From June 2026 wages, Malaysian employers had to deduct a new contribution from their staff: 0.75% of pay, up to a wage ceiling of RM6,000, for SOCSO’s Lindung 24 Jam scheme covering accidents outside work. About nine million workers were affected. Payroll teams changed their settings, explained the smaller payslip, and remitted the money.
On 9 July 2026 the Human Resources Minister announced that the Cabinet had made the contribution voluntary for Malaysian employees, with immediate effect. It stayed mandatory for foreign workers. Lawyers advised employers to review their deductions and follow SOCSO’s latest instructions.
Five weeks, two rule changes, and different treatment by citizenship. This is the first thing to understand about AI in HR. Payroll is a set of rules, and the rules change when someone makes an announcement. A language model does not know what the Cabinet decided last week.
Three kinds of HR work
Most of what an HR team does falls into one of three kinds. AI suits one of them.
Rules. Salary, overtime, EPF, SOCSO, EIS, monthly tax deductions (PCB), the HRD Corp levy, leave entitlements. Each has a correct answer that depends on dates, wages, age and citizenship. Another recent example: since October 2025 wages, employers and their non-Malaysian employees each contribute 2% to the EPF, a change the fund announced on 25 June 2025. This is work for payroll software that somebody keeps up to date, with tests. The same inputs must always give the same payslip.
Reading. CVs, medical certificates, expense receipts, new-joiner forms, resignation letters, the staff handbook. Documents and questions in ordinary language. This is where AI earns its place.
Judgement. Who to hire, who to promote, how to handle a grievance, whether to dismiss. This is for people, on the record.
Keep the three apart and most of the risk goes away. Trouble starts when a tool built for reading is allowed to make a judgement or do the arithmetic.
Where AI helps today
Turning documents into records. A new joiner sends a photographed identity card, a bank statement header and an EPF number. A claim comes in as a crumpled receipt. A model can read each one into the right fields, and a person confirms before it is saved. We use the same pattern in production for bank statements of account: the model extracts, a person confirms, and code does the calculating.
Answering the questions HR answers every day. How many days of annual leave do I have left? What is the claim limit for outpatient treatment? Can I carry leave forward? An assistant can answer from the company’s own handbook and leave records, show which clause it relied on, and pass anything it is unsure of to a person. Staff will ask in a mixture of Malay and English, often on WhatsApp. See how well models handle Malaysian language and AI agents on WhatsApp.
Drafting. Job advertisements, offer letters from a template, interview question sets, a plain-language summary of a policy change. In Jobstreet by SEEK’s 2025 report, job ad creation was the task recruiters most often named among their uses of AI. A person edits and sends.
Finding the exceptions. A claim submitted twice. A medical certificate dated on a public holiday. Overtime that exceeds what the roster allows. These are checks a system runs across every record once the documents have been read.
Recruitment is where to slow down
Screening is the obvious use, and it is already happening. Jobstreet’s 2025 survey of 2,279 HR professionals and employers found 26% of recruiters using AI tools, with candidate screening among the top uses. It also found 47% of businesses hesitant to adopt AI, citing a lack of personalisation.
They have reason to be careful. In an earlier Hays survey of 832 Malaysian employers, reported in February 2024, 57.4% of respondents believed AI resume screening could be biased. Among those already using AI, 26.6% were not actively assessing bias in their tools, and only 26% of HR staff had a policy governing their use. Those figures are more than two years old, and we have not found a newer Malaysian survey that asks the same questions.
Bias gets in easily here. A Malaysian CV often carries a photograph, an identity card number, age, race, religion and marital status. A name alone often signals ethnicity and gender. A model asked to “rank the best candidates” will use whatever is on the page, including patterns from the past that nobody intended to repeat.
A safer design has four parts.
- Extract, do not judge. Ask the model for facts against the job’s stated criteria: years in a relevant role, licences held, languages, notice period. Not a score for “fit”.
- Hide what should not count. Remove the photograph, name, identity card number, age, race and religion before anyone, or anything, ranks the applications.
- Never auto-reject. The system sorts and summarises. A recruiter decides who goes forward, and can see every application.
- Keep the record. What the model was shown, what it returned, and who made the decision.
What the regulator now expects
On 30 April 2026 the Personal Data Protection Commissioner issued guidelines on automated decision-making and profiling. According to a summary by Allen & Gledhill, they cover automated decisions that significantly affect a person, and list employment opportunities among the examples. Profiling is defined to include predicting a person’s work performance.
Three points matter for HR.
- Tell people. Individuals should be given a notice that automated decision-making or profiling is involved, what kinds of decisions are made, and the possible consequences.
- Keep a person in it. The guidelines’ best practices say AI should not be the sole basis for a decision about an individual, and that trained staff should be designated to review how it is used.
- Treat health data with extra care. Explicit consent is expected where the processing involves sensitive personal data such as health information. Medical certificates and medical claims are exactly that.
Underneath this sits the amended Personal Data Protection Act, which made breach notification mandatory and put direct obligations on vendors that process data. An HR system holds identity numbers, salaries, bank accounts and medical records for every employee. We cover what that means for a system in our note on the amended PDPA.
One practical consequence: staff records do not belong in a consumer chatbot. If your team is pasting salary lists into a free AI tool to fix a formula, start with a one-page AI policy.
What to do, in order
- Get the rules into a system someone maintains. For payroll this usually means buying, not building. Ask the vendor how quickly it shipped the July 2026 change.
- Pick one reading job. Expense receipts or new-joiner forms are good candidates. Measure how many fields a person has to correct.
- Add a handbook assistant that answers from your own documents, cites them, and hands over when unsure.
- Leave recruitment until last. When you do it, strip the fields that should not count, keep the decision with a recruiter, and tell candidates how their application is processed.
- Write down where the data goes. Which provider, which country, what is retained. Staff are entitled to ask.
Read with AI. Pay with rules. Hire with people. An HR team that keeps those three apart gets the time back without creating a new kind of dispute.
Sources
- SOCSO 24-hour protection scheme to raise payroll deductions for nine million workers from June (The Vibes)
- Perkeso’s Lindung 24-hour scheme remains mandatory for foreign workers (The Vibes)
- SOCSO reverses position on Lindung 24 Jam: what employers need to know (Shearn Delamore & Co, via Conventus Law)
- Mandatory EPF contributions for non-Malaysian employees to begin in October (Human Resources Online)
- Nearly half of Malaysian employers plan to expand permanent workforce: Jobstreet survey (The Sun)
- Nearly half of Malaysian employers don’t plan to use AI in recruitment: survey (HRD Asia)
- PDP Commissioner launches guidelines on data protection impact assessment, automated decision-making and profiling, and data protection by design (Allen & Gledhill)
This article is general information and is not legal advice.