PDPA and AI agents: a practical guide for Singapore SMEs

Why this comes up in every conversation

A WhatsApp thread with a customer is personal data: a name, a phone number, often an address, an order history, sometimes a photo of their living room or their invoice. Put an AI agent on that thread and you haven't created a new category of law — you've added a new processor to an existing obligation. That's the honest frame: the PDPA applied to your enquiries before the agent arrived; the agent makes it worth checking that your house is actually in order.

The obligations that matter for an AI front desk

The Personal Data Protection Act's requirements are broader than this, but five areas do most of the work for a service SME:

• Consent and purpose. Collect and use personal data for purposes a reasonable person would consider appropriate, and tell customers what those purposes are. A customer messaging you about a booking expects their details to be used to handle the booking — not to be quietly added to a marketing list.
• Protection. Make reasonable security arrangements: encryption, access control, and a clear answer to "who can read these conversations?"
• Retention. Keep personal data only as long as a business or legal purpose exists, and be able to delete it when that ends.
• Breach notification. Assess data breaches, and notify the PDPC and affected individuals when a breach is likely to cause significant harm or reaches significant scale.
• Accountability. Every organisation must designate someone responsible for data protection — the DPO obligation applies whether or not AI is involved.

And if you send marketing messages to Singapore telephone numbers — including on WhatsApp — the Do Not Call provisions apply: check the DNC registry, honour withdrawals, and keep records.

What the PDPC has said about AI, in plain English

In March 2024 the PDPC published Advisory Guidelines on the use of personal data in AI recommendation and decision systems. They are advisory rather than binding law, but they signal how the Commission thinks about enforcement, and three threads matter for SMEs:

• You still need a lawful basis — typically meaningful consent, or one of the PDPA's exceptions (such as business improvement), each of which comes with conditions.
• Transparency is expected — be able to explain to a customer, without theatrics, that an AI system helps handle enquiries and what data it uses.
• Accountability doesn't outsource — engaging a vendor doesn't transfer your obligations; it adds a party you're responsible for choosing well.

Separately, IMDA's Model AI Governance Framework (including its generative-AI edition) is a useful, free reference for what "responsible deployment" looks like operationally: human oversight, monitoring, and clear accountability for what the system sends. If those phrases sound familiar, it's because approval-first deployment is that framework made concrete.

Eight questions to ask any AI vendor

Whether you talk to us or anyone else, these eight questions separate real answers from badges:

• Is our workspace isolated from other customers', or is it a shared multi-tenant pool?
• Is data encrypted at rest and in transit — including the connector tokens and API keys that reach into our systems?
• Who on your side can access our conversations, and under what controls?
• Is there an audit trail we can inspect — every message sent, every action taken, who approved it?
• Where is data processed, and can the deployment region be chosen if we have residency requirements?
• What happens to our data when we leave — export, deletion, timelines?
• Which AI model providers see our data, and do we control which models are used for which tasks?
• Does anything outbound leave without a human approving it — and can we tighten that per action category?

How Olano approaches it

Our answers to those questions, on the record: each customer runs in a private, isolated workspace — no shared multi-tenant data. Conversation data is encrypted at rest and in transit; connector tokens and API keys are encrypted; access is role-based; and an immutable audit trail records every action and approval. Outbound sends are approval-gated by default, with trust levels 0–4 to raise autonomy deliberately. For residency requirements, custom AWS/GCP regions — or customer-controlled cloud/VPC deployments — are available, and the client controls which model providers handle which tasks and where data is processed. The same architecture is described on our security section and in the Singapore WhatsApp guide.

What stays your responsibility

No vendor can make you compliant by invoice. Even with good infrastructure, the SME still owns:

• the privacy notice and consent wording customers actually see;
• the DPO designation and someone who owns data questions day to day;
• retention decisions — how long enquiry threads and lead records should live;
• DNC compliance for any outbound marketing the team approves;
• a simple breach response plan: who assesses, who notifies, on what clock.

A deployment conversation is a good forcing function for all five — we walk through them during onboarding rather than after an incident.

This article is general information as of July 2026, not legal advice. The PDPC's published guidelines and advisories are the authoritative source.