Short answer: A workable AI policy is short, plain and specific. It states who may use AI for what, what data may be entered, how outputs are checked, who is accountable and how incidents are reported. Start with one or two pages, share it with staff, and review it at least yearly. Below is a template you can adapt, which is general information and should be checked against your legal and sector obligations. Review it with legal counsel, and check PDPC and sector regulator guidance (listed below).
Why have a policy?
Staff are likely using AI already. A policy sets boundaries, shows customers and regulators that you take care, and helps people use tools confidently. It supports, but does not replace, compliance with law. The PDPC's guidance on personal data in AI systems, listed below, explains how the PDPA applies.
Template: sections to include
1. Purpose and scope. Who the policy covers (employees, contractors, volunteers) and which tools (company-approved and personal accounts used for work).
2. Approved uses. Examples: drafting, summarising public information, brainstorming, code assistance with review.
3. Prohibited or restricted uses. For example: entering customer personal data, health data, identification numbers or confidential client information into unapproved tools; fully automated decisions about hiring, discipline or credit without human review; presenting AI output as verified without checking. The PDPC's guidelines on AI systems and on NRIC and other national identification numbers, listed below, explain the rules for personal data.
4. Data rules. What may be entered into which tools. Define categories such as public, internal, confidential and personal data.
5. Approval of new tools. Who reviews a new tool, what they check (data terms, security, location of processing, vendor reputation) and where the approved list is kept.
6. Human review. Which outputs must be checked before use, and by whom. Higher-impact decisions need more review. The IMDA Model AI Governance Framework has a section on determining the level of human involvement in AI-augmented decision-making.
7. Transparency. When you will tell customers, members or staff that AI is involved, and how they can reach a person.
8. Intellectual property and confidentiality. Staff should not input material they do not have the right to share, and should check outputs for third-party material before publishing.
9. Incident reporting. What counts as an incident (data leak, harmful output, wrong decision), who to tell, and how quickly.
10. Accountability and review. Name the policy owner, the review date and how staff can ask questions.
How do we roll it out?
- Draft with input from the people who actually use the tools.
- Have a lawyer or privacy adviser review it.
- Brief staff in 20 minutes with examples.
- Ask staff to confirm they have read it.
- Collect questions and update the policy.
What mistakes should we avoid?
- Copying a large organisation's policy that no one will read
- Banning everything, which pushes use underground
- Having no owner
- Never reviewing it
Keep learning
Free monthly newsletter at asia-ai.org. Membership options at asia-ai.org/join.
This article is general information, not advice.