In this guide · 7 sections
- 01What an AI risk register is, and why a small company needs one
- 02The seven columns of a one-page AI risk register
- 03Worked example: a typical 15-person services firm
- 04How to build the register in an afternoon
- 05Controls that work for small teams
- 06Do it yourself or get help?
- 07Frequently asked questions
Short answer
List each AI tool or use in a row with seven columns: use, data involved, owner, what could go wrong, impact, controls and next review date. Score impact simply (low, medium, high), fix the high ones first and review the page every quarter or when you add a new tool.
What an AI risk register is, and why a small company needs one
A risk register is a list of things that could go wrong, how bad they would be and what you are doing about them. An AI risk register does the same for AI tools: chat assistants, website chatbots, AI features inside your accounting or CRM software, and any automation that calls an AI model.
Singapore's approach is guidance rather than law. IMDA's Model AI Governance Framework for Generative AI (30 May 2024) starts with accountability: someone answers for each AI system. PDPC's Advisory Guidelines on Use of Personal Data in Generative AI, published 20 July 2026, treat businesses that use AI tools as System Deployers with primary responsibility for PDPA compliance. A register is the simplest way to show you have thought about both.
Our IMDA AI governance framework guide and PDPC generative AI guidelines explainer cover the source documents. This page turns them into a working table. It is general guidance, not legal advice.
The seven columns of a one-page AI risk register
Keep it to seven columns so it fits on one landscape page. Anything longer tends not to get updated.
| Column | What to write | Example |
|---|---|---|
| 1. AI use | Tool and what it is used for | ChatGPT Business for drafting proposals |
| 2. Data involved | Types of data that go in or come out | Client names, project scope; no NRIC or payment data |
| 3. Owner | One named person who answers for it | Operations manager |
| 4. What could go wrong | The realistic failure, in a sentence | Wrong figures in a proposal; client data pasted into a personal account |
| 5. Impact | Low, medium or high, with a reason | Medium: embarrassing, fixable, possible PDPA issue |
| 6. Controls | What you do to reduce the risk | Company workspace only, human review before sending, training |
| 7. Next review | Date, plus trigger events | Jan 2027, or when the vendor changes terms |
If you want more detail later, add columns for vendor, where data is stored and whether a data processing agreement is in place. Start without them.
Worked example: a typical 15-person services firm
Notice that the risk is rarely the model itself. It is usually the data that goes in, who sees the output and whether a person checks it before it reaches a customer.
| Impact level | Rule of thumb | Typical uses |
|---|---|---|
| High | Customers see output without a human check, or sensitive personal data is involved | Website or WhatsApp chatbot, AI screening of job applicants, AI agents that send emails |
| Medium | Output is used externally after a human check, or client data goes in | Drafting proposals, meeting notes, Copilot over shared files |
| Low | Internal use with no personal or confidential data | Brainstorming, rewriting internal notices |
How to build the register in an afternoon
Ask staff what AI they use
Send a two-question form: which AI tools do you use for work, and what do you put into them? Promise no blame. Our guide to finding shadow AI has more ways to spot unapproved tools.
Check your software for AI features
Accounting, CRM, helpdesk and email tools often switch AI features on by default. Note each one and whether it uses your data.
Write one row per use
Fill in the seven columns. If nobody wants to be the owner, that is a sign the tool should not be used.
Rate impact and pick controls
Use the low, medium, high rule above. For each high row, add at least one control that stops bad output reaching customers unchecked.
Agree the top three fixes
Typical first fixes: move staff from personal to company accounts, add human review to the chatbot, switch off training on your data where vendors allow it.
Set the review date
Put a quarterly review in the calendar and note trigger events: new tool, new vendor terms, new type of data, or a complaint.
Controls that work for small teams
Large-company frameworks list dozens of controls. For a small business, these cover most of the risk in practice:
Baseline AI controls
For agents that take actions by themselves, IMDA's January 2026 agentic AI framework adds two practical rules: limit what an agent can access from the start, and set checkpoints where a person must approve before it acts. Our AI policy and staff training checklist turns the people side into a short policy.
Do it yourself or get help?
Most small companies can build the register themselves in an afternoon using the columns above. The harder parts are getting honest answers from staff about the tools they use and turning the high-risk rows into working controls.
IT Sorted is a Singapore-based team of developers with over 10 years of experience. Security Awareness Training (S$390) covers safe AI use alongside phishing and passwords, and the PDPA Compliance Pack (S$690) puts your data inventory, policies and notices in order so the AI register has something to sit in. We do not give legal advice; for regulated activities, check with a lawyer.
Frequently asked questions
No. Singapore has no general AI law requiring a register. IMDA's Model AI Governance Frameworks are voluntary, and PDPC's generative AI guidelines are advisory. However, PDPC expects organisations using AI tools with personal data to understand and document how data is used, and a register is the simplest way to do that.
At minimum: each AI use, the data it involves, a named owner, what could realistically go wrong, an impact rating, the controls in place and a next review date. Larger organisations add vendor, data location, contracts and testing results, but seven columns on one page are enough for most small companies.
Review it every quarter, and also whenever you add a new AI tool, a vendor changes its terms or data handling, staff start using AI for a new type of data, or something goes wrong. A register that was accurate a year ago is likely to miss tools staff have started using since.
Yes. Any AI system that customers interact with should be in the register, usually rated high impact because customers see answers without a human check and it may collect personal data. Controls include answering only from your approved content, logging and reviewing conversations, and an easy way to reach a person.
One person, usually the data protection officer, operations manager or owner, keeps the register itself. Each row also has its own owner, the person responsible for that tool day to day. In a small company these may be the same person, which is fine as long as the reviews actually happen.