Last updated 23 October 2026 — by Adrian Tan, SDM. Marketing and website guidance, not legal advice. References are to the Personal Data Protection Act 2012 and the Consumer Protection (Fair Trading) Act 2003 as shown on Singapore Statutes Online on 23 October 2026, and to PDPC, IMDA and MAS guidance published up to that date. Advisory guidelines are not law; where we interpret how a rule applies, we say so.
An aesthetics clinic adds a WhatsApp bot that books consultations. A renovation firm’s website widget gives instant quotes. An online store’s assistant answers questions about delivery and returns at 2am. A tuition centre’s bot collects parents’ numbers for trial-class follow-ups.
Each of these is cheap to set up and genuinely useful. Each is also a system that collects personal data, makes statements on the business’s behalf and, increasingly, generates its own answers. Singapore has no single statute that regulates AI chatbots; we found none in our research. But the existing rules apply in full, and in July 2026 two regulators published guidance aimed squarely at businesses that deploy them: the Personal Data Protection Commission’s (PDPC) advisory guidelines on generative AI, and IMDA’s transparency guidelines for generative AI chatbots.
This guide is for business owners and marketers adding a chatbot to a website, WhatsApp or Messenger, and for the agencies that build them. It covers the data protection rules, the Do Not Call trap, who is liable when the bot says something wrong, the new transparency guidance, sector overlays and a launch checklist, with four worked examples.
The short version
- You are responsible for your bot, even if a vendor built it. PDPC’s July 2026 guidelines say the business that deploys a generative AI system bears primary responsibility for PDPA compliance, including when it buys the system as a service.
- Tell people what you collect and why, before you collect it. Consent under the PDPA only counts if the purposes were notified first.
- Do not ask for NRIC numbers to book or quote. NRIC collection is limited to where the law requires it or high-fidelity identity verification is needed.
- A phone number in a chat is not consent to market. Do Not Call rules still apply to follow-up messages, including on WhatsApp.
- “The bot said it” is not a defence. Misleading statements by a chatbot are statements by the business.
- Disclose that it is a bot. We found no Singapore statute requiring it, but IMDA’s voluntary guidance recommends it, and the EU AI Act requires it for businesses serving EU users.
How common are AI chatbots in Singapore SMEs?
Adoption is rising fast from a low base. IMDA’s Singapore Digital Economy Report 2025 found that 14.5% of SMEs used AI in 2024, up from 4.2% in 2023, against 62.5% of larger firms. Customer service was one of the top three business functions where firms used AI, alongside IT and finance, and 84% of AI-adopting firms used off-the-shelf generative AI tools. That last figure matters for compliance: most SME chatbots are built on a vendor’s model and platform, which is exactly the arrangement PDPC’s new guidelines address.
The PDPA, stage by stage
The easiest way to apply the Personal Data Protection Act to a chatbot is to follow a conversation from greeting to follow-up.
Before the first question: notify and get consent
Section 13 of the PDPA says you may not collect, use or disclose personal data without consent, deemed consent or an exception. Section 20 requires you to tell people your purposes on or before collection, and section 14(1) says consent only counts if that notification was given. Section 14(2) adds that you cannot make consent a condition of service beyond what is reasonable, or obtain it with misleading information.
For a chatbot, that means a short notice at the start of the chat, before the bot asks for a name, number or email: what you will collect, what for, and a link to your privacy policy. “We’ll use your name and number to arrange your consultation and reply to your enquiry” is the kind of line that works. If you also want to use chat transcripts to improve the bot, say so. PDPC’s 2024 guidelines on AI recommendation and decision systems recognise a Business Improvement exception for improving products and understanding customer behaviour, but only where the purpose cannot reasonably be achieved without identifiable data.
During the chat: collect only what you need
The most common over-collection is the NRIC number. Since 1 September 2019, PDPC’s NRIC guidelines limit collecting NRIC numbers, or copies, to where the law requires it or where identity must be verified to a high degree of fidelity. The same applies to FIN, birth certificate and work permit numbers. Booking a consultation, sending a quote or reserving a trial class does not meet that bar. See our guide to NRIC rules in marketing for the detail.
