AI policy template for Singapore organisations, with plain-language guidance

Singapore's recent breaches show why an AI policy must be part of everyday data security, not just a list of approved chatbots. In early November 2026, breaches disclosed by Singapore Airlines' technology supplier, SPH Media and a school meal-ordering platform exposed different kinds of personal information, and Fullerton Health gave notice of a separate incident. These reports do not establish that AI caused the attacks, and at SPH Media AI involvement is unconfirmed. Separately, the Cyber Security Agency (CSA) warns that AI is changing the speed, scale and sophistication of cyber threats.

A PDPC case gives a direct AI-use lesson: a marketing script built with generative AI exposed 95,364 Bee Cheng Hiang customers' email addresses after inadequate testing. The practical response is to control which tools staff use, what data those tools can reach, who reviews outputs and code, and who approves consequential actions. The PDPC's generative AI guidelines and its advisory on unauthorised AI tools point the same way. A related joint advisory on personal AI agents from CSA, GovTech and IMDA followed on 6 November 2026.

An AI policy tells your staff four things: which AI tools they may use for work, what information must never go into them, who checks AI output before it reaches a customer, and what to do when something goes wrong. A small or mid-sized organisation can fill in the template below, sign it and start using it.

01What Singapore's official guidance says

A July 2026 parliamentary reply said that "many AI risks are already addressed through existing legislation" (MDDI, 7 July 2026). Three kinds of official material shape a sensible policy: the law you must follow, frameworks you can adopt, and tools you can use to test AI systems.

The law: the Personal Data Protection Act

The PDPA is Singapore's baseline law for personal data, enforced by the Personal Data Protection Commission (PDPC) (PDPA overview). It does not stop you using AI, but it applies whenever staff put personal data into an AI tool. Its obligations include consent, purpose limitation, notification, protection, retention limitation, transfer limitation, data breach notification and accountability (Advisory Guidelines on Key Concepts in the PDPA). Three points matter most for an AI policy:

• Protect the data. Section 24 requires organisations "to make reasonable security arrangements to protect personal data in their possession or under their control" (PDPC advisory, 30 Sep 2026).
• Name a data protection officer. "Every organisation in Singapore that handles personal data must appoint at least one DPO", who can be an employee or an outside provider (PDPC: Getting started as a DPO).
• Report notifiable breaches on time. A breach is notifiable when it is likely to result in significant harm to individuals, or when it affects 500 or more individuals. You must notify the PDPC no later than three calendar days after determining that the breach is notifiable (PDPC: Report your organisation's data breach).

PDPC guidance on personal data and AI

• [Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems](https://www.pdpc.gov.sg/organisations/regulations-decisions/regulatory-guidance/advisory-guidelines-on-use-of-personal-data-in-ai-recommendation-and-decision-systems) (1 March 2024). These explain when you need consent and when exceptions such as business improvement or research may apply, and encourage organisations to describe their AI safeguards in written policies, "especially where the outcome has high impact on consumers".
• [Advisory Guidelines on Use of Personal Data in Generative AI](https://www.pdpc.gov.sg/organisations/regulations-decisions/regulatory-guidance/advisory-guidelines-on-use-of-personal-data-in-generative-ai) (20 July 2026). These treat prompts, inputs, generated outputs and agent activity as data to protect. They expect organisations that deploy generative AI to educate users "on the specific types of personal data that should be input into their systems", and encourage them "to develop clear written policies and document processes".
• [Advisory on unauthorised AI tools](https://www.pdpc.gov.sg/media-events/advisory-for-data-protection-risks-when-using-unauthorised-ai-tools-for-the-development-of-software-or-scripts) (30 September 2026). This warns that "without clear policies and asset management controls governing the adoption of AI tools, employees may independently introduce unauthorised AI tools", which it calls "Shadow AI". It recommends an inventory of AI tools, clear written policies, staff training on what data may go into AI tools, and human review of AI output before use.

