This deep dive expands on the Loop Asia conversation with Koo Ping Shung, Practicum Director at Data Science Rex.

Koo Ping Shung said something every AI governance committee member will recognise: getting to a final decision can take six months to a year — different functions in the room, each with their own agenda, no easy way to reach one conclusion. It's not that the committee is badly run. Consensus-based ethics resolution as the gate before anything ships is structurally too slow for what it's governing. "There's no way AI innovation can wait for you six months to a year," he said. Teams ship anyway, which means the governance model on paper isn't the one operating in practice.

If you're the executive accountable for that outcome — CEO, COO, head of professional services, whoever owns the "move on AI" mandate without a technical background to lean on — treat this as an operating design problem, not a compliance one.

Why the ethics-first model stalls

Koo's argument for why full ethical consensus is the wrong gate isn't a hand-wave. He doesn't believe you can build an "ethical machine": no human is ethical 100% of the time, and humanity hasn't agreed on foundational questions like the trolley problem after decades of trying. If we can't resolve it among ourselves, we don't have the standing to encode one answer into a machine. A model can't hold competing values in tension the way a human can — it optimises for whatever the designer fixed as the target and trades off against the rest.

That's a genuinely hard problem, and Koo isn't proposing you solve it before you act. He's proposing you stop treating it as the gate.

Task-level trust as the actual operating model

The distinction he draws instead — trustable versus trustworthy — is a decision architecture a committee can use now. Trustworthy is a claim about an entity: reliable across the board — the standard the six-month committee is implicitly certifying, and why it takes six months, since entity-level trust means resolving every value trade-off at once. Trustable is a claim about a task: can I predict, within reason, what this use case will do. Narrower, faster, answerable case by case, without agreeing on ethics in the abstract first.

In mechanics, that's risk-tiered delegation rather than an ethics tribunal. Low-stakes, well-understood tasks — small blast radius, well-characterised failure mode — get evaluated fast against Koo's other idea: a library of prior use cases, weighed against precedent instead of first principles. High-stakes, novel tasks still go to deeper review — the exception, not the default gate.

What this means for the executive holding the mandate

If you're accountable for AI outcomes without owning the technical detail, your leverage isn't resolving the ethics debate — nobody resolves that on a committee timeline. It's deciding what "trustable enough to ship" looks like for a given task, and building the muscle to make that call quickly and repeatedly instead of convening the full committee every time.