AI for HR in Hong Kong: useful workflows and data limits

AI for HR in Hong Kong is most useful when it removes the first pass of a repetitive task: drafting a job advert, turning an approved policy into a manager briefing, or preparing an onboarding checklist. It is much less convincing as a judge of candidates or staff. The practical question is not whether HR can use AI. It is which material may enter the tool, what it produces, and who checks the result before it affects a person.

I would begin with work that has a clear source and a human owner. A team can test a workflow on fictional or cleared material, agree what a good output looks like, then decide whether the same process belongs inside an approved company environment. That gives HR a way to learn without making an applicant's CV the first experiment.

What can AI for HR in Hong Kong do well?

Start with a job advert. Give the tool an approved role brief, reporting line, location, essential requirements and a house style. Ask for a plain-language draft and a separate list of claims that were not in the brief. A recruiter then checks the requirements against the real job, removes language that could deter suitable applicants and confirms the salary or benefits wording with the right owner. The machine saves drafting time. The recruiter still owns what the firm says publicly.

Another useful job is policy explanation. An HR team may have a long, approved leave policy and a manager who needs a one-page guide for handling a common request. AI can turn the source into a draft checklist with section references. HR then checks every entitlement, exception and deadline against the policy before circulation. This is similar to checking an AI draft for invented facts: the polished sentence is the beginning of review, not evidence that the rule is right.

Onboarding is a third candidate. Use a role and start date to draft a sequence of tasks, then have the manager and HR assign owners for equipment, access, introductions and mandatory training. The output is a working checklist, not an automatic decision about the new employee. The free prompt library has examples that can be adapted to approved, non-sensitive material.

Let AI prepare the paperwork; keep the judgment with the person accountable for it.

Recruitment needs a tighter boundary

An applicant's CV contains personal data. It may also contain details irrelevant to the role. Pasting a folder of CVs into an unapproved chatbot is a poor starting point, even if the resulting ranking looks tidy. The Privacy Commissioner's Code of Practice on Human Resource Management applies the Personal Data (Privacy) Ordinance to recruitment, current employment and former employees' matters. Its recruitment guidance concerns what data employers collect, why they collect it and how applicants are informed. Adding an AI tool to the process does not remove those duties.

If a team wants help reviewing applications, start with the job criteria, not the candidates. Ask AI to turn an approved role specification into a consistent interview question bank and a scoring rubric for human review. A hiring manager should inspect the rubric for vague proxies such as "culture fit" and for criteria unrelated to the work. Test it against fictional profiles before using it in a real process. If candidate information will be processed, HR and the firm's privacy or legal owner need to settle the permitted tool, purpose, access and retention first.

Do not ask a model to infer personality, loyalty, health or likely family plans from a CV or an interview note. Those guesses are not job evidence. A summary can still be useful, but it should quote the relevant experience and show where it came from. Someone must be able to challenge it by returning to the application, rather than accepting a confident label.

Employee data is not ordinary prompt material

Performance notes, pay, medical information and grievances carry different stakes from a job-ad draft. The PCPD's March 2025 checklist for employee use of generative AI recommends that organisations define permitted tools, the information employees may enter, review requirements and reporting routes. The PCPD's 2024 AI Model Personal Data Protection Framework also calls for governance, risk assessment and human oversight where AI systems involve personal data. These are guides for setting controls, not approval for every use of an AI product.

For an HR workflow, write the boundary in operational terms. Which tool is approved? May it receive personal data, or only synthetic examples? Who can see prompts and outputs? How long are records kept? How does an employee correct a mistake that entered a report? The answer depends on the firm's system and purpose. Our guide to using ChatGPT with work data explains why a general claim that a product is "secure" cannot answer those questions on its own.

De-identifying a record can reduce risk, but replacing a name with "Employee A" does not automatically make the rest anonymous. A distinctive role, incident and date may still identify the person inside a small team. If those details are not necessary for the task, remove them. If they are necessary, use the firm's approved route and have the data owner decide whether the processing is appropriate.

A one-week test that proves something

Choose one low-risk task: for example, an onboarding checklist built from an approved template. On day one, record how the task works now, including who supplies information and who signs off. Prepare two fictional examples, one routine and one awkward. Keep the source material beside every AI draft so a reviewer can trace each item.

Then run the same prompt with the same inputs in the tool the firm permits. Ask it to separate facts supplied by the source from assumptions. A reviewer checks omissions, wrong owners, made-up policy steps and any instruction that should have been escalated. Revise the prompt once, run the two examples again, and keep the changes that genuinely help. Avoid calling a faster draft a time saving if HR spends longer repairing it.

At the end of the week, the deliverable is modest: a tested prompt, two reviewed examples, a list of material the tool may receive, and a named person who checks the output. If the task still requires extensive repair, stop there. Perhaps the source policy is unclear, or the workflow has too many exceptions. More prompting will not repair a broken process.

Where the person must stay in charge

AI can prepare an interview guide, summarise an approved policy or organise a checklist. It should not be the unexamined reason somebody is hired, rejected, disciplined or denied an entitlement. A person who knows the role and the evidence must make and explain the decision. They also need a route to correct source data when the model's summary is wrong.

This distinction is where a small HR pilot becomes credible. It gives staff permission to remove drudgery while making the data boundary and review obligation explicit. I work with professionals and teams to build that kind of repeatable method inside their real workload. If your HR team has one process that keeps eating the week, let's examine it together before you buy another tool or hand over another decision.