Putting company data into a public cloud AI chat tool can breach Singapore’s PDPA when that data identifies people and you have not secured consent or controlled where it goes. The paid enterprise tiers reduce the risk. Running a private model on your own infrastructure removes the question entirely.

The question every Singapore team is quietly asking

Someone on your team has already pasted a customer list, a contract, or a support transcript into a cloud AI chat tool to save an hour. It worked, so they did it again. Now nobody is quite sure how much company information has left the building, or where it went.

This is the honest starting point for most teams we meet. The tool is useful, the staff are not being reckless, and the policy was never written down. The worry only surfaces when a client asks how their data is handled, or when someone reads the words “Personal Data Protection Act” on a renewal form.

So the real question is not whether these tools are clever. It is whether the way your team uses them would survive a question from a customer or the PDPC.

What the PDPA actually requires

Singapore’s Personal Data Protection Act governs how organisations collect, use, and disclose personal data. Personal data means anything that can identify a living individual, on its own or combined with other information you hold. A name, an email, an NRIC, a phone number, or a support ticket tied to a customer all count.

Three obligations matter most here. You need consent or a valid basis to use that data. You need to protect it with reasonable security. And you remain accountable for it even after it leaves your hands, including when a third party processes it for you.

In March 2024 the PDPC issued specific guidance on using personal data in AI systems. The direction is clear: feeding customer data into a tool you do not control, without a basis and without knowing where it is stored, is the kind of use that creates exposure. Breaches can lead to financial penalties tied to your annual turnover and to the loss of client trust that is harder to repair.

What happens to your data when you paste it into a cloud AI tool

When you use the free or personal version of a public cloud AI chat tool, your conversations can be retained and may be used to improve the model unless you turn that setting off. The data leaves Singapore and sits on servers run by a company you have no contract with. For personal data, that is the part that conflicts with the PDPA, because you can no longer say where the data is or who has seen it.

The paid business tiers are different, and it is fair to say so. The enterprise plans from the major providers operate under terms that exclude your inputs from training and add contractual protections. For many companies that is a reasonable step up.

What those tiers still do not change is location and control. Your data is processed outside your premises by a third party. You are trusting a contract rather than holding the data yourself, and you are dependent on a subscription that can change in price or in terms.

How to use AI without handing over your data

The way to remove the question is to run the model where the data already lives. A private AI setup keeps the model on your own hardware, inside your own network, so customer data never leaves your control. There is no third party to vet, no overseas server to explain, and no subscription that resets every month.

Open models have made this practical. Mistral, Qwen, and Gemma now run capably on a single workstation-class GPU, and they handle the everyday work most teams want from AI: drafting, summarising, answering questions from your own documents, and processing support requests. The output is good enough for real workflows, and the data stays in the building.

Most teams stall here because they assume private AI needs a server room and a data science team. It does not. A single machine handles a small team, and the setup is built around the tools your staff already use.

What it looks like in practice

A typical first step is narrow on purpose. We pick one workflow that currently leaks data into public tools, a support inbox, a document review process, or an internal knowledge search, and move it onto a private model.

Staff get the same chat experience they are used to, pointed at a system that runs locally. The customer data they work with stays on your infrastructure, and you can finally answer the question of where it goes with one sentence: it does not leave.

From there, teams usually expand to a second and third use case once they trust the first. The point is to start with the workflow that carries the most risk today, not to boil the ocean.

Where the public tools are still fine

Private AI is not the answer to everything, and pretending otherwise would not help you. For brainstorming, drafting marketing copy, writing code that touches no customer records, or summarising public information, the public tools are quick and perfectly reasonable.

The line is personal and confidential data. If a task involves information that identifies your customers, your staff, or your commercial position, that is the work to keep in-house, on a private and cloud-free setup. Draw the line there, write it into a short policy, and most of the PDPA worry goes away.

MT Labs helps companies across Singapore deploy AI tools they actually own. Private infrastructure, no recurring cloud subscriptions, and a setup built around how your team already works. Whether you’re exploring your first AI use case or consolidating scattered tools into one system, we’ll walk you through it. Get in touch and let’s figure out what makes sense for your business.