Who Is Liable When Staff Paste Customer Data Into a Public AI Chatbot?

If an employee pastes a customer list, a signed contract or a client email thread into a free AI chatbot, your company is the one answerable under Singapore's Personal Data Protection Act. Not the employee. The PDPA places its obligations on the organisation that controls the data, and an employee acting in the course of employment is treated as acting for the organisation. "A staff member did it without telling us" is not a defence. It is evidence that no control was ever put in place. That is the whole exposure in one sentence, and it is fixable in an afternoon.

What does the PDPA actually require when data goes into an AI tool?

Three obligations do most of the work here, and none of them mention artificial intelligence.

The purpose limitation obligation says you may only use personal data for purposes the individual was told about and consented to, or that a reasonable person would consider appropriate. Your customer gave you their address so you could deliver an order. Running their file through a chatbot to draft a marketing email is a different purpose.

The protection obligation requires reasonable security arrangements to prevent unauthorised access or disclosure. Typing data into a consumer service with no business agreement behind it is difficult to defend as reasonable, particularly once a cheaper, contractually covered option exists in the market.

The transfer limitation obligation is the one most SMEs miss. Almost every public AI chatbot processes data on servers outside Singapore. Sending personal data overseas requires you to satisfy yourself that it will be protected to a standard comparable to the PDPA, normally through contract terms. A free consumer account gives you no such contract, which means the transfer is happening with nothing underneath it.

Penalties were raised in 2022. An organisation with Singapore turnover above S$10 million can face up to 10 per cent of that turnover; below that threshold the ceiling is S$1 million. For most SMEs the realistic damage is smaller than the headline but still unpleasant: a PDPC investigation consumes weeks of management time, and the finding becomes a public document that your enterprise customers will read.

Which AI tools are a real problem and which are fine?

This is the distinction that saves money, because the honest answer is that banning AI is both unenforceable and unnecessary.

Consumer tiers are the problem. Free and personal-subscription chatbots generally reserve the right to use your conversations to improve their models, the account belongs to the individual rather than the company, and you cannot audit or delete what went in. When an employee leaves, that history leaves with them.

Business and enterprise tiers are a different proposition. Paid workspace plans from the major vendors contractually exclude customer content from model training, sit under administrative control, and come with data processing terms you can show a client. The tool is the same. The paperwork is not.

So the practical control is rarely a ban. It is a licence, plus a short written rule about what may go into it.

What should you write down before January?

One page, signed by whoever owns the business, is enough to move you from no position to a defensible one. It needs four things.

• The approved tools, by name. "Use the company Copilot account" is a rule. "Be careful with AI" is not.
• A red list. Full names with contact details, NRIC or FIN numbers, bank and payment details, medical or employment records, signed client contracts, and anything a client has marked confidential. These never go into any AI tool without written approval.
• A green list. Public marketing copy, your own product descriptions, anonymised or aggregated figures, generic drafting help. Say what staff may do, or they will guess.
• One named owner. A person, not a department, who approves exceptions and keeps the list current.

Then brief the team once, in a meeting, and keep the attendance note. The note matters: it is the difference between an isolated staff error and a systemic failure to train, and the PDPC treats those two very differently.

What do you do if it has already happened?

Assume it has. Most businesses that go looking find at least one instance.

Work out what went in, whose data it was, and how many people were affected. Delete the conversation history and, where the tool allows it, turn off training on your inputs. Then assess whether the incident is notifiable. Under the data breach notification obligation, you must notify the PDPC if the breach is likely to cause significant harm to affected individuals, or if it affects 500 or more individuals. Assessment is expected to be prompt, with 30 days treated as the outer limit, and notification to the PDPC must follow within three calendar days of concluding it is notifiable. Where significant harm is likely, affected individuals must be told as well.

Most single-employee pasting incidents will not clear that bar. Document the assessment anyway. An undocumented decision not to notify looks identical to never having noticed.

Why is October the right month to deal with this?

Because the money decision and the compliance decision are the same decision, and both land now.

Software renewals for 2027 are being quoted this quarter, and both Microsoft and Google have been bundling AI features into higher-priced tiers. If you are going to be asked for more per seat regardless, the time to decide which staff get a licensed, contractually covered AI account is while the quote is still open, not in February after the invoice has been accepted. A licensed seat typically runs in the region of S$30 to S$45 per user per month. Covering your five or six heaviest users is a four-figure annual line item set against a liability ceiling of S$1 million.

There is a second reason. Larger customers have started asking about AI usage in vendor questionnaires alongside their cyber insurance and Cyber Essentials questions. A one-page AI policy with a date on it is now a sales document, not just a compliance one. Businesses that can produce it on request win procurement rounds against businesses that need two weeks to write one.

Frequently asked questions

Can we just ban AI tools at the office?

You can write the rule, but you cannot enforce it. Staff have the same tools on their personal phones, and a ban simply moves the activity somewhere you cannot see, audit or delete. A named approved tool with a clear red list produces far better compliance than a prohibition nobody follows.

Is an AI meeting note-taker covered by the same rules?

Yes, and more strictly. A recording captures voices, names and commercial discussion, which makes it personal data belonging to everyone in the room, including people outside your company. You need consent before recording starts, a clear statement of where the transcript is stored and for how long, and a check that your client's own confidentiality terms permit it. Several professional services contracts do not.

Who should own the AI policy in a company with no IT department?

Whoever already owns customer data decisions, usually the owner, the operations lead or the finance manager. It does not need technical skill. It needs the authority to say no to an exception and the discipline to keep a short written record of what was approved and when.

If you want this done properly rather than quickly, Digital Perpetual runs a short AI exposure review for Singapore SMEs: what staff are already using, what has already gone out, the one-page policy, and the licence decision costed against your 2027 renewal. It takes about a week and it closes before the year-end rush.