The frameworks: voluntary, and useful

• [Model AI Governance Framework](https://www.pdpc.gov.sg/organisations/resources/guidance-by-topic/singapores-approach-to-ai-governance) (second edition, 21 January 2020), from IMDA and the PDPC. Its two principles are that AI decisions should be explainable, transparent and fair, and that AI systems should be human-centric. It covers four areas: internal governance (clear roles, procedures and staff training), the level of human involvement, operations management, and stakeholder communication, including making AI policies known to users. A self-assessment guide, ISAGO, goes with it.
• [Model AI Governance Framework for Generative AI](https://aiverifyfoundation.sg/resources/mgf-gen-ai/) (30 May 2024), from IMDA and the AI Verify Foundation. Its nine dimensions include accountability, data, incident reporting, testing and assurance, security and content provenance.
• [Model AI Governance Framework for Agentic AI](https://www.mddi.gov.sg/newsroom/singapore-launches-new-model-ai-governance-framework-for-agentic-ai--/) (launched 22 January 2026), from IMDA, for AI agents that take actions. Its four dimensions are bounding agents' access and powers, human approval checkpoints, controls across the agent's lifecycle, and user training.
• [CSA Guidelines and Companion Guide on Securing AI Systems](https://www.csa.gov.sg/resources/publications/guidelines-and-companion-guide-on-securing-ai-systems/) (15 October 2024), on securing AI across its lifecycle.

The testing tools: AI Verify and Project Moonshot

AI Verify is a testing framework and software toolkit for assessing traditional and generative AI systems against 11 AI governance principles. Project Moonshot is a toolkit for testing and red-teaming applications built on large language models. Both come from IMDA and the AI Verify Foundation. Treat them as testing aids, not as certification or proof of legal compliance. They matter most when you build, customise or sell AI systems. An organisation that only buys off-the-shelf tools will rarely run them, but still needs the rules in this policy.

Sector rules

The Monetary Authority of Singapore published Guidelines on Artificial Intelligence Risk Management on 7 October 2026. These guidelines set out MAS's supervisory expectations for all financial institutions and take effect on 7 October 2027. For law firms, Annex C of the Ministry of Law's Guide for Using Generative AI in the Legal Sector (6 March 2026) has a sample internal generative AI policy. If you are regulated, check with your regulator or professional body first.

02Start with a one-page card

A policy only works if people remember it when they open an AI tool. So start with a one-page card of six one-line rules. Our journal post Put your AI rules in plain language shows how to write a rule clearly enough that a tool can check a draft against it.

The six rules are section 3 of the template. For a 20-minute exercise to draft your own one-page version first, see Write the policy before choosing the filter.

03The template at a glance

The template has twelve short sections. Some draw on the PDPA, which is law. Others draw on voluntary guidance from IMDA, the PDPC and the CSA. The purpose, the scope, the six headline rules and the review cycle are your own choice. The mapping below is Metacamp's reading of the guidance, not an official crosswalk.

04The AI policy template (copy and adapt)

Copy everything from the policy title to the staff acknowledgement into your own document. Replace every item in [square brackets], delete anything that does not apply, and add the rules your sector needs. If a new joiner cannot follow a sentence, rewrite it.

[Company name] AI Use Policy

Version [1.0] · Approved by [name, role] on [date] · Next review due [date]

Policy owner: [name, role] · Data Protection Officer: [name, business contact]

1. Purpose

[Company name] uses AI tools to work faster and serve customers better. This policy explains how we use them safely, so that we protect personal data and confidential information, stay within the law, and keep a person responsible for every piece of work. It is written to help us meet our obligations under Singapore's Personal Data Protection Act 2012 (PDPA), and it draws on the Model AI Governance Framework published by IMDA and the PDPC.

