PDPC's Generative AI Guidelines: What System Deployers Must Do

Most Singapore enterprises using generative AI are not training foundation models. They are deploying them: connecting a model to internal documents, customer tickets, CRM records or email through a workflow platform or an AI agent. On 20 July 2026, the Personal Data Protection Commission (PDPC) published guidance that speaks directly to this situation, and its central message for deployers is uncomfortable but clear: the organisation deploying the system bears primary responsibility for PDPA compliance. This article explains what the guidance says about roles, what it expects of deployers, and how to turn it into operating practice for AI workflows.

What PDPC Published

PDPC, with support from the Infocomm Media Development Authority (IMDA), issued the Advisory Guidelines on Use of Personal Data in Generative AI on 20 July 2026. According to PDPC's announcement, the guidelines cover three things:

• How personal data can be collected and used to develop GenAI models, including web-scraping and the re-use of data originally provided for non-GenAI purposes.
• How data protection responsibilities are allocated across the GenAI lifecycle.
• How organisations should handle individuals' requests about the processing of their personal data for GenAI.

The guidelines are organised around three lifecycle stages: development, deployment and post-deployment. PDPC says they incorporate feedback from a public consultation that closed on 1 July 2026, and that they build on, and should be read with, two earlier documents: the Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems (published 1 March 2024) and the Advisory Guidelines on Key Concepts in the PDPA. If your organisation already aligned with the 2024 guidelines for recommendation or decision systems, the GenAI guidance extends that work rather than replacing it.

For context on the PDPA's general obligations as they apply to AI in customer-facing marketing, see our earlier guide to PDPA compliance for AI marketing. This article focuses on the new GenAI-specific guidance and on the deployer role.

Three Stakeholders, Three Sets of Responsibilities

The most practically useful part of the announcement is its split of responsibilities in the deployment stage:

Many organisations will occupy more than one role. A bank that buys a workflow platform (system provider) running on a third-party foundation model (model provider) and configures it to process customer emails is the system deployer. An enterprise that builds its own internal assistant on top of a model API may be both system provider and deployer. Map your role per use case before assuming someone else holds the obligation.

What Deployers Are Expected to Do

PDPC's summary names three deployer responsibilities. Each translates into concrete work.

1. Define clear purposes for processing

Every GenAI workflow that touches personal data should have a written purpose: what the system does, which personal data it needs to do it, and what it must not be used for. "General productivity" is not a purpose. "Summarise inbound customer service emails and draft replies for agent review" is. A clear purpose makes it possible to test whether the data flowing into the system is necessary, and whether consent or an exception covers it.

2. Safeguard data flowing through the system

Generative AI workflows move data in ways that traditional applications do not: into prompts, retrieval indexes, conversation logs, tool calls and outputs. Deployers should know where each of those stores sits, who can access it, and how long it is kept. The announcement notes that model providers should pay particular attention to data retention; deployers should ask their providers directly how prompts, outputs and logs are retained, and configure their own retention accordingly.

3. Review safeguards regularly, especially for agents

PDPC singles out agentic AI systems for regular review. An agent that can read a CRM, send emails or update records has far more ways to expose personal data than a chatbot that only returns text. Reviews should cover which tools and data sources each agent can reach, whether permissions are narrower than the user's own, and whether actions involving personal data are logged. Our page on data protection and enterprise security sets out the principles we apply here.

The provider-side "best practice" of sharing safeguard information is only useful if deployers ask for it. Build requests for model-level and system-level safeguard documentation into procurement, alongside your usual security questionnaire.

Re-Using Customer Data for GenAI Development

If your organisation fine-tunes or otherwise develops a model using data customers gave you for another purpose (PDPC calls this "User Data"), the announcement states that where exceptions to consent do not apply, organisations must obtain consent through "AI-Specific Notifications". These should explain the purpose of use, what data will be used, how it will be used, and how individuals can decline or withdraw consent.

For web-scraped data, PDPC indicates developers may rely on the Publicly Available Exception to collect publicly accessible personal data without consent, but data behind digital barriers such as paywalls or registration requirements must be carefully assessed to see whether it still qualifies as publicly available. Enterprises that only deploy third-party models will rarely scrape data themselves, but the question is still relevant when evaluating a model provider's data practices.

Handling Access and Correction Requests

Individuals can request access to and correction of their personal data even after it has been used in GenAI development. PDPC acknowledges this is difficult, because training data is vast and is not stored in a traditional repository, but it still expects organisations to adopt best practices: good upstream data handling, case-by-case review of requests, and tracking and adopting appropriate technical measures.

For deployers, the more common scenario is simpler: a customer asks what data your AI workflow holds about them. That is answerable only if you know where prompts, retrieved documents and outputs are stored. Designing for this up front is much cheaper than reconstructing it after a request arrives.

A Deployer's Working Checklist

Where personal data decisions sit alongside wider AI risk questions, such as human approval before an agent acts, our overview of AI governance and human oversight explains how we separate AI suggestions from human-approved execution.

Frequently Asked Questions

Are PDPC's GenAI guidelines legally binding?

They are advisory guidelines that clarify how PDPC interprets the PDPA in the GenAI context. The PDPA itself is binding, and the guidelines indicate how PDPC expects organisations to meet it.

We only use a vendor's GenAI tool. Are we still responsible?

If you deploy the system to process personal data, PDPC's guidance treats you as the system deployer, which bears primary responsibility for PDPA compliance, including defining purposes and safeguarding data.

Do the 2024 AI recommendation guidelines still apply?

Yes. PDPC says the GenAI guidelines build on and should be read together with the Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems and the Advisory Guidelines on Key Concepts in the PDPA.

Why does PDPC mention agentic AI specifically?

The announcement asks deployers to regularly review safeguards especially for agentic AI systems, which can take actions and reach more data sources than systems that only generate text.