Shadow AI at Work: What It Is and How Malaysian Firms Handle It
Shadow AI is staff using AI tools the company has not approved. Why it creates PDPA risk, how to find it without a witch hunt, and how to replace it.
Shadow AI means employees using AI tools for work without the company's knowledge or approval, for example a personal ChatGPT account for client emails. It creates PDPA risk because you cannot assess transfers, notify customers or investigate leaks for tools you do not know about. The fix is an approved alternative and a clear rule, not a ban.
Part of the guide: PDPA and AI in Malaysia: What Companies Must Know (2026 Guide)

On this page
Shadow AI is what happens when staff adopt AI tools faster than the company approves them.
Part of the guide: PDPA and AI in Malaysia. Checked 22 October 2026.
What counts as shadow AI?
- A personal ChatGPT, Gemini or Claude account used for work.
- An AI browser extension that reads emails or web pages.
- A free transcription app used for client meetings.
- AI features switched on inside software you already own.
- A WhatsApp-connected bot set up by one enthusiastic team member.

Why is it a PDPA problem?
Because the PDPA obligations attach to the data, and you cannot meet them for tools you do not know exist. Specifically: you cannot tell customers about disclosures (Notice and Choice), you cannot show security controls (Security principle), you cannot check where data goes (the section 129 transfer conditions), and you cannot report a breach you never saw. The PDPA guideline on data breach notification sets a 72-hour deadline to notify the Commissioner once a notifiable breach is identified, so unknown tools make that deadline hard to meet. Health information counts as sensitive personal data, which raises the stakes for clinics and healthcare suppliers. We do not cite adoption percentages here because we have not verified a Malaysian figure; ask your own staff.
How do you find it without a witch hunt?
- Announce an amnesty. Say you want a tool list, not names for blame.
- Run a five-question survey. Which tools, for what tasks, on what accounts, with which data, how often.
- Check the technical traces. Browser extensions, SaaS sign-ups with company email, expense claims for subscriptions.
- Sort what you find. Approve, replace or stop.

What do you do with what you find?
| Finding | Action |
|---|---|
| Useful tool, business plan available | Buy the business plan, move users to company accounts |
| Useful tool, risky data | Approve for non-personal data only; add rules |
| Redundant tool | Replace with an approved one |
| Risky tool with no safe plan | Stop, explain why |
What does shadow AI cost compared with an approved tool?
You pay either way. The difference is control. A worked comparison with stated assumptions: 20 staff each pay for a personal plan at a price you set yourself, say USD 20 a month. That is 20 subscriptions you cannot see, cannot cancel and cannot move to a different account when someone leaves. A single team plan for the same people, with admin controls, is usually priced per seat too, so the licence cost is similar, but you gain a central owner, a tool list and an exit route. Check the current price and terms on the vendor page before you decide. Anthropic, for example, publishes commercial terms stating it does not train models on customer content from its commercial services, while consumer terms differ. Read the plan you actually buy.
What makes people comply?
An approved tool that is as good as the one they use today, a rule short enough to remember, and a manager who uses the same tools. Without the approved alternative, policies lose.
Hypothetical example: a 40-person logistics firm in Shah Alam finds three personal AI accounts used for customer emails. It buys a team plan, writes a one-page rule and runs a 45-minute session, so staff have an approved tool in place of the personal accounts.
The fastest cure for shadow AI is an approved tool and staff who know how to use it well, which is what Generative AI at Work delivers in two days.
What next?
Start with the AI policy template and the employee version. We offer an AI opportunity audit that finds the tools and the use cases. Book a call.
Sources
- Personal Data Protection Department Malaysia, Act 709 (checked November 2026)
- PDPA guideline on cross-border personal data transfer
- PDPA guideline on data breach notification
- Anthropic, Commercial Terms of Service
General information, not legal advice.
Frequently asked questions
How do I find out which AI tools staff use?
Ask. An anonymous survey and a short team conversation find more than blocking software. Add a review of browser extensions, expense claims and SaaS sign-ups.
Should we block ChatGPT on the office network?
Blocking moves use to phones and home laptops. Pair any block with an approved alternative, or people will go around it.
What should the AI rule say?
Name the approved tools, list what data may never go in (customer names, IC numbers, health details, contracts), say who to ask, and say how to report a mistake. One page is enough.
Is shadow AI a PDPA breach?
Not automatically. It becomes one when personal data is processed without a lawful basis, proper security or a transfer basis, or when a breach follows. Unknown tools make it hard to show you complied.