PDPC’s July 2026 generative AI guidelines add a point specific to chatbots: user prompts, inputs, outputs and agent activity are data the business must protect, and businesses should tell users what personal data they should, and should not, enter. A line such as “Please don’t share medical records, NRIC or card numbers here” is a practical way to do that. One of PDPC’s worked examples is a chatbot whose prompts are scanned for personal identifiers and whose logs are access-restricted.
After the chat: protect, retain, transfer
- Protection (s 24): make reasonable security arrangements for chat logs, including who on your team and at your vendor can see them.
- Retention (s 25): delete transcripts once the purpose is served and there is no legal or business reason to keep them. Set a retention period in the chatbot platform, rather than keeping everything forever by default.
- Overseas transfer (s 26): many chatbot platforms and model providers process data outside Singapore. Transfers must meet the prescribed requirements so the data gets comparable protection, which in practice means checking your vendor contract.
If something goes wrong: breach notification
A notifiable data breach must be reported to PDPC no later than three calendar days after you assess it as notifiable (s 26D), and affected individuals must also be told. A breach is notifiable if it is likely to cause significant harm or is of significant scale, meaning 500 or more people. Under the breach regulations, some data types, such as card and bank account numbers, are deemed to cause significant harm. A chatbot that has collected thousands of names and numbers can reach that threshold quickly. Our guide to data breaches and marketing data covers the response steps.
Financial penalties for PDPA breaches go up to 10% of annual Singapore turnover for organisations with turnover above $10 million, or $1 million otherwise (s 48J(3)).
The Do Not Call trap
Many chatbots exist to capture leads, and the natural next step is a promotional WhatsApp or SMS. That is where Part 9 of the PDPA, the Do Not Call (DNC) provisions, comes in.
Section 43 requires a valid DNC Register check before you send a marketing message to a Singapore number, unless the person has given clear and unambiguous consent, evidenced in written or other accessible form (s 43(4)). A number typed into a chatbot to get a quote is consent to be contacted about that quote. It is not, by itself, clear and unambiguous consent to receive promotions. PDPC’s DNC guidance treats messages sent through data apps such as WhatsApp to a Singapore number as within scope.
The fix is simple: an unticked opt-in inside the chat (“Send me offers and updates on WhatsApp”), a record of the timestamp and wording, and a DNC check for anyone who did not tick it. DNC breaches carry penalties of up to $1 million for organisations (s 48J(4)). Our guides to WhatsApp marketing in Singapore and telemarketing and DNC rules cover the full regime.
When the bot gets it wrong: who is liable?
Consumer protection
Section 4 of the Consumer Protection (Fair Trading) Act 2003 makes it an unfair practice for a supplier to say or omit anything that might reasonably deceive or mislead a consumer, or to make a false claim. The Act’s definition of supplier includes employees and agents. Several items on the Second Schedule’s list of unfair practices map directly onto chatbot errors: misrepresenting price or availability, charging materially more than an estimate, and misrepresenting a consumer’s rights or remedies, such as refunds. In our reading, a business cannot escape those provisions because the misleading statement was generated by software it chose to deploy. We found no Singapore case yet on a chatbot misstatement under the Act.
The Air Canada case
The best-known illustration is from Canada. In Moffatt v Air Canada (2024 BCCRT 149, decided 14 February 2024), the airline’s website chatbot told a customer he could apply for a bereavement fare retrospectively, which the airline’s actual policy did not allow. The British Columbia Civil Resolution Tribunal found negligent misrepresentation and rejected the argument that the chatbot was a separate legal entity responsible for its own actions. The airline was responsible for all the information on its website, chatbot included. The award, about CA$812 including interest and fees according to published reports, was small, but the principle travels: it is not Singapore law, but it is how a tribunal reasoned about the same question.
Singapore’s algorithmic contract case
Singapore’s own leading case on automated systems is Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 02. The Court of Appeal held that, for deterministic algorithms, knowledge of a mistake is judged by the state of mind of the programmers at the time of programming, up to the point the contract was formed. The court also noted that the law may need to develop for systems that could be said to have “a mind of their own”. Generative chatbots are not deterministic, so how Quoine applies to them is open. The practical lesson is that the decisions you make when you configure a bot, such as what it may quote, what sources it draws on and when it hands over to a person, are likely to be where responsibility is assessed.
