Privacy Impact Assessment Philippines: Who Needs One and How to Run It

Worked AI Privacy Impact Assessment: Customer Reply Assistant

Fictional educational example — prepared 3 October 2026. “Example Retailer” and “Example AI Vendor” are invented. All consultation notes, test results and decisions below are illustrative, not findings about a real business or product. This is a completed assessment with an unresolved-risk decision, not an NPC form, certification or permission to process personal data.

Scope · Data flow · Purpose and basis · Risk register · Decision · Adapt this assessment

1. Assessment Record and Scope

• Reference: EXAMPLE-AI-PIA-01, version 1.0. Assessment commenced 1 October and completed for management review 3 October 2026.
• Controller and project owner: Example Retailer; Customer Support Manager. The DPO leads privacy review, IT owns technical controls, Procurement checks vendor terms, and the General Manager makes the business decision.
• Purpose: test whether an assistant can draft useful replies to order-status enquiries without exposing unnecessary customer information.
• Proposed scope: three support agents, one business workspace and a two-week pilot capped at 50 enquiries. These limits are illustrative design choices, not legal thresholds.
• Excluded: marketing, staff scoring, customer profiling, fraud screening, refunds, automated sending, model training and direct mailbox or order-system access.
• People affected: customers whose enquiries are selected, people mentioned in messages, and staff whose access is logged. Child-related, health, identity-document, payment-card and other sensitive content goes to the manual process.
• Current decision: synthetic-data testing only. The live-data proposal is on hold pending the actions in Section 7.

2. Data Inventory and Flow

Proposed live flow: customer enquiry → retailer helpdesk → agent prepares a minimal excerpt → vendor returns a draft → agent verifies and sends through the helpdesk → temporary copies expire. The model cannot call tools or retrieve other customer records.

1. Collection: the existing helpdesk receives names, email addresses, order references and free-text messages. The helpdesk remains the record system; the AI project does not change its independently justified retention schedule.
2. Preparation: the agent substitutes a random case token and removes names, addresses, contact details, order IDs, signatures and attachments. The retailer alone keeps the token mapping. Redacted free text may remain identifiable, so the assessment still treats it as personal data.
3. Vendor processing: send only the token, enquiry category, minimal excerpt and approved generic shipping policy. No full inbox, customer database or payment information is accessible. Proposed role: processor acting on instructions; independent training or reuse would require a fresh role and purpose analysis.
4. Output: a draft returns to the agent. The agent checks facts against the helpdesk and approved policy before manually sending. The AI output cannot determine entitlement to a refund.
5. Disposal and access: proposed AI-workspace retention is seven days; content-free security logs, 30 days. These are example limits requiring justification and verification, not NPC defaults. Exported files are disabled. Vendor backups, support access, locations and subprocessors are not yet verified, which blocks live use.

3. Necessity, Alternatives and Lawful Basis

Alternative assessed: existing reply templates with manual handling. They require no new AI recipient and remain available. The fictional team wants to test drafting quality, but has not established a benefit that outweighs additional exposure; speed alone does not justify disclosure.

Underlying order service: processing necessary to fulfil the customer contract may fall under RA 10173 Section 12(b). That does not automatically justify sending the entire conversation to an AI vendor. Optional AI-assisted drafting: the example proposes a separate Section 12(f) legitimate-interest assessment covering purpose, necessity and the balance with customer rights. It is unfinished, so no basis is treated as established for live AI processing. A privacy notice is not consent.

Sensitive or privileged information is outside this pilot; ordinary personal-information grounds do not substitute for the separate conditions in Section 13. Unexpected sensitive content must stay in the manual channel and any accidental upload must enter incident review. See RA 10173, Sections 11–13 and 21.

Proportionality decision: test invented messages first. If the benefit is not demonstrated, keep the manual workflow. If it is, reassess only the smallest live-data scope supported by the evidence.

4. Transparency, Rights and Stakeholder Input

The proposed notice will identify AI-assisted drafting, its purpose, data categories, recipients or recipient categories, relevant retention, safeguards and the retailer’s rights contact. Provide a manual support route. Requests for access, correction, objection or erasure go to the DPO, who uses the internal case token to locate records and coordinates any vendor action; responses must account for applicable rights and lawful retention duties.