2. Scope

• This policy applies to all employees, interns and contractors [and representatives] who use AI tools for [Company name] work, on company or personal devices.
• An "AI tool" is any software that writes or analyses text, images, audio, video or code, or that makes recommendations or decisions. This includes AI features built into software we already use, such as email, office suites, meeting apps and our customer relationship management (CRM) system.
• [Optional: AI systems that we build or sell to customers follow [separate process] as well as this policy.]

3. Our six rules

1. 01 / 06

   A person owns every piece of AI-assisted work. If your name is on it, you have checked it.

2. 02 / 06

   Use only the tools on our approved list, with your work account.

3. 03 / 06

   Never put Restricted information (see section 5) into an AI tool.

4. 04 / 06

   Check AI output before a customer sees it, and before it is used in a decision about anyone.

5. 05 / 06

   Tell people when they are dealing with AI.

6. 06 / 06

   Report problems the same day. Nobody is blamed for reporting early.

4. Approved tools

• The AI Tool Register [link] lists every approved tool, what it may be used for, which information it may handle, who owns it and when it is next reviewed. If a tool is not on the register, it is not approved.
• Use company accounts only. Do not use personal or free accounts for work.
• [Where a tool allows it, the account owner switches off the use of our data to train the vendor's models.]
• To ask for a new tool or AI feature, contact [AI policy owner]. Before approving it, they check what information the tool will see, where that information is stored and processed (including outside Singapore), whether the vendor uses it to train its models, the security settings, and the cost.
• Do not connect an AI tool to email, calendars, file storage, customer systems or payments (through plug-ins, browser extensions or "agent" features) unless the register says that connection is approved.

5. Information rules

We sort information into three classes before it goes into any AI tool.

• Restricted: never put into an AI tool. NRIC, FIN and passport numbers; bank, card and account numbers; passwords and access keys; health and medical information; information covered by a confidentiality agreement; customer or employee files you have not been approved to use; [add your own].
• Internal: approved tools only, and only what the task needs. Internal plans and reports; customer names and contact details for an approved task; draft contracts and pricing; meeting notes and recordings; [add your own].
• Public: fine in any approved tool. Information we have already published; public news and reports; general questions; your own drafts that contain no personal or confidential details.

If you are not sure which class something belongs to, treat it as Restricted and ask [AI policy owner or Data Protection Officer]. Remove names and other identifying details whenever the task allows. Only use personal data with AI for purposes the people concerned have been told about and, where needed, have agreed to.

6. Checking AI output

• Check every AI output before you use it. You are responsible for it as if you wrote it yourself.
• Check facts, figures, names, dates, links, quotes and any legal or regulatory statement against a reliable source. AI tools can give confident answers that are wrong.

Match the check to the risk:

• Internal drafts and notes: you check it.
• Anything a customer or the public will see: you check it [and, for [list of content types], a second person reviews it].
• Decisions about people or money (hiring, performance, pricing, credit, claims, advice to clients): AI may help, but a qualified person makes the decision and records how it was made.

Do not let an AI tool send, publish, pay, delete or share anything on our behalf without a person approving that specific action.

7. Being open about AI

• Tell customers when they are talking to an AI chatbot or assistant rather than a person, and give them a way to reach a person.
• If AI helps make a decision that affects a customer, be ready to explain in plain language what the AI did and how a person was involved.
• Label AI-generated or AI-edited images, audio and video where people could mistake them for real. Never use AI to imitate a real person without their permission.
• Respect copyright and confidentiality. Do not upload material we do not have the right to use.

8. Security and AI agents

• Protect every AI account with a strong password and two-factor authentication.
• Every [month], [AI policy owner] reviews which AI tools are connected to company accounts and removes access that is no longer needed.
• AI tools can be tricked by hidden instructions inside documents, emails and web pages. Take extra care with AI output based on outside content, and never approve an action you do not understand.
• AI agents (tools that take actions, not just write text) need [AI policy owner]'s approval before they connect to any company system, and must ask a person before any action that cannot easily be undone.