What that means in practice
- Ground the bot’s answers in your actual policies and price lists, not the model’s general knowledge.
- Stop it from giving firm prices, refund promises or legal, medical or financial advice; route those to a person.
- Label estimates as estimates, with what they include and exclude.
- Review transcripts regularly for wrong answers, and fix the source.
Refund promises are the commonest problem. If your bot tells customers they have a “30-day full refund” you do not offer, or that faulty items cannot be returned, it is misstating their rights. Our guide to “no refund, no exchange” and the lemon law explains what the correct position is.
The July 2026 guidance
PDPC: the deployer is responsible
PDPC’s Advisory Guidelines on Use of Personal Data in Generative AI, published on 20 July 2026, are the most directly relevant document for chatbot owners. Their central point (paragraph 9.1) is that system deployers bear primary responsibility for PDPA compliance, even when they obtain the system as software-as-a-service or through an API. The guidelines encourage written policies on what personal data may be entered, how prompts and outputs are protected, and the safeguards for agentic features that take actions, such as booking or ordering, on a user’s behalf. They are advisory, not binding, but they show how PDPC will read the Act.
IMDA: the Chatbot Info Card
IMDA’s Transparency Guidelines for Generative AI Chatbots, also dated 20 July 2026, are voluntary. They recommend that deployers publish a “Chatbot Info Card” answering four questions: what the bot can and cannot do, how reliable and safe it is, how data is handled, and how users can report issues. At first use, the chatbot should show a short safety statement linking to the card. The guidelines say they do not replace sector rules.
For an SME, an info card is a short page on your website, not a compliance project. A version for a renovation firm might say: “Our assistant answers questions about our services and gives indicative price ranges. It cannot give final quotations, which are issued only by our designers after a site visit. Conversations are stored for 90 days to handle your enquiry and are processed by our chatbot provider. Spot an error? Tell us at [email].”
Do you have to say it is a bot?
We found no Singapore statute that requires a chatbot to disclose it is automated. IMDA’s guidance recommends disclosure at first use, and it is good practice: a customer who thinks they are messaging a person is more likely to rely on what they are told. If you serve customers in the European Union, Article 50 of the EU AI Act requires that people are informed they are interacting with an AI system, unless that is obvious, by the time of the first interaction. That obligation applies from 2 August 2026.
Sector overlays
| Sector | Extra rules for a chatbot |
|---|---|
| Financial institutions | MAS finalised its Guidelines on AI Risk Management on 7 October 2026. They apply to all financial institutions and all forms of AI, and take effect from 7 October 2027. |
| Healthcare and aesthetics | The bot’s claims are advertising under the healthcare services advertising regulations. MOH has said licensees must not reproduce testimonials, so a bot should not quote patient reviews. Health details typed into chats raise the stakes of a breach. |
| Education and children | PDPC has separate advisory guidelines on children’s personal data (March 2024). Collect parents’ details, not children’s, wherever possible. |
| All consumer businesses | CPFTA s 4 and the Second Schedule apply to everything the bot says about price, availability and rights. |
For the related advertising rules, see our guides to aesthetic clinic advertising and children’s data in marketing.
Four worked examples
1. An aesthetics clinic’s WhatsApp booking bot
The bot asks for full name, NRIC and “any medical conditions” before offering slots. Remove the NRIC request: booking does not need high-fidelity identity verification. Move medical questions to the consultation itself, so health data is not sitting in chat logs. Add a first-message notice of purposes and a line telling users not to send medical documents in the chat. Make sure the bot’s answers about treatments follow the clinic’s advertising rules and never quote patient testimonials.
2. A renovation firm’s instant-quote widget
The bot tells a homeowner a kitchen will cost “$18,000 all-in”. The final bill is $26,000. That is the scenario the CPFTA’s estimate and price items cover, and the Air Canada reasoning suggests “the bot said it” will not help. Reconfigure the bot to give ranges, state what is excluded, and say final quotations are issued only after a site visit. Keep transcripts so you can see exactly what was said. Our guide to renovation contractor advertising covers the wider rules.
