Adding an AI chatbot to your WhatsApp or Telegram line is one of the easier wins a Singapore SME can make this year — faster replies, fewer missed enquiries, a business that answers at 11pm. But the moment that assistant reads a customer's name, mobile number or booking details, you are handling personal data, and the Personal Data Protection Act (PDPA) applies. That is not a reason to panic. It is a reason to set the thing up properly.
Most of what the PDPA asks of a small business is common sense written down: tell people what you collect and why, collect only what you need, keep it safe, and do not hoard it. This is a practical checklist for owners wiring an AI assistant into a channel where customers already message them — a salon, clinic, tuition centre or agency, not a compliance department with a legal team.
Does the PDPA even apply to your chatbot? Short answer: yes
If your assistant collects or stores data about an identifiable person — a name, a mobile number, an email, an NRIC, an appointment tied to a customer — that is personal data, and the PDPA governs it. The regulator is the Personal Data Protection Commission (PDPC). A chatbot gets no exemption because it is automated: you, the organisation, remain responsible for what it collects and what your vendor does with it. Picture a PDPA-compliant chatbot as a member of staff who happens to never forget a conversation.
Consent and purpose limitation
Two obligations do the heavy lifting. Consent means you should have a person's agreement to collect and use their data. Purpose limitation means you only use it for the purposes you told them about. A customer who messages you first to ask about a service is generally taken to be consenting to you using that message to answer them — but that does not stretch to adding them to a marketing blast next month, or passing their number to a partner. To do more, say so and let them opt in.
Practically: make sure a customer can see a short, plain notice — a line in your WhatsApp greeting, or a link on your booking page — explaining what you collect, why, and who to contact about it. You do not need a ten-page policy; you need one honest paragraph and a real answer when someone asks.
Collect only what you actually need
The cheapest way to reduce risk under the PDPA is to hold less customer data. Every field your bot asks for is something you must protect, justify and eventually delete. Before you let an assistant collect an NRIC, a date of birth or a full address, ask whether the job requires it. To book a haircut you need a name, a number and a time — not an identity card. Ask for the minimum, and resist capturing everything just in case.
Where is the data stored and processed?
When an AI assistant answers a message, the text usually travels to a language model to generate the reply. Ask your vendor where that happens and on what terms. For an SME weighing up PDPA and AI in Singapore, three questions matter: is customer data processed in-region or shipped overseas, is it encrypted, and is there a written agreement that the provider will not keep or train on it? Your vendor is your data intermediary under the PDPA, so their handling is effectively yours in the eyes of the PDPC. Clinics and anyone touching health information should be especially careful — see our notes for clinics.
Sensitive data: NRIC, card numbers, and redaction
The PDPA has no formal sensitive-data tier the way some laws do, but the PDPC treats certain data with extra care, and NRIC is the clearest example. Its guidelines say organisations should generally not collect, use or disclose NRIC numbers — or copies of the card — unless the law requires it or it is genuinely necessary to verify identity to a high degree of accuracy. Booking a table or a property viewing is not that. The same caution applies to full credit-card numbers, which you rarely have a good reason to keep in a chat log.
This is where redaction earns its keep. A well-built assistant strips things like NRIC and card numbers out of a message before it is stored or sent onward, so a stray number a customer volunteers does not sit in your database. microcrew does this at ingestion — sensitive identifiers are redacted before the message is saved or passed to a model — so an over-sharing customer does not create a liability you never asked for.
The safest customer data is the data you never collected. The second safest is the data your tool redacts before it is ever stored.
Retention: don't keep it forever
The Retention Limitation Obligation is simple: stop keeping personal data once the purpose is done and there is no legal reason to hold it. Decide how long chat transcripts and customer details live in your systems, write it down, and make sure your tool enforces it rather than quietly accumulating years of conversations by default. "We keep everything indefinitely" is not a policy — it is a breach waiting to be discovered.
Don't train models on your customers' chats
One quiet risk with AI tools is that customer conversations get fed back to improve a shared model. For an SME that is a bad trade: you lose control of where the data ends up, and cannot honestly promise a customer their booking chat will not resurface elsewhere. Ask your vendor whether your conversations train any model — theirs or a third party's — and get the answer in writing. A no-training arrangement should be the default, not a paid upgrade.
A practical PDPA checklist for your chatbot
Do one pass before going live — each item is something you can check yourself in an afternoon.
- Publish a plain-language notice of what you collect, why, and who to contact about it — one paragraph in your greeting or on your booking page.
- Collect only the fields the task needs. Drop any request for NRIC, date of birth or address the job does not genuinely require.
- Confirm where data is processed — in-region or overseas, encrypted in transit and at rest, and who your vendor's sub-processors are.
- Make sure NRIC and card numbers are redacted. Test it: send a fake NRIC and card number through and check they are masked before storage.
- Set a retention period and verify it is enforced — how long transcripts and contact details are kept, and that the tool deletes on schedule.
- Get no-training in writing — confirm your conversations are not used to train the vendor's or a third party's model.
- Keep a human handover for anything sensitive — complaints, medical or legal questions, and payment disputes should reach a person.
- Know your breach plan. Notifiable data breaches must be reported to the PDPC and, in some cases, to affected individuals — know who does that, and how fast.
Getting it right without overthinking it
None of this should stop you from adding an assistant to WhatsApp. The PDPA is not trying to catch out the salon owner who wants faster replies — it is asking you to be deliberate about data you were already collecting by hand. Handle the basics (notice, minimisation, redaction, retention, no training) and you are ahead of most businesses messaging customers with no policy at all. Our other guides cover the practical side of running AI on WhatsApp.
If you would rather these defaults were handled for you, microcrew is built for Singapore SMEs with them in place — NRIC and card numbers redacted at ingestion, AI processing under a no-training agreement, and bounded retention — with 50% off the first month. See how it works or message the team on WhatsApp at +65 8466 3236 to talk it through.
This article is general information for Singapore SMEs, not legal advice. For how the PDPA applies to your specific business, check the PDPC's guidelines or speak to a qualified professional.