9. Roles

• AI policy owner, [name, role]: keeps this policy and the AI Tool Register up to date, approves new tools and answers questions.
• Data Protection Officer, [name]: advises on personal data and leads the response to any suspected data breach.
• Managers: make sure their team has read this policy and follows it.
• Everyone: follows this policy, asks when unsure and reports problems quickly.

10. When something goes wrong

If you put Restricted information into an AI tool, find that wrong or harmful AI output has already gone out, or notice a tool behaving unexpectedly:

1. 01 / 05

   Stop using the tool for that task.

2. 02 / 05

   Report it to [AI policy owner] and [Data Protection Officer] the same day, through [channel].

3. 03 / 05

   Write down what happened: which tool, what information, when, and who may be affected. Do not delete anything.

4. 04 / 05

   If personal data may be involved, the Data Protection Officer assesses whether it is a notifiable data breach under the PDPA and, if it is, notifies the PDPC and affected individuals as the law requires.

5. 05 / 05

   Fix the cause, and update this policy or the AI Tool Register if needed.

We do not punish anyone for reporting a mistake promptly.

11. Training

• Everyone completes [AI basics training] within [timeframe] of joining, and a refresher every [year].
• People who use AI for [customer-facing work or decisions about people] also complete [role-specific training].

12. Review and breaches

• [AI policy owner] reviews this policy every [six or twelve] months, and sooner when we adopt a major new tool, when the law or official guidance changes, or after an incident.
• Breaking this policy may lead to action under our [disciplinary policy].

Staff acknowledgement

I have read and understood the [Company name] AI Use Policy, version [1.0].

Name: [name] · Signature: [signature] · Date: [date]

05How to put the policy in place, step by step

Six steps take you from the blank template to a signed policy.

1. 01 / 06
   List the AI tools already in use.

   Ask each team, and include AI features inside software you already pay for. The PDPC's Shadow AI advisory recommends this inventory.

2. 02 / 06
   Sort your information.

   With your data protection officer, agree what is Restricted, Internal and Public, starting from section 5.

3. 03 / 06
   Fill in the template.

   Add sector rules from your regulator or professional body.

4. 04 / 06
   Get it approved.

   A director or owner signs and dates it.

5. 05 / 06
   Brief the team.

   Walk through the one-page card, then ask everyone to sign the acknowledgement.

6. 06 / 06
   Set the next review date.

   Put it in the calendar now.

06What information should never go into an AI tool

Sort information into three classes before anyone pastes it into a tool.

The Ministry of Law's sample policy for law firms has a similar rule: "Do not input client names, case details or sensitive data into public GenAI platforms."

• Use your own examples. Staff follow real document types ("signed client agreements", "payroll exports") more reliably than abstract categories.
• Strip identifiers first. Many tasks, such as rewording an email, work as well with names and numbers removed.
• Check where the data goes. Many AI tools process data outside Singapore, and the PDPA's Transfer Limitation Obligation applies when personal data is transferred overseas. Ask your data protection officer before approving a tool for personal data.
• Watch the free accounts. A personal account on a free AI service sits outside your company's settings and contracts, which is why rule 2 asks for work accounts.

07How closely should people check AI output

The Model AI Governance Framework describes three levels of human involvement, and suggests choosing between them by weighing how likely harm is and how severe it would be.

• Human in the loop: the person keeps "full control and the AI only provid[es] recommendations or input".
• Human over the loop: the person is "in a monitoring or supervisory role, with the ability to take over control" when something unexpected happens.
• Human out of the loop: "there is no human oversight over the execution of decisions".

A sensible default is a person in the loop for anything a customer will see and any decision about people or money. Use "over the loop" for internal tools whose results someone reviews regularly, and "out of the loop" only for low-stakes tasks that are easy to undo. The PDPC's Shadow AI advisory warns against "excessive trust in AI-generated outputs without human verification". If your team approves actions proposed by AI agents, our 10-second approval test gives four questions to ask first.

