Meet the Banking Friction Killer

Your most loyal customer has a personal AI agent. One morning, that agent receives your bank’s painstakingly crafted home-insurance pitch. It does not call the customer. It does not pause to admire the copy. It scans the market in seconds, picks the cheapest comparable policy from a competitor, and clicks buy.

The customer relationship you spent a decade building? Disintermediated by code, before breakfast.

This is the unsettling vision Chris Barford laid out to a room of Hong Kong’s banking elite during an EY keynote titled “The Taming of Autonomous AI in FSIs” at the recent FSI & AI Hong Kong Summit. Barford is the rare consultant whose CV starts with computer science — a Manchester graduate who came up through PwC and IBM, moved from EY London to Hong Kong in 2016, and now leads the firm’s financial services AI and data practice from a desk that doubles as a board seat at the FinTech Association of Hong Kong. He has spent the past decade watching banks flirt with automation. Agentic AI, he argues, is not the next phase of that flirtation but a structural rupture.

From GenAI to judgment machines

Generative AI has mostly sped up existing processes. Cleaner emails, quicker summaries, and modest productivity bumps that satisfied a CFO (and bored a CDO). Agentic AI is something else. It shifts the human role from execution to judgment — and increasingly to oversight of judgment made by something else.

The numbers Barford put up are staggering. Ops and middle-back-office leaders across Hong Kong, he said, are walking around with CFO mandates to strip 15-40% of costs out of their organizations over the next couple of years. Nobody has fully explained how. The unspoken assumption is that AI will simply do it.

To approach the conceptual ceiling of 50-80% cost reduction, firms must move past copilots and rebuild around the six core capabilities Barford calls for. Classify unstructured inputs, extract clean data, retrieve context-grounded knowledge, make rules-bound decisions, generate content, and search the open web. Skip any one of these, and your agents plateau.

Microservices, reborn

Barford’s witticism of the day, delivered in consultant deadpan: microservices architecture is back. It has just been rebranded as agentic.

Banks, he argued, will need to expose APIs not for their partners but for their customers’ own agents to negotiate with. Agent-to-agent commerce is no longer hypothetical. And it will reorder how products get discovered, priced, and bought.

For data leaders in the Asian Finance industry, that rewires the entire stack conversation. The old prize of a 360-degree customer view gets inverted. Soon, the customer’s agent will hold a 360-degree view of you. The institution that cannot present its products in clean, machine-readable, semantically reliable form will be invisible at the moment of decision.

The data moat is yours to lose

Barford offered an optimistic note, and it should land hard for every CDO in Asia. Financial services firms still own the most valuable training material on the planet. Anyone can rent a frontier model access for a few thousand dollars a month. Hardly anyone can replicate decades of transactional, behavioral, and relationship data, locked away behind regulatory walls. That asymmetry is the single greatest strategic advantage incumbents have left.

The catch: most of that data sits unstructured, ungoverned, and unloaded. Barford pressed the audience with a simple question. How many had truly fed their proprietary data into their internal GPT deployments? Not many, by his telling. The firms that win the agentic decade will be the ones doing the unglamorous data plumbing now.

When models drift, things break

Barford then offered an anecdote that should worry every AI leader. He had built himself a personal agent to scan business cards and enrich each contact with LinkedIn data. It worked beautifully, until it didn’t. The LinkedIn lookups quietly stopped returning results. Somewhere in a faraway data center, the underlying foundation model had decided to deprioritize that source. Drift, silently injected upstream.

Now scale that single point of failure across an autonomous credit-assessment workflow. One subtle model change inflates declared incomes by a percentage point. A downstream verification agent, recalibrating around the new normal, lowers its standards. Lending decisions degrade. Spotless credit records start to wobble.

This, Barford warned, is the cascade-of-failures problem. It is an AI architecture risk and a data leader’s risk.

Asia’s regulatory split

The continent’s regulators are not aligned, and Barford was direct about who is leading. Singapore’s MAS, through its Veritas Initiative, has built what he called gold-standard guidance through years of industry collaboration.

Hong Kong has moved more cautiously but pragmatically: principle-based, sector-specific, refreshingly permissive. “Nothing is prohibited” in Hong Kong, Barford noted, in pointed contrast to the E.U. AI Act’s prohibited-use list. China sits at the other extreme. Models must be approved up front, which is a brake on independent experimentation but can be an accelerator once they clear for state-sector deployment.

For AI leaders in Asia, the takeaway is operational. Build once, deploy many. But design for the strictest regime you might enter, because regulatory arbitrage is brittle when foundation models change behavior overnight.

Where humans actually belong

Barford reserved his most provocative argument for the human-in-the-loop orthodoxy. He believes it is on the way out. Not because humans do not matter, but because asking junior staff to validate sophisticated AI outputs is “fantasy.” The outputs are too polished, the underlying logic too opaque, the reviewers too inexperienced to push back. The future, he argued, is human control, not human verification of every transaction.

That is the message AI and data leaders in Asia need to carry to their boards this quarter. Not “we are deploying agents.” Not “we have a copilot strategy.” Something sharper.

Here are the processes we are removing from human execution. Here is where accountability now sits. Here is the data foundation that makes it safe.

The age of agentic AI is here. The agents are already shopping. The question is whether your bank shows up in the results.

Image credit: iStockphoto/Byrdyak