When I do one-to-ones with a client's staff, or with other business contacts, and we open their AI settings, about six in ten have model training switched on.
They aren't careless people. Most of them don't know the setting exists. In a 2026 survey of 1,944 American adults by DuckDuckGo, which sells a private AI chat product of its own, 53% didn't know, or weren't sure, that most chatbots train on their conversations by default. My own count is closer to home, and it's the one that worries me.
So here's the uncomfortable starting point for any founder. Your staff are already using AI at work. Some of them are using it on personal accounts, on their phones, with the training switch on. And what they paste in isn't just the financials. It's customer lists, payroll, a client's contract, an NRIC number, the password to the bank portal.
Banning AI doesn't fix this. It just moves it out of sight. This playbook is what I recommend instead: know what counts as sensitive, control the account, and put one page of rules in front of everyone.
This is practical guidance, not legal advice. For a specific situation, speak to a lawyer who knows the Personal Data Protection Act.
What Counts as Sensitive
When people hear "data privacy", they think of customer records. That's one class out of six.
- Personal data about customers. Names with phone numbers, emails, addresses, NRIC numbers, purchase histories. Anything that identifies a person.
- Employee data. Payroll, salaries, CPF, medical certificates, appraisals, disciplinary notes.
- Financials. Unpublished results, management accounts, forecasts, bank statements, account numbers.
- Client information under NDA. A client's contract, their pricing, their strategy deck. Often you've signed a promise not to share it with any third party, and an AI provider is a third party.
- Credentials. Passwords, API keys, one-time codes, the login to the accounting system. These should never go into any AI tool, on any plan.
- Trade secrets. Your pricing models, supplier terms, recipes, formulas, source code, the bid you're about to submit.
A useful test for staff: if this landed in a competitor's inbox, or a newspaper, would it hurt someone? If yes, it's sensitive.
The Risk Is the Account, Not the AI
Every major assistant has two worlds: a consumer plan for individuals, and a business plan sold to companies. The model is often the same. The rules on your data are not.
Three things follow from that table.
First, the switch only covers the future. Turning training off stops new chats being used. It doesn't pull back what was already shared.
Second, a personal account is outside your control. You can't see what was pasted, you can't set the rules, and when the employee leaves, the chat history leaves with them.
Third, the fix is cheap. A business seat costs far less than one breach notification. I recommend every client starts on, at minimum, a Team plan or its equivalent, so training is off for everyone, stays off, and isn't left to each person to remember.
What Singapore Expects
The Personal Data Protection Act doesn't stop applying because the tool is new. If staff paste customer or employee data into an AI tool, the usual obligations still bite: use personal data only for the purpose it was collected for, protect it with reasonable security, and take care when it leaves Singapore, which it almost always does with an AI provider.
On 20 July 2026 the PDPC published its Advisory Guidelines on Use of Personal Data in Generative AI. They're advisory, not law, but they show what the regulator expects. Most SMEs fall into the role the guidelines call a deployer: you're not building the model, you're putting it to work in your business, and you're answerable for how your people use it.
Two numbers are worth knowing. A data breach has to be reported to the PDPC within three calendar days of assessing that it's notifiable, which applies if it's likely to cause significant harm or affects 500 or more people. And the PDPC can fine an organisation up to S$1 million or 10% of its Singapore turnover, whichever is higher.
Separately from the law, there's the contract. If a client's NDA says their information goes to no third party, pasting it into a personal chatbot may breach that promise, whatever the privacy settings say.
A Traffic Light for Staff
Rules people remember are short. I use three colours.
- Green: any approved tool. Public information, your own drafting, general questions, and data that's been properly anonymised so no person or client can be identified.
- Amber: only the company's business account. Internal documents, management accounts, forecasts, customer analysis with names removed, anything you'd mark "internal".
- Red: never, on any plan. Passwords and access codes, NRIC numbers, medical information, and a client's confidential material unless the client has agreed in writing.
When in doubt, it's amber. When it's a password, it's red, always.
Rolling It Out in a Week
- Day 1: pick the tool and buy the seats. One approved assistant on a business plan, for everyone who uses AI at work. Cheaper than one breach notification.
- Day 2: adapt the policy. Fill in the tool and the owner in Figure 02. Add anything specific to your industry.
- Day 3: name an owner. One person who approves new tools and takes the "something went in" call. In a small company, often the founder or the finance lead.
- Day 4: brief everyone. Thirty minutes. Show the traffic light, show them where the training setting is on their phones, and let them ask the awkward questions.
- Day 5: check the personal accounts. Ask staff to turn training off on any personal AI app they keep. Most will, once someone shows them where the switch is.
- Every quarter: review. Tools and terms change. The table in Figure 01 will be out of date within a year.
Where This Breaks
- A policy with no tool. If the company doesn't provide an approved assistant, staff will use their own, and the policy becomes a document nobody follows.
- A ban. Same result, with less honesty. People don't stop using AI. They stop telling you.
- A founder who exempts themselves. If the boss pastes the board pack into a personal app, that's the real policy, whatever the page on the wall says.
- Meeting note-takers and extensions. The quiet leak. A free browser extension or a note-taking bot can see more than any chat window. Rule 7 exists for this.
- Treating anonymisation as magic. Removing names helps. But a customer list sorted by postcode, with order values and dates, can still identify people. When in doubt, it's amber.
The Switch Most People Never Find
The most useful five minutes in this whole playbook is the one-to-one where you ask someone to open the app and show you the training setting. Six times in ten, in my experience, it's on, and the person had no idea.
Control the account. Keep the red list out. Put one page in front of everyone. Then let your team use AI properly, because the answer was never to stop them.
Checked against each provider's published terms, November 2026.