AI pilot decision tool

How should a Hong Kong enterprise define and accept an AI pilot?

Treat the pilot as a controlled business decision, not a model demonstration. Fix one workflow, its baseline, authority boundary, evidence set, acceptance thresholds, and decision owners before testing begins.

One workflowNamed users, baseline, and consequence
Evidence gatesCapability, control, and operations
Signed decisionScale, redesign, hold, or stop

What the charter fixes before work starts

A credible pilot makes the decision boundary explicit enough that a polished demo cannot move the goalposts.

Decision and owner

Name the operating decision, sponsor, workflow owner, technical owner, risk approver, and final decision authority.

Scope and baseline

Record users, frequency, current cycle time or quality, systems of record, included cases, exclusions, and the counterfactual.

Authority and data

Separate read, recommend, draft, approve, and execute rights; map personal data, retention, access, cross-border movement, and prohibited uses.

What the acceptance record must show

The final decision should be reproducible from evidence rather than confidence or presentation quality.

Capability evidence

Use representative common, edge, high-impact, missing-data, and adversarial cases with metric definitions and minimum thresholds.

Operational evidence

Measure latency, cost, availability, failure recovery, audit completeness, overrides, escalations, and user adoption in the complete workflow.

Decision disposition

Record scale, redesign, hold, or stop; list unmet gates, accepted residual risks, conditions, accountable signatories, and the next review date.

How to use the pilot charter

Draft and sign the charter

Approve the problem, baseline, data, authority, evaluation set, thresholds, owners, timeline, and stop conditions before build.

Run the bounded pilot

Test the complete workflow with controlled users and preserve versions, inputs, evidence, failures, interventions, and costs.

Review against every gate

Mark each criterion pass, fail, inconclusive, or not tested; do not hide severe failure behind an average score.

Sign the disposition

Scale only the approved authority tier; otherwise redesign, hold, or stop with a documented reason and evidence owner.

Regulatory sources and scope

These are the primary regulator materials used by this implementation guide. This page is not legal advice.

Frequently asked questions

Is a successful demo enough to pass the pilot?

No. A demo may establish technical possibility, but acceptance also needs representative quality, data and authority controls, operational reliability, human review, and a defensible business measure.

Should every metric use the same threshold?

No. Thresholds should follow task consequence and risk. Averages must not compensate for a failed privacy, permission, safety, or other mandatory gate.

Is this legal or regulatory advice?

No. It is a provider-authored implementation template grounded in public Hong Kong guidance. Adapt it with your legal, privacy, security, risk, procurement, and sector specialists.

Start with one concrete operating problem.

We will first clarify the workflow, evidence, data boundary, and outcome standard before recommending the next step.

Contact Onyx