3. An online store’s support assistant
Asked about a faulty blender, the assistant replies that “all sales are final after 7 days”. That misstates the buyer’s lemon-law rights. Ground the bot in your returns policy, with a separate rule for faulty items, and test it with the ten questions customers actually ask. Add an escalation path to a person for any defect complaint.
4. A tuition centre’s trial-class bot
The bot collects the child’s name, school and level, plus a parent’s mobile number, then the centre adds every number to a weekly promotions broadcast. Collect the parent’s details and the level only; the school is rarely needed. Add an unticked opt-in for promotions and check the DNC Register for everyone else before broadcasting. Check where the chatbot vendor stores data and set a retention period for transcripts.
A launch checklist
- Map what personal data the bot collects, and remove anything you do not need, starting with NRIC.
- Show a purpose notice and privacy policy link before the first data request.
- Tell users what not to enter (NRIC, medical records, card numbers).
- Ground answers in your own policies and price lists; block firm quotes, refund promises and professional advice.
- Add a clear hand-over to a person.
- Disclose it is a bot, and publish a short Chatbot Info Card.
- Use a separate, unticked opt-in for marketing messages, and keep the record.
- Check your vendor contract for overseas processing, security and data use for model training.
- Set a transcript retention period and restrict log access.
- Add the bot to your breach response plan, with the three-day notification clock.
- Review a sample of transcripts every month.
Your chatbot is part of your website, so the general website obligations still apply; see our guide to legal requirements for business websites in Singapore, and to PDPA rules for marketing and tracking. Avoid designs that pressure users into sharing data; our guide to dark patterns explains why. For using AI to produce marketing content rather than conversations, see AI in content marketing in Singapore.
Where this leaves you
An AI chatbot is not legally different from a salesperson who never sleeps: what it collects is your data responsibility, and what it says is your statement. The July 2026 guidance from PDPC and IMDA makes that explicit. Collect less, notify first, ground the answers, separate marketing consent, disclose that it is a bot, and keep the logs safe and short-lived. Done that way, a chatbot is one of the most useful tools an SME website can have.
If you want a website, with or without a chatbot, built to convert and to handle customer data properly, our web design team in Singapore can help, and you can see how we work in our client case studies. For the bigger picture, start with our complete guide to web design in Singapore.
Frequently asked questions
Is there a law that regulates AI chatbots in Singapore?
We found no Singapore statute aimed specifically at AI chatbots. The PDPA and the Consumer Protection (Fair Trading) Act apply in full, and in July 2026 PDPC published advisory guidelines on generative AI and IMDA published voluntary transparency guidelines for generative AI chatbots.
Who is responsible if a chatbot vendor mishandles data?
PDPC’s July 2026 guidelines say the business that deploys a generative AI system bears primary responsibility for PDPA compliance, even when it buys the system as a service or through an API. Check your vendor contract for security, overseas processing and use of data for training.
Can my chatbot ask for an NRIC number?
Only where the law requires it or where identity must be verified to a high degree of fidelity. Booking appointments, sending quotes or registering for trial classes does not meet that bar, so the bot should not ask for NRIC, FIN or similar numbers for those purposes.
Can I send promotions to numbers collected by my chatbot?
Not automatically. Under the PDPA’s Do Not Call provisions you need a valid DNC Register check, or clear and unambiguous consent evidenced in writing or another accessible form. Use a separate, unticked opt-in in the chat and keep the record.
Am I liable if my chatbot gives a wrong price or refund promise?
In our reading, yes. Misleading statements to consumers are unfair practices under the CPFTA, and a business cannot shift responsibility to software it deployed. In Moffatt v Air Canada, a Canadian tribunal held the airline liable for its chatbot’s wrong statement about a fare policy.
Do I have to tell users they are talking to a bot?
We found no Singapore legal requirement, but IMDA’s voluntary guidelines recommend a safety statement at first use and a published Chatbot Info Card. If you serve users in the EU, the EU AI Act requires disclosure from 2 August 2026.