Illustrative consultation record: Support requested faster drafting but rejected automatic replies; IT recommended removing connectors; Procurement could not verify backup deletion; the DPO required a completed basis assessment. Customer consultation has not occurred. Before a live pilot, test notice comprehension with a small volunteer panel using the proposed text and invented messages, record concerns and revise the design. Participation in that exercise is not consent for live processing.

Registration check: DPO must assess the actual system against NPC Circular 2022-04, including Section 5(A), and record whether registration or an update and automated-processing notification are needed. The absence of profiling in this design is not a blanket registration exemption. Adding scoring or autonomous decisions reopens the assessment.

5. Risk Register and Acceptance Method

Illustrative internal scale: likelihood 1 = unlikely, 2 = plausible, 3 = likely; impact 1 = limited inconvenience, 2 = material privacy harm, 3 = serious financial, identity or other harm. Multiply them: 1–2 low, 3–4 medium, 6–9 high. These are editorial working assumptions, not an NPC scoring system. Assess harm to people. An unresolved legal basis or material unknown blocks live use regardless of score. Residual scores below are targets after controls are evidenced, not proven current results.

6. Validation and Evidence Register

Fictional test findings: in a 20-message synthetic test, one message retained an address and two drafts misstated the returns policy. No customer data was used. The team therefore did not pass the proposed data-removal and answer-verification gates. These invented figures illustrate how to document failure; they are not a benchmark or a result from a real model.

• E1: versioned data-flow and access diagram — illustrative design recorded.
• E2: purpose/basis analysis and alternative comparison — basis review open.
• E3: vendor terms, subprocessors, locations, training and deletion evidence — incomplete.
• E4: synthetic tests, expected answers, failures and retest results — failures recorded; retest pending.
• E5: notice, rights-request drill and staff briefing — not yet completed.
• E6: registration assessment, incident procedure and management decision — assessment open; decision below recorded.

Keep the evidence under access control. Record the reviewer, date and document version for each item; do not put raw customer messages into a widely shared PIA report.

7. Action Plan and Management Decision

• IT and Support: fix the redaction and draft-review failures and rerun the synthetic cases before any live-data request.
• Procurement: obtain vendor training, deletion, backup, location and subprocessor evidence before contract approval.
• DPO: complete the basis, necessity, rights and registration assessments before a live-data decision.
• Support QA: test the notice and manual route, brief staff and document an incident drill before pilot access.

Illustrative decision dated 3 October 2026: the General Manager accepts the DPO’s recommendation to hold live processing. Only invented-message testing in the isolated workspace may proceed. Procurement cannot approve customer-data access, and no medium/high residual risk has been accepted for a live pilot. Each owner must supply closure evidence; the DPO reassesses risk and management records a new decision. Internal risk acceptance cannot cure unlawful processing.

Sign-off record: Prepared by DPO role; technical findings checked by IT lead role; operational findings checked by Support Manager role; decision made by General Manager role. These are fictional role entries, not actual signatures. In a real report, record the authorised people, dates, recommendations, objections and signed decision.

8. Monitoring, Incidents and Review

If a live pilot is later approved, Support QA reviews every draft before sending and reviews the issue log daily; the DPO reviews the pilot after two weeks. Suspend the affected use case on an unexpected upload, disclosure or unapproved action. Preserve relevant evidence, restrict access and invoke the breach-response process, including a separate assessment of notification duties.

Reopen the PIA before adding a connector, new data category, training use, vendor/subprocessor, processing location, longer retention, profiling or autonomous sending; also after an incident or relevant legal change. Keep versions and reasons for each decision. The proposed cadence is a design choice, not a universal statutory schedule.

9. Use This Structure for Your Own Assessment

Copy Sections 1–8 into your internal document. Replace every fictional fact, role entry, score and test result with your own evidence. Keep “unknown” where verification is incomplete. Attach the flow map, basis analysis, vendor evidence, risk/action register, tests and decision record; assign a review date and owner.

Start with the AI inventory and approval record, obtain evidence with the vendor evidence and decision pack, and carry approved restrictions into your staff AI policy.

Legal anchors checked for this example: NPC PIA guide; NPC Advisory 2024-04; NPC Circular 2022-04; and RA 10173. The older general-guidance discussion retains its separate review date. This example is general educational information and needs adaptation and professional review for the actual processing.