In this guide · 8 sections
- 01Before the session: what to prepare
- 02The 45-minute AI training session plan
- 03Part 1: approved tools and accounts
- 04Part 2: what never to paste into an AI tool
- 05Part 3: checking AI output before you use it
- 06Parts 4 and 5: telling customers, and reporting mistakes
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Run one 45-minute session in five parts: approved tools (5 min), what never to paste (10 min), checking output (10 min), telling customers (5 min), and reporting mistakes (5 min), with a 10-minute hands-on exercise. Base it on the CSA-IMDA advisory's three practices and finish with a signed one-page checklist.
Before the session: what to prepare
Training works when it matches your actual rules. Spend 30 minutes getting these ready, ideally with whoever owns IT and the person responsible for data protection.
Prepare
Check the training settings first; our guide to AI training opt-out settings shows where they are in each major tool.
The 45-minute AI training session plan
Keep it practical. Fewer slides, more examples from your own work.
| Time | Topic | What to cover | Activity |
|---|---|---|---|
| 0 to 5 min | Why we are doing this | AI is allowed and useful here; these rules protect customers and staff | Ask who used AI at work this week |
| 5 to 10 min | Approved tools | Which tools and accounts; no personal accounts, unofficial apps or extensions | Show the approved list |
| 10 to 20 min | What never goes into a prompt | Personal data, NRIC, financials, client documents, passwords | Sort ten examples into OK / not OK |
| 20 to 30 min | Checking the output | Hallucinations, made-up sources, outdated facts, tone | Find the errors in a real AI answer |
| 30 to 35 min | Telling customers | When AI-drafted work needs review or disclosure | Agree the house rule |
| 35 to 40 min | Reporting mistakes | Who to tell, how fast, no blame | Walk through one example |
| 40 to 45 min | Wrap-up | Questions; hand out and sign the checklist | Sign-off |
Part 1: approved tools and accounts
The CSA-IMDA joint advisory AD-2026-008 tells users to choose reputable tools with clear privacy policies and terms, and otherwise to assume the provider can access what you share and may use it to train models. It specifically warns that unofficial mobile apps and browser extensions offering AI features may also collect your data.
Turn that into two rules for staff. Use only the tools on the approved list, signed in with the company account. And do not install AI browser extensions or apps without asking, especially ones requesting permissions that do not match their job, such as camera or location access.
Part 2: what never to paste into an AI tool
The advisory says to avoid disclosing personal details such as full name, address and NRIC number, as well as financial or medical records, and that workplace information and confidential business data should never go into AI tools without proper authorisation.
Generally fine in an approved tool
- Rewriting your own draft email with names removed
- Summarising a public article or government page
- Brainstorming headlines or social posts
- Explaining a formula or an Excel error
Never, unless specifically authorised
- Customer names with phone numbers, NRIC or addresses
- Patient, student or employee records
- Contracts, bank statements, payroll or pricing for a tender
- Passwords, API keys or screenshots of admin pages
For the sorting exercise, write ten short items on cards (a customer complaint with a phone number, a public price list, a supplier contract, and so on) and have the group sort them. The discussion is where people learn the grey areas.
Part 3: checking AI output before you use it
The advisory warns that generative AI produces 'hallucinations': fictional or partly correct information presented convincingly, including fabricated references, quotes and statistics. It advises cross-checking with official and trusted sources and treating AI content as a starting point, not the final word.
Check every fact, number and name
Open the source. If the AI cannot point to one, treat the fact as unverified.
Check dates and rules against official sites
Prices, laws and grant terms change. Use gov.sg, IRAS, PDPC or the vendor's own page.
Read it as the customer would
Is the tone right? Does it promise anything we cannot deliver?
Keep a human name on it
The person who sends it owns it, whether AI drafted it or not.
For the exercise, show an AI answer about your own industry that contains a wrong fee, an outdated rule or a made-up reference, and ask the group to find the errors. Our guide to AI hallucinations as a business risk has more examples.
Parts 4 and 5: telling customers, and reporting mistakes
When to tell customers
Agree a simple house rule. A workable version: AI-assisted drafting of routine emails needs no disclosure, but anything presented as professional advice, a report or a design must be reviewed by a qualified person, and AI-generated images or chatbots should be labelled as such. If an AI tool processes customers' personal data, the PDPC's Advisory Guidelines on Use of Personal Data in Generative AI (issued 20 July 2026) treat your business as a deployer with primary responsibility; our summary of the guidelines explains what that means.
Reporting mistakes
People hide mistakes they think will get them punished. Say clearly that reporting fast is what you want. If someone pastes personal data into the wrong tool, the business may need to assess whether it is a data breach under the PDPA, and the PDPC expects that assessment to be done quickly.
Do it yourself or get help?
A manager can run this session with the plan above and the CSA-IMDA advisory summary as background reading. Repeat it for new joiners and once a year, or when you approve a new AI tool.
If you want ongoing training without running sessions yourself, our Security Awareness Training (S$390) sets up a platform in your own account with a 12-month plan of 5-minute lessons, Singapore-specific scam topics, completion tracking and a recorded walkthrough for new joiners. You can add your AI rules to that programme so they are repeated through the year.
Frequently asked questions
For most small businesses, one 45-minute session covers the essentials: approved tools, what never to paste, checking output, telling customers and reporting mistakes. Repeat it for new joiners and refresh it once a year, or whenever you approve a new AI tool or change your rules.
Personal details such as full names with NRIC numbers, addresses or phone numbers, financial and medical records, client documents, contracts, passwords and confidential business data, unless the business has specifically authorised it. The CSA-IMDA advisory says workplace and confidential information should never be shared without proper authorisation.
There is no mandatory curriculum, but two official documents are useful. The CSA-IMDA joint advisory of 28 July 2026 sets out safe-use practices for individuals, and the PDPC's Advisory Guidelines on Use of Personal Data in Generative AI, issued 20 July 2026, explain organisations' PDPA duties when using AI tools.
Report it straight away to the named person, saying which tool, what data and when. Delete the conversation if the tool allows it. The business then decides whether this is a data breach under the PDPA that needs assessing and, if notifiable, reporting to the PDPC within the required time.
Set a house rule. Routine drafting with AI usually needs no disclosure, but chatbots and AI-generated images should be labelled, and anything presented as professional advice should be reviewed by a qualified person. If AI processes customers' personal data, explain this in your privacy notice and follow the PDPC guidelines.