AI and PDPA in Malaysia: What to Check Before Connecting HR Data

Your HR manager wants an AI agent to collect claim receipts, check missing information and chase approvals. The proposed shortcut is to connect the entire HR folder.

That folder also contains salary letters, identity documents, medical certificates and disciplinary records. The workflow needs a small part of it, not everything.

The useful question is not simply whether AI is safe. It is what this particular agent can read, what it can do, where the information goes and who remains responsible.

For a Malaysian SME, those boundaries should be agreed before connecting live employee records. This checklist is a practical starting point, not a legal opinion or a complete PDPA compliance assessment.

Know what is law, guidance and still a proposal

Malaysia's Personal Data Protection Department lists seven principles covering matters including notice, disclosure, security, retention, data accuracy and access. Using an AI service does not make those underlying responsibilities disappear.

There is also specific regulatory guidance to consider. The Personal Data Protection Commissioner's Automated Decision-Making and Profiling Guideline addresses automated processing with legal or similarly significant effects on individuals. Its recommended AI practices include human review and avoiding reliance on AI as the sole factor in decisions concerning someone.

Separately, as at 22 October 2026, JPDP's proposed AI and Personal Data Protection Framework is under public consultation, scheduled from 5 to 23 October 2026. It is a proposal, not a final framework that should be presented as an already enforceable checklist.

The practical distinction matters. Checking whether a receipt has a date is different from automatically rejecting an employee's entitlement or scoring their reliability. Get qualified advice before introducing consequential automated decisions, profiling or sensitive-data processing.

Map the workflow before granting access

Consider an illustrative distributor with employees in Klang Valley and Johor. Staff submit travel claims through email, and HR records them in Excel before managers approve them.

A sensible first workflow could extract the claim amount, identify missing receipts, find the correct approver and send reminders. It does not need to inspect every employee document.

Write down the intended data journey:

Where the request enters: an approved form, mailbox or employee portal.
Which fields are needed: employee reference, expense date, category, amount and receipt.
Which policy applies: the current approved claims policy, with a version date.
Where the result goes: a draft claims register and the assigned approver.
What requires human action: unclear receipts, policy exceptions and approval decisions.

Then list what the agent should not access. For this example, exclude disciplinary files, unrelated medical records and other employees' salary documents.

Ask your privacy adviser whether existing notices and the legal basis for processing cover the proposed use. Do not assume that an employee's earlier submission of a document permits every new AI use.

Give the agent access to the workflow it needs, not the entire department it serves.

Separate shared context from shared permissions

A business may want Finance, HR and Operations agents to work from consistent information. That does not mean each agent should see the same records.

For example, Finance may need an approved reimbursement amount and payment reference. A branch manager may need to know whether a staffing request is approved. Neither necessarily needs the supporting medical certificate.

Ask your implementation team to demonstrate these boundaries using test accounts:

Can an employee retrieve only their own claim status?
Can a manager access only requests within their responsibility?
Can Finance receive payment information without unrelated HR attachments?
Can administrators investigate failures without routinely reading confidential files?
Does removing a user's access also remove access through the AI interface?

Apply the same thinking to notifications. A WhatsApp message saying that three claims need approval is different from posting receipts and employee details into a branch group. Prefer a short notification with an authenticated route to the underlying record.

Treat these as design checks, not proof of legal compliance. Have permissions tested rather than accepting a presentation slide saying that access is controlled.

Ask the vendor where employee data goes

Before uploading live records, ask for written answers about the actual service configuration. A general assurance that the platform is secure is not enough to make a deployment decision.

Use these questions in your procurement discussion:

Which organisations receive employee data, including model providers and subcontractors?
Where are processing, storage and backups located?
Is submitted information used to train or improve any provider's models?
What is retained in prompts, attachments, conversation histories and diagnostic logs?
How are access, deletion and contract termination handled?
Who handles a suspected breach, and how will your organisation be informed?

Ask your adviser to assess any overseas processing against the applicable Malaysian cross-border transfer requirements. Do not treat either local hosting or an overseas provider's security certification as a complete compliance answer.

For retention, distinguish original business records from extra AI copies. Your organisation may need to retain a claim record, but that does not automatically justify retaining every prompt, extracted attachment and troubleshooting copy indefinitely. Malaysia's published retention principle says personal data should not be kept longer than necessary.

Agree a documented schedule that accounts for applicable record-keeping obligations, then test how deletion works in practice.

Keep policy calculations and consequential decisions separate

For an illustrative RM180 mileage claim, AI could extract the journey description and flag a missing receipt. A defined calculation should apply the approved mileage rate. A named approver should handle exceptions.

Do not ask a language model to invent the applicable rate, infer a missing policy or decide whether an employee deserves reimbursement.

Likewise, an agent can collect leave requests and route them to managers without deciding that frequent medical leave makes someone unreliable. Avoid turning an administrative workflow into employee profiling by accident.

Human review must also be usable. Give the reviewer the source record, relevant policy and reason for the flag, with a clear way to correct or override the result. An unexplained recommendation followed by an approval button is a weak control.

Where a form, fixed rule or existing HR system already handles the task reliably, keep it. Use AI where interpreting varied documents or messages adds value, not merely because the workflow involves people.

Pilot with deliberately awkward cases

Begin with synthetic examples or appropriately redacted records in a controlled environment. Include difficult cases, not just clean receipts.

For example, test a blurred image, an outdated policy, two employees with similar names and a request sent to the wrong manager. Also test whether an employee can persuade the agent to reveal someone else's claim.

Before rollout, agree acceptance checks:

No unauthorised employee-record access in the agreed tests.
Missing or uncertain information is flagged rather than invented.
Every approval is linked to an authorised person.
Corrections reach the destination record without creating duplicate entries.
The team can pause automation and continue manually.

The real effort includes cleaning employee identifiers, resolving inconsistent policies and teaching managers how exceptions work. Budget staff time for those jobs, not just the software connection.

Solver AI's approach is to validate the workflow, test it in a controlled environment and then roll it out. For HR, that validation should include data boundaries and human responsibility alongside time saved.

Explore workflow automation with controlled access

The aim is straightforward: remove the chasing and rekeying without giving an agent unnecessary access or authority. A narrower, well-tested workflow is a better starting point than connecting the entire HR drive.