Short answer
When Hong Kong companies adopt AI, what must be managed is data, judgement, authority, action and accountability. Harry Lui sets out a four-layer AI judgement system to help management turn experience into governable capability.
Many companies assume the first step in adopting AI is choosing a model, buying software or teaching staff to write prompts. That work has its place, but it is not the hardest part of enterprise AI transformation.
The hard part is this: once AI begins reading customer data, analysing operations, suggesting prices, screening opportunities or even triggering the next action, does the company clearly know what the judgement is based on, who has authority to approve it, and how to trace it when the result is wrong?
After AI enters a company, what changes is not only the speed of work but the allocation of decision rights. What the company must manage is no longer just “are staff using AI?”, but “which judgements can AI assist with, and which decisions must a person own?”
From automated workflows to a judgement system
Traditional automation usually handles rule-based work, such as sending a confirmation when a form arrives or generating a report on a set date. AI takes part in understanding, classifying, comparing, inferring and recommending. That work often carries uncertainty and is affected by data quality and context.
Companies therefore need to build four connected layers.
Layer one: facts
What data does the AI use? Is it complete, current and authorised for use? Have customers’ own words, contract terms, product information and internal records been mixed up with opinion or speculation?
Without reliable facts, AI simply produces complete-looking answers faster. Rather than first chasing a smarter model, a company should make sure important data has a source, a version and an owner.
Layer two: judgement
The same data can lead to different conclusions. A company needs to break down how its best managers and frontline staff judge: which signals they notice, how they handle contradictions, when more information is needed, and when no conclusion should be drawn.
Only when judgement criteria are clearly expressed can AI become an assistant rather than a black box producing answers out of nowhere.
Layer three: action
Recommendation and action must be kept separate. AI may suggest adjusting a quote, prioritising a customer or requesting missing documents, but whether it may send, modify or commit directly should depend on permissions set according to risk.
Low-risk work can run automatically; work involving money, contracts, customer commitments, personnel or sensitive data should keep clear human review and escalation.
Layer four: learning
After each judgement, the company needs to keep the outcome: whether the recommendation was adopted, what actually happened, and whether errors came from missing data or flawed judgement rules. Without this layer, AI starts from scratch every time; with outcome feedback, the company gradually builds its own judgement assets.
Three gaps Hong Kong companies most often overlook
The first gap is scattered data. Much important information sits in WhatsApp, email, personal computers and colleagues’ memories. Whether AI can help is often not a model problem; the company simply lacks a trustworthy entry point for data.
The second gap is a mismatch between authority and accountability. Companies ask staff to own results without making clear what data they may use, what they may promise, and who decides on exceptions. Adding AI only amplifies this.
The third gap is recording answers but not reasons. Six months later, management sees a decision without knowing the constraints and alternatives at the time, so discussions repeat and mistakes that have already been paid for are made again.
Management can start with one workflow
A company does not need to build a large AI platform at the outset. It can first choose one workflow that is frequent, time-consuming, error-prone and measurable, such as quoting, customer follow-up, complaint classification or management reporting.
Map the current workflow, then answer five questions: which facts are used, who makes which judgement, who approves the next step, how results are measured, and how records flow back into the system. Only then decide which parts suit AI assistance.
The order matters. If the workflow itself is chaotic, adding AI only makes the chaos run faster.
An AI-native company does not hand everything to AI
A truly mature AI-native company does not let AI handle everything automatically. It clearly allocates roles between people and AI: AI organises large volumes of data, finds patterns, proposes options and carries out low-risk tasks; people own goals, trade-offs, exceptions, commitments and final accountability.
Competitiveness will not come only from which model a company uses, but from whether it can connect facts, judgement, action and learning into a traceable, improvable operating system.
Every competitor can buy AI tools. The judgement methods, customer understanding and operating feedback a company has built over years are the assets that are harder to copy.
Practical supplement
AI governance checklist for one workflow
Choose one workflow first and complete each field; leave anything unconfirmed open rather than asking AI to guess.
Your entries remain only on this page and will not be saved after you refresh or leave; use Copy worksheet to keep the text.
Discussion question
In your current workflow, what most needs clarification: data, approval authority, or feedback records?
Smark Global perspective
Smark Global helps Hong Kong companies move from AI training and hands-on work on real workflows towards AI agents, ERP/CRM, data integration and measurable operating outcomes.
Learn more: Smark Global AI Academy →Source: Original commentary by Harry Lui
Last updated: 2026-10-02（first published: 2026-10-02）