Is an AI Agent Safe for Customer Data? PDPA Checklist

An AI agent can be used safely with customer data in Malaysia if the business treats it like any staff member handling personal data: collect only what is needed, tell customers why, restrict access, keep records and have a clear route for corrections and breaches. The PDPA applies to the business, not just the software, so the checklist below is for both.

Is an AI agent safe for customer data under the PDPA?

An AI agent is as safe as the rules around it. The Personal Data Protection Act 2010 puts the responsibility on the business that collects the data, so you remain accountable even when an AI answers the chats.

That sounds heavy, but it is the same standard you already apply to a receptionist who sees customer names, phone numbers and sometimes IC numbers. You would train that person, limit what they can open and check their work. An AI agent needs the same treatment, and the good news is that software is easier to limit and easier to audit than people.

This article is a practical checklist for Malaysian SMEs. It is not legal advice, and the law has been updated recently, so read the official sources linked here and ask a lawyer about your own situation. For the wider idea of AI employees with defined jobs, see what an AI agent office is.

What customer data does an AI agent actually see?

An AI agent sees whatever customers type or send, plus whatever you connect to it. On WhatsApp that usually means a phone number, a profile name, the messages, and sometimes photos, voice notes or documents.

Map this before you launch. Many owners are surprised by how much arrives without being asked: a customer sends a photo of their IC to prove identity, a screenshot of a bank transfer, or a voice note describing a health issue. Each of these is personal data, and some can be sensitive.

The PDP Department explains what counts as sensitive personal data in its frequently asked questions. Read it before you decide what the AI may handle.

Which PDPA principles matter most when you use an AI agent?

The PDPA sets out principles covering how personal data is collected, used, disclosed, secured, kept and corrected. The full text is on the department's PDPA 2010 page, and the key ideas translate into simple chat habits.

• Notice and purpose. Tell customers what you collect and why. A short privacy line in the first message or on your page is a start.
• Use only for that purpose. Data collected to book a service should not quietly become a marketing list.
• Security. Control who and what can open the data.
• Keep it only as long as needed. Decide a retention period and stick to it.
• Access and correction. A customer can ask what you hold and ask for corrections.

The law was also amended. The department publishes the Personal Data Protection (Amendment) Act 2024, and legal summaries such as DLA Piper's note on breach notification and data protection officer guidelines describe the newer duties around breach reporting and appointing a data protection officer. Check which of these apply to a business of your size.

What is a PDPA checklist for using AI agents?

The checklist below has ten items. Work through them before going live and again every few months.

1. List the data. Write down every kind of personal data the AI will see, including images and voice notes.
2. Write the purpose. One sentence for each: why do we collect this?
3. Cut the questions. Remove any request for data the next step does not need.
4. Give each AI employee its own scope. The booking role needs a name and time slot. It probably does not need a salary slip.
5. Set a human rule for sensitive cases. Health, financial hardship, children and complaints go to a person.
6. Add a privacy notice. Link it from your WhatsApp profile and mention it when you first collect data.
7. Control staff access. Give each user only the chats they need and remove access when they leave.
8. Keep logs. Be able to show what the AI said and what it did.
9. Decide retention. Set when old chats and files are deleted or archived.
10. Plan for a breach. Know who decides, who tells customers and who reports, so nobody improvises under stress.

How should access be limited between AI employees?

Each AI employee should see only the information its job needs, the same way a receptionist does not open the accounts file. This is the strongest practical idea in an AI agent office, because roles are separate by design.

A report to the boss rarely needs names. Ask for summaries such as how many enquiries and bookings came in, with personal detail available only on request.

What does privacy-aware chat look like?

What the conversation looks like

Illustrative example.

The AI did four things right. It did not comment on a medical condition. It asked for the minimum needed to book. It told the customer who would see the photo. And it passed the case to a human. In regulated trades such as beauty and aesthetics, the AI informs and books while licensed people advise.

What should you ask an AI vendor before you sign?

Ask for written answers, because verbal assurances are hard to rely on later. These are the questions a careful owner asks.

• Where is customer data stored, and in which country?
• Who at the vendor can see our chats, and under what conditions?
• How long is data kept, and can we have it deleted when we leave?
• Is customer data used to train AI models for other companies?
• Which AI model providers receive the message text?
• How do you handle and report a breach?
• Can we export or erase one customer's data on request?
• Are you using the official WhatsApp Business API?

That last point is a privacy question too. Unofficial tools can put your number and your customers' data at risk, as explained in our guide to unofficial WhatsApp API risks. Also be careful about consent for marketing, which has its own rules in WhatsApp's Business Messaging Policy and our opt-in guide.

Who should be extra careful?

Businesses that hold sensitive data should be most careful, because mistakes there cost more. This includes clinics, aesthetic centres, moneylenders and anyone who collects IC copies or payslips.

For example, shops that take IC and payslip photos during phone instalment enquiries should read our post on handling IC and payslips on WhatsApp, and clinics should read the guide on clinic reminders and the PDPA.

A business that only answers general questions about opening hours and price lists carries much less risk. Not every AI agent needs to touch sensitive data, and the safest design is often to keep it out of the chat altogether.

How does a ChatsHero AI Agent Office handle customer data?

ChatsHero, a Malaysian AI agent platform, runs each AI employee from your own SOP, price list and FAQ, so you decide what each role is told to collect and what must go to a person. Human staff can step in, and conversations are tagged so your team can see what happened.

For data questions such as hosting location, retention and model providers, ask us directly and get the answers in writing, using the vendor questions above. A free 30-minute Zoom demo is a good place to raise them while you see a working AI Agent Office.

Frequently Asked Questions