What Makes a Good AI Policy? Seven Tests for Malaysian Firms
Seven practical tests to judge whether your company AI policy works: specific tools, data classes, named owner, review duty, incident path, training and PDPA fit.
A good AI policy is specific, short and used. It names approved tools, lists data that must never be entered, assigns an owner, says who reviews output, gives staff a way to report mistakes and is backed by training. If staff cannot tell you what it says, it is not working.
Part of the guide: PDPA and AI in Malaysia: What Companies Must Know (2026 Guide)

On this page
Many policies exist and almost none are read. These seven tests separate the ones that work.
Part of the guide: PDPA and AI in Malaysia. Checked 22 October 2026.
What are the seven tests?
| # | Test | Passes when |
|---|---|---|
| 1 | Specific tools | Named tools and plan types, not "AI tools" |
| 2 | Data classes | A list of data that is never allowed, allowed with care, free to use |
| 3 | Named owner | A person (not a department) updates it |
| 4 | Human review | It says what output must be checked and by whom |
| 5 | Incident path | A phone number or WhatsApp group, a time limit, no blame for early reporting |
| 6 | Trained | Everyone has seen worked examples, not just a PDF |
| 7 | PDPA fit | Links to the privacy notice, vendor contracts and cross-border check |

Why do specific tools matter?
A rule like "no confidential data in public AI" fails because staff do not know which tools count. "ChatGPT free, Gemini free and any personal account are not allowed for client work; the company Team plan is" can be followed.
What belongs in the data classes?
Three lists are enough: never (IC numbers, health records, payroll, passwords, unpublished results), only in approved tools (customer emails, contracts with the names removed), free (public web content, your own drafts). The PDPA treats health and biometric data as sensitive, so put them in the "never" column unless a specific system is cleared.

What does human review mean in practice?
Name the cases. Examples: any message that goes to a customer; anything with figures, legal clauses or medical content; any decision about a person such as hiring or credit. The Data Integrity and Access principles of the PDPA support this: wrong data about a person is your problem, not the tool's.
What makes an incident path work?
Staff report if they will not be punished for it. A 24-hour reporting rule plus an instruction to escalate suspected personal data breaches immediately, because notification to the Commissioner can be due within 72 hours for breaches likely to cause significant harm or affecting more than 1,000 individuals, as the PDPA data breach notification guideline sets out.
What does a policy that fails look like?
A worked example with a made-up policy: "Staff must use AI responsibly and must not share confidential information." Score it against the table. Tools: fail, none named. Data classes: fail, "confidential" is undefined. Owner: fail. Human review: fail. Incident path: fail. Trained: unknown. PDPA fit: fail. It passes none of the seven. Rewriting it means naming the company plan, adding the three data lists, putting a person's name on it and adding a reporting contact, which fits on one page. For a broader risk structure, the NIST AI Risk Management Framework offers voluntary guidance on governing AI use; a small firm can borrow its idea of an owner and a regular review without adopting the whole framework.
How do you test it?
Ask three people at random: "May I paste this customer email into our AI tool?" If you get three different answers, rewrite. Do this every review.
Directors write the first version of these rules in AI Strategy for Leaders, and staff test them against real tasks in Generative AI at Work.
What next?
Start from the AI policy template, then run a staff session. See our courses or book a call.
Sources
(checked November 2026)
- PDPA guideline on data breach notification
- Personal Data Protection Department Malaysia, Act 709
- NIST, AI Risk Management Framework
General information, not legal advice.
Frequently asked questions
How can I tell if my AI policy is too vague?
If it says 'use AI responsibly' or 'protect confidential data' without naming tools, data types or people, it is too vague. A staff member should be able to decide in 30 seconds whether a task is allowed.
Should an AI policy ban ChatGPT?
Usually no. Bans push use into private accounts. Offer an approved tool and set rules for what data it may hold.
Who should own the AI policy?
One named person with authority, such as the managing director, operations lead or DPO. A committee that owns it in name tends to leave it unchanged. The owner updates the tool list and runs the review.
How often should we update it?
At least every six months, and whenever you add a tool, change a vendor's plan, or a relevant law changes.