer where advantage lives. The question becomes what you choose to build, and that is a matter of taste and judgment, not headcount.

Then he identifies the deeper mechanism, the most quietly important idea in the episode. Generative AI, by its statistical nature, normalizes - it produces the most probable next answer. Feed the same public data into a thousand similar models and they converge on similar conclusions. An industry that adopts the same intelligence everywhere does not become smarter relative to itself; it becomes uniform. The research, the strategy memos, the risk reads all begin to rhyme.

Which is exactly why the edge migrates to whatever cannot converge: proprietary data no competitor holds, the specific team you assemble, the judgment to decide what deserves building at all, and the organizational design that lets you deploy it faster than the bank next door. The institutions that thrive, Gunawardana argues, will be those that guard and compound the data moat while expressing genuine human judgment on top of it - because everything else is becoming everyone's.

Where the edge goes: When every model gives the same answer, alpha lives in what is not public - your data, your people, your judgment.

07. The Bank of 2035

Asked to picture banking a decade out, Gunawardana offers three answers, each more provocative than the last.

What disappears is the busy work: the manual document-chasing, the static periodic reviews, the laboriously assembled product pitches. Automating those inefficiencies frees bankers for the core value of the job - applying human judgment and creating new economic value.

What gets dramatically better is personalization, at a level that was never previously economical. Advice that is contextual, personal, near-instant, around the clock - and compliant. His striking formulation: every client will get a family-office level of attention, not just the ultra-high-net-worth layer at the top.

And the prediction he would put his name to, the one most peers would resist: the winner of the AI race will not be the bank with the biggest AI budget or the best model. AI, he argues, is becoming like electricity - a commodity. Give two banks an identical model, and the one with clean, connected data and guardrails agreed with the regulator will run away from the one with a cutting-edge model sitting atop fifteen years of silos and a long approval cycle. Leaders will earn more on the same AI dollar simply by being organized better.

The contrarian bet: The AI race will not be won with AI spend. It will be won with plumbing.

08. What Makes a Person Worth Keeping

Underneath every prediction sits the question everyone is quietly asking, from the finance student about to graduate to the banker twenty years in: when machines take the routine work, what makes a person worth keeping?

Gunawardana's dividing line is clean. If your job is gathering information to find the answer, parts of it are at risk; the go-out-and-summarize work is disappearing. If your job is knowing which answer matters, you become more valuable, not less. The people worth keeping are the ones prized for judgment under ambiguity - for knowing when something doesn't add up even though the model says it does, for catching the confident answer that is wrong.

His practical advice is immediate. Get your hands dirty with the tools now - not to become an engineer, but to build the instinct for what AI is genuinely good at and where it fails, an instinct earned only through repetition. And if you sit at the gathering end of a workflow, make the deliberate move toward its decision point. The people who thrive, he says, will not be the ones racing AI on speed. They will be the ones brilliant at catching what AI misses - a standard, he adds with characteristic candor, that he applies to his own job too.

The career rule: Move from the gathering point to the decision point. Judgment under ambiguity is the last moat.

Conclusion: The Choice, Stated Plainly

Strip away the keynotes and the graveyard of pilots, and the episode's conclusion is almost austere. The banks that win the next five years will not be the ones with the most pilots or the loudest announcements. They will be the ones willing to do the slow, unglamorous work underneath - the plumbing, the governance, and the hard conversations about trust and about people.

That is what makes the episode's title less a threat than a description. Banking is entering its fastest structural change in generations, and it is arriving first in Asia, inside institutions where a mistake is not a bug report but a headline. The technology, Gunawardana has shown from the inside, is the easy part. The reinvention is everything else. Reinvent or disappear is not a warning shouted from a stage. It is simply the choice, stated plainly - and the institutions now quietly fixing their data, rewriting their org charts, and drawing their red lines have already made it.