08Keep one register of approved AI tools

The register answers the question staff ask most often: may I use this tool for this job?

Keep it in a shared spreadsheet that everyone can read and only the policy owner can edit. For each tool, record the account type (a company plan, not a personal one), what it is approved for, which information classes it may handle, who owns it and when it will next be reviewed. The PDPC's Shadow AI advisory suggests capturing each tool's "purpose, data flow information, and intended users" and reviewing the inventory regularly. List free personal accounts explicitly as not approved, so nobody has to guess. Our monthly access audit checks what each tool can still reach.

09What to do when something goes wrong

People need to report mistakes quickly, and you need to know which ones the law requires you to report onwards.

The template's same-day reporting rule is your company's own rule, not a statutory deadline. A wrong figure in an internal draft is a quality problem to correct. When personal data may have been exposed, treat it as a possible data breach:

• Assess it quickly. Organisations must take "reasonable and expeditious steps" to assess whether a breach is notifiable "within 30 calendar days" (PDPC: Guide on Managing and Notifying Data Breaches).
• Notify the PDPC on time. A breach is notifiable when it is likely to cause significant harm to the people affected, or when it affects 500 or more people. If it is, notify the PDPC no later than three calendar days after you determine that it is notifiable.
• Tell individuals where significant harm is likely. Do this at the same time as, or after, you notify the PDPC, unless an exception applies. A breach that reaches 500 people does not by itself mean individuals must be notified.

The PDPC's advisory on unauthorised AI tools gives an illustration: an employee built an email script with a public AI tool on a personal account, skipped testing and review, and exposed every customer's email address to the whole mailing list. The advisory calls this "a potential breach of the organisation's obligations under the PDPA".

10Keep the policy current

Review the policy on a fixed date, and early when you adopt a new AI tool or feature, when the law or official guidance changes, when something goes wrong, or when staff start using AI for a new kind of task.

At each review, update the AI Tool Register and the one-page card as well as the full policy, so the three never disagree.

11Training your team on the policy

Staff training is one of the Model AI Governance Framework's internal governance measures, and the PDPC's generative AI guidelines and Shadow AI advisory both recommend educating staff on what data may go into AI tools. Section 11 of the template sets basic training for everyone and extra training for customer-facing work or decisions about people.

For outside help, Metacamp's GAI101 Applied Generative AI Fundamentals (course code TGS-2026065683) is an 18.5-hour, in-person course for working professionals, with no coding needed. Learners turn a real workplace task into an AI assistant or workflow and learn to check the result. Check the dates and timetable for the run you choose.

• Fee and funding. Each course has a full fee of S$1,750. Metacamp is not GST-registered, so no GST is added. After course fee funding from SSG, now part of SWDA (IRAS)), Singaporeans aged below 40 pay S$875 (50% subsidy) and Singaporeans aged 40 and above pay S$525 (70% subsidy). Singapore Citizens aged 25 or over may use up to S$500 of Opening SkillsFuture Credit towards the remaining fee. Indicative fees checked 11 November 2026; course-run and learner eligibility apply; confirm at enrolment.
• Private team runs. A private GAI101 run for your team gets the same subsidies as a public run, provided it is the full 18.5-hour course as registered with SSG. See Metacamp for Teams.
• SkillsFuture Enterprise Credit (SFEC). The current SFEC expires on 30 November 2026. From 1 December, the redesigned SFEC gives eligible companies a fresh S$10,000 credit in an online wallet for supported workforce transformation. Check programme and employer eligibility before booking (MOM written answer, 3 November 2026).
• Enterprise Innovation Scheme (EIS). EIS) may apply to qualifying employer-paid training. Check your exact course code and supporting period on the official list and consult your tax adviser. Eligibility is not a blanket promise.

12FAQ