Generative AI has spent the last two years oscillating between hype and caution in financial services — celebrated for its efficiency gains, distrusted for its opacity. Hong Kong’s regulators appear to be betting there is a middle path, and Manulife Hong Kong has just become one of the test cases for what that path looks like in practice.

A Four-Regulator Experiment

The GenA.I. Sandbox++ is not a typical single-agency pilot. It brings together the Hong Kong Monetary Authority, the Securities and Futures Commission, the Insurance Authority and the Mandatory Provident Fund Schemes Authority, working alongside Cyberport, to create a controlled environment where financial institutions can trial GenAI use cases with regulatory visibility built in from the start rather than bolted on after deployment. That cross-sector structure matters: it suggests Hong Kong’s regulators see GenAI risk as a shared problem spanning banking, securities, insurance and retirement savings, not something any single authority can supervise in isolation.

Manulife Hong Kong is one of only five insurers admitted to this first cohort, and it enters with two separate use cases — one on the insurance side, one on the retirement side — giving it a rare vantage point across both halves of the sandbox’s scope. The company’s participation builds on its earlier selection as a Core Participating Insurer in the Insurance Authority’s separate AI Cohort Programme, reinforcing a pattern of Manulife positioning itself early in Hong Kong’s regulatory AI conversations rather than waiting for standards to settle.

Claims: AI as a Reference Tool, Not a Decision-Maker

The first use case, BetterClaims AI, targets a genuinely difficult moment in the customer relationship: the point at which someone is filing a medical or critical illness claim, often under stress, and needs accurate answers fast. The tool is built on Manulife’s own approved policy content and functions as a knowledge assistant for advisors — surfacing relevant coverage details, policy definitions and claims guidelines, each linked back to authoritative source material, at the point of customer engagement.

What’s notable is where the company has deliberately drawn the line. Claims assessments and final decisions stay with qualified claims professionals; the AI’s role is confined to giving advisors faster, better-referenced information, not to adjudicating claims itself. That separation of “information retrieval” from “decision-making” is likely to be one of the more closely watched design choices to come out of the sandbox, since it offers a template for how insurers elsewhere might deploy GenAI in customer-facing roles without triggering the accountability problems that come with letting a model make judgment calls.

Fraud Detection: Flagging Patterns, Not Automating Enforcement

The second use case moves into less familiar territory: a GenAI-assisted red-flag tool, developed with Cyberport, designed to support fraud risk review of intermediary-related MPF transactions. The tool combines AI, document intelligence and analytics to help investigators spot potential risk patterns in Mandatory Provident Fund dealings — a corner of Hong Kong’s retirement system where intermediary misconduct can have outsized consequences for scheme members’ savings.

Here again, the guardrails are explicit. The tool is described as a decision-support system only; it does not automate investigations or enforcement actions. Human investigators remain responsible for interpreting flagged patterns and deciding what happens next. As life expectancies rise and retirement planning grows more complex, earlier and more consistent detection of risk signals is one of the more tangible ways AI could improve outcomes for ordinary savers — provided, as Manulife’s retirement leadership emphasizes, that human judgment stays central to every decision.

The Bigger Signal: Compute Access and Cross-Industry Learning

Beyond the two use cases, participation in the sandbox gives Manulife access to Cyberport’s AI Supercomputing Centre — a meaningful benefit in itself, since compute access remains a real constraint on how quickly financial institutions can experiment with GenAI at scale. The programme is also explicitly framed as a collaboration space between financial institutions, technology vendors and regulators, with participants expected to share learnings rather than treat their pilots as proprietary black boxes.

That knowledge-sharing mandate may end up being the sandbox’s most consequential feature. Individual GenAI pilots are common across the industry; what’s rarer is a structured mechanism for the resulting lessons — on governance, on where AI should and shouldn’t touch a decision, on what “responsible adoption” looks like in practice — to circulate back to regulators and competitors alike. If it works as intended, Hong Kong’s approach could offer a reusable model for how other financial hubs balance GenAI experimentation against the accountability that customers and regulators rightly expect from an industry handling people’s health claims and retirement savings.