TL;DR: It depends on two things: which plan the person is using, and what kind of data goes in. A company account on a business plan is a reasonable place for most internal work. Customer personal data in a free personal account is the one thing to stop. Under Singapore's PDPA, your company stays responsible for personal data your staff type into a tool. A one-page rule sheet for staff is at the end.
Most companies we talk to have the same quiet worry. Staff are already using ChatGPT, Claude or Copilot to draft emails, tidy spreadsheets and summarise documents. Nobody has said what is allowed to go in. The options are to ban it, ignore it, or set clear rules. A ban rarely stops anyone. Staff just use the tools on their own phones, and then you have no idea what is going in. Setting rules is the option that works, and the rules only need to answer three questions.
Question 1: what kind of data is it?
Sort what staff might paste into three piles. Most of the confusion goes away once everyone uses the same piles.
| Pile | Examples | In a company AI account |
|---|---|---|
| Public | Your website text, published prices, a job advert | Fine anywhere |
| Internal | Draft proposals, process notes, meeting notes, internal numbers | Fine in a company account on a business plan |
| Personal or confidential | Customer names with contact details or account numbers, staff records, NRIC numbers, contracts under confidentiality, anything a client told you in confidence | Only if IT has approved that tool for that use. Otherwise, leave it out or remove the names first |
In short: public material can go in anywhere, internal material belongs in a company account, and personal or confidential data needs approval first.
Question 2: which plan is your team using?
The promises depend on the plan, not just the brand. A free personal account and a company account on a business plan are different products, with different terms about training, storage and location. This is the check most companies skip.
Here is what three of the major business offerings say in their own documentation:
- OpenAI (ChatGPT Business, Enterprise and the API). OpenAI says that by default it does "not use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or our API platform, including inputs or outputs, for training or improving our models". Its data-residency options, which let a company choose where stored data is kept, include Singapore, but OpenAI limits them to ChatGPT Enterprise, Edu, Healthcare and API customers. ChatGPT Business is not on that list. Source: OpenAI.
- Claude (Anthropic). Anthropic says that "by default, we will not use your inputs or outputs from our commercial products" to train its models, with an exception for feedback a user chooses to submit, such as the thumbs up or down button. Feedback conversations can be kept for up to 5 years. Owners of Team and Enterprise plans can switch the feedback option off for the whole organisation. Source: Anthropic.
- Microsoft Copilot (Microsoft 365). Microsoft's documentation says prompts, responses and the company data Copilot reads "aren't used to train foundation LLMs" (the large AI models behind the tool). The conversations are stored alongside your other Microsoft 365 content, and administrators can set how long they are kept. Copilot only shows a person documents they already have permission to open. On location, Microsoft says requests are sent to the nearest data centre in the region, but can go to other regions when demand is high, and that customers outside the EU may have requests processed in the US, EU or other regions. Source: Microsoft Learn.
These promises apply to the business plans only. Free and personal versions have different terms, and some use your chats to improve the AI unless the user turns that off. Terms also change often, so if your staff use a free version, check its current terms rather than assuming it works like the business plan.
Where your data is kept is a separate question. OpenAI lets companies keep stored data in Singapore, but only on its Enterprise, Edu, Healthcare and API plans, not ChatGPT Business. Microsoft says Copilot requests from outside the EU can be processed in the US, EU or other regions. If a client contract or an industry rule requires your data to stay in Singapore, get the vendor to confirm it in writing.
Question 3: what does the PDPA say?
Using someone else's AI tool doesn't move the responsibility to them. If your staff enter customer personal data and something goes wrong, the PDPA treats your company as the one responsible for protecting it.
- You are responsible for the tools you choose. Singapore's PDPC published its Advisory Guidelines on the Use of Personal Data in Generative AI on 20 July 2026. They are guidance, not new law. A summary by the law firm Allen & Gledhill says companies that choose and use a generative AI tool "bear primary responsibility" for making sure it meets their PDPA obligations, and are encouraged to write down their safeguards and review them regularly.
- You must protect personal data you hold, and that includes data your staff put into an AI tool. A written rule and a named person in charge are how you show you are doing it.
