Data Privacy Act and AI: what to check before you deploy an AI assistant

RA 10173 and its implementing rules govern personal-data processing; AI does not create an exception. Advisory 2024-04 explains their application to AI systems. Start by identifying the information in messages, retrieved documents, generated answers and records sent to connected tools.

A customer’s inquiry can contain more information than the workflow needs. A retrieved document can include details about someone other than the person asking. For planning purposes, map those paths rather than assuming a public-facing assistant handles only public information. The same exercise matters for internal assistants using employee or client records.

Separate the legal assessment from the implementation decision. Your lawyer or DPO can assess the processing and applicable obligations. Your engineering team then needs a concrete specification of the approved data, access boundaries, review points and request-handling mechanisms. A useful assistant and a lawful processing arrangement are related questions, but neither should be inferred from the other.

Use NPC Circular 2022-04 for the registration discussion, rather than Circular 17-01: section 40 expressly supersedes the older circular. Section 5 addresses mandatory registration, including data processing systems involving automated decision-making or profiling. Sections 26–27 address notification through registration information. Check the system’s facts with your DPO before deciding how the rules apply.

Calling a system an assistant does not settle whether it performs automated decision-making or profiling. Document whether it only prepares information for a person or independently determines a result affecting someone. A lead tag, for example, needs to be understood in the context of how staff use it; a label alone is not the legal assessment.

Registration is also a different question from whether the processing is otherwise appropriate. A certificate or a vendor statement should not replace examination of the information flow and safeguards. Ask the responsible adviser to review the current circular and any later NPC guidance before launch, especially if the workflow or role of human review changes.

Use this as a preparation checklist for discussion with your DPO, lawyer and engineering team. It is not a declaration that completing a checklist establishes compliance. Record the decisions and owners so the approved workflow can be compared with the implementation and reviewed after changes.

1. Map personal data. List what enters through messages, documents and tools, what is retrieved, and what is stored or handed to staff. Include test examples and generated summaries, not only the final answer shown to a user.
2. Set the lawful basis. Ask the responsible adviser to assess the basis for each processing purpose under RA 10173, rather than assuming every activity uses consent or that an existing permission automatically covers a new AI workflow.
3. Write a plain-language notice. Explain the actual processing, purpose and relevant AI involvement in a way the intended audience can understand. Check the consent guidelines where consent is the basis; a notice and consent are distinct considerations.
4. Minimize accessible data. Identify the fields and documents needed for the task. Remove unrelated records from the assistant’s scope, and test that retrieval follows the user’s permissions rather than simply asking the model to avoid sensitive answers.
5. Limit tools and actions. Separate reading from updating records. Identify actions that need approval and ensure the system executing an action enforces that approval instead of relying only on the assistant’s wording.
6. Define human review. Name the people handling exceptions and consequential decisions, what context they receive and how they can intervene. A nominal review step is not useful if the reviewer cannot understand or change the proposed result.
7. Run a PIA. Use NPC PIA guidance and the applicable security rules with your DPO to examine the processing and risks. Consider the workflow before launch and revisit material changes rather than treating the assessment as a one-time document.
8. Plan data subject requests. Define the contact point, record locations and responsible team. Consider what happens when a request concerns a message, summary or information already used by the workflow, and what technical support the vendor can provide.
9. Perform vendor due diligence. Ask where information is processed, which organizations process it, what is retained, how access is constrained and how incidents and data subject requests are supported. Review the answers for the actual proposed configuration.
10. Test security boundaries. Consider a scoped AI security review for hostile instructions, retrieval access and tool permissions. Keep its findings and limitations separate from a legal compliance conclusion.
11. Plan monitoring and incidents. Assign ownership of unexpected behavior, access changes and suspected disclosures. Agree how the workflow can be paused, investigated and corrected, and how the responsible team will assess any reporting obligations under the applicable rules.

Ask questions about the proposed system rather than accepting a broad claim that a product is safe or compliant. The answer should identify the processing arrangement and the responsibilities around it. If the configuration changes, revisit the answers with the people responsible for the workflow and privacy review.

• Where is client data processed, and which organizations and underlying providers process it?
• What information is sent for the task, what is stored, and what retention and deletion arrangements apply?
• Which users can access each source, and how are those permissions enforced before retrieval?
• Which tools can the assistant call, and which proposed actions require a person’s approval?
• What security testing has been performed on this configuration, and what remains outside its scope?
• How can the business investigate, locate, correct or remove information when handling a data subject request?
• Who responds to incidents, system changes and unsuccessful hand-offs after launch?

ChuuniX runs AI on its managed AI service/API layer, so client data is processed through that service and its underlying model providers. We design assistants with data-access limits and human review in mind. That is a description of the delivery approach, not a guarantee of compliance or a substitute for your organization’s assessment.

The AI security review examines agreed technical boundaries and failure modes. A free discovery call for AI consulting can help define the workflow before a build, while AI integration services cover implementation. Clients in Metro Manila can request in-person meetings; we work mostly remotely across the Philippines and worldwide.

Post-launch support is available on a retainer covering monitoring, updates, model changes and security updates. Your team should retain an owner for processing decisions, approved information and human review. A technical support agreement should make it easier to operate the agreed workflow; it does not transfer every organizational responsibility to the builder.

For concrete planning examples, consider the personal data that may appear in property lead qualification or retail product inquiries. The approved task may be narrow while the incoming message contains more information. Return to the homepage to explore ChuuniX’s services, and involve your adviser before approving the processing arrangement.