Closing Hong Kong's AI ROI gap: from pilots to payback

Most Hong Kong boards have already said yes to AI. Budgets are approved, pilots are running, and the demos look impressive. What hasn't happened, in most cases, is the part the board actually asked for: a visible, durable move in the P&L.

That gap — between AI that demonstrates well and AI that pays — is now measurable.

In a 2026 survey of more than 100 C-suite executives across Hong Kong and mainland China, the Deloitte–HKU AI Adoption Index found that 69% of organisations are experimenting with or scaling AI pilots, yet only 23% have moved anything into operational deployment and just 4% have reached what the study calls transformational maturity. Many executives say returns have fallen short of expectations.

Read that curve the way a CFO would: nearly everyone is spending, a quarter are operating, and almost no one is transforming. The enthusiasm is real. The realisation of value is not — yet.

The gap is rarely the technology

When returns disappoint, the instinct is to question the models. That's usually the wrong place to look. Three operating-model failures explain most of the gap.

Pilots are built to impress, not to run. A proof-of-concept tuned for a steering-committee demo skips the unglamorous 80% — data plumbing, integration, monitoring, ownership — that production actually requires. The demo clears; the rollout stalls.

There's no problem worth scaling. A surprising share of pilots are technology looking for a use case. Without a sharply defined business problem and a named owner with a P&L line, even a working pilot has nowhere to land.

The organisation isn't wired to absorb it. Deloitte's 2026 State of AI report notes that while worker access to AI rose sharply in 2025, only about a third of organisations are using it to genuinely re-imagine how they work — and many feel underprepared on data, infrastructure and talent. AI that no one owns, trusts, or changes their workflow for produces activity, not outcomes.

Treat AI as an operating capability, not a project

The enterprises closing the gap make one mental shift: AI stops being a portfolio of pilots and becomes a capability the business actually operates. In practice that looks like four moves.

Start from the problem, not the model. Choose a small number of problems where value is concrete and measurable, each with an accountable owner. The technology choice comes after.

Design the production pathway on day one. Every pilot should carry a credible route to production — data, integration, security, monitoring — before it is greenlit. A pilot with no pathway is a science experiment.

Put senior engineering next to the business. The fastest way across the pilot-to-production chasm is to embed senior engineers alongside the people who own the problem, building for the real environment rather than a sandbox. This is the forward-deployed engineering model, and it exists precisely because slideware doesn't survive contact with production.

Run it, don't just ship it. Value shows up only after go-live, which means models need stewardship — monitoring, recalibration, governance — as upstream data and foundation models keep moving underneath them.

What a board can do this quarter

• Fund two or three problems with clear owners and a P&L line, not ten exploratory pilots.
• Require a production pathway in every pilot business case before approval.
• Budget for the unglamorous 80% — data, integration, change management — not just the model licence.
• Measure value realised, not models shipped or tools rolled out.

None of this requires a moonshot. It requires treating AI the way Hong Kong's best operators already treat any capability that touches the P&L: with clear ownership, a path to production, and people accountable for the result.

That pilot-to-production gap is the work ASTRA was built for — an AI-enablement programme and forward-deployed engineers who turn promising pilots into systems that run, in Hong Kong, under your governance.