- Personal data sent overseas must get the same level of protection. The PDPA requires the organisation receiving it to protect it to a standard comparable to Singapore's. Many AI tools process data outside Singapore, so this rule covers more tools than people expect. Keep personal data on business plans, which come with a data-protection agreement between the vendor and your company. A free personal account has no such agreement with your company, which is another reason to keep personal data out of it.
- Breaches have a deadline. The PDPC's rule is that a breach that must be reported has to reach the PDPC no later than 3 calendar days after you decide that it is reportable. In general, that means a breach likely to cause significant harm, or one that affects 500 or more people. Three days goes quickly, so decide now who makes that call. The rule sheet below has a space for their name.
AI tools aren't off limits. Treat a chat window with the same care you give a customer spreadsheet.
Your options
There are three levels of control. Most small companies should pick the middle one.
- Light. Give staff the rule sheet below and ask them to follow it. It costs nothing and is better than silence, but you are relying on everyone remembering it.
- Middle. Pay for one business AI plan, usually the one that comes with the email and documents you already use. Ask staff to use only that for work, and give them the rule sheet. Everyone is then on a plan whose terms you have checked, and IT can manage the settings.
- Strict. Allow AI only for public and internal data, never personal data, and have IT block other tools. This suits companies handling health, financial or children's data, or ones whose clients have asked for it in writing.
Whichever you pick, find out which AI tools your staff actually use, this week if you haven't already. The answer is often a surprise. Our checklist for IT has a tool inventory you can use for it.
The one-page rule sheet for staff
Download the rule sheet (it opens in Word or Google Docs), fill in the square brackets and send it to everyone who uses AI at work. The text is below if you'd rather copy it.
The sheet asks you to name an AI contact: one person staff go to before putting in sensitive data, with questions, or when something goes in by mistake. They don't need to be technical. They need to be easy to reach and able to make a decision quickly, including whether a mistake has to be reported to the PDPC. In a small company this is usually the owner, the operations manager or whoever looks after IT. The PDPA requires every organisation to appoint a Data Protection Officer, so if you have one, they are the natural choice.
Using AI tools at work: [Company name]
Your AI contact: [name, job title, email]. This is the person who approves AI use, answers questions and handles mistakes.
Use only our approved tools, signed in with your work account. Approved: [tool names]. Don't use a personal account, a personal phone app or a free website for work material.
Fine to put in:
- Public material, such as our website text or published prices.
- Internal drafts and notes that contain no customer or staff personal details.
Ask your AI contact first:
- Anything with customer or staff names plus contact details, account numbers or other identifiers.
- Contracts, quotes and anything a client gave us in confidence.
Never put in:
- NRIC or passport numbers, bank or card details, passwords or access keys.
- Health records or anything about children.
Check before it leaves. A person reads and checks anything that goes to a customer, a regulator or a bank, or is saved as the official record. Check numbers, dates, names and quotes against the original.
If something goes in by mistake, tell your AI contact straight away, on the same day. Nobody is in trouble for reporting. The company has to decide within days whether the law requires it to report the mistake, and cannot do that if it hears late.
Questions or new tools: ask your AI contact.
What we don't know
- Vendor terms change. We used OpenAI's, Anthropic's and Microsoft's own documentation, but they update it regularly. Confirm the terms for your own plan. We could read OpenAI's page only as a search extract, not in full.
- We haven't covered every tool. This post doesn't cover the free or personal versions of any tool, or Google Gemini and others. Ask IT to check any other tool your staff use.
- We haven't read the PDPC's full guidelines. We relied on a law firm's summary, which doesn't cover staff typing data into a tool or how overseas processing is handled. The breach-reporting thresholds above are the general rule, not the full detail.
- This is not legal advice. If you have a real incident, speak to a lawyer.
- No setting makes an AI tool safe. The rule sheet and the checks reduce the risk. They don't remove it.
Related reading: How to Prevent AI Tools From Exposing Company Data covers tools that can share files on their own, and Why Dropping Files Into ChatGPT Isn't a Business Process covers why a chat window isn't a substitute for a built process.