AI Vendor Due Diligence Checklist Philippines

Last materially reviewed: September 4, 2026

Direct Answer

Before adopting an AI vendor, Philippine businesses should review more than price and functionality. The due diligence should cover what data the vendor receives, whether prompts or files are used for model training, retention and deletion, security controls, subprocessors, processing locations, incident notification, service changes, output ownership, confidentiality, audit evidence and exit terms. If personal data is involved, the Data Privacy Act and NPC requirements for controllers and processors also apply.

Key Takeaways

Identify exactly what data will enter the AI system and why.
Check whether prompts, files or outputs are retained or used to train models.
Review subprocessors, hosting locations and cross-border processing.
Require appropriate security, breach notification and deletion obligations.
Clarify output rights, confidentiality, model changes and termination/export options.

AI Vendor Evidence Pack: What to Request and Verify

Updated 3 October 2026. Use this pack for a defined use case and the exact product, plan and deployment. Record “not supplied” or “not verified” instead of assuming a vendor claim is true. Keep commercial and security documents in your own controlled records.

Evidence matrix · Decision record · Annotated sample clauses

Evidence Matrix

Read each row across: request evidence, verify it, assign an owner and record any reason to hold the proposed use. The hold indicators are recommended review gates, not statutory categories.

Vendor claims, verification and approval gaps
Topic	Request	Verify	Owner / hold indicator
Training and reuse	Terms for the exact plan/API; no-training commitment; workspace settings; exceptions for feedback, safety review and fine-tuning.	Save dated terms and settings; ask who can enable training and whether deletion reaches derived datasets.	Procurement + privacy lead. An opt-out switch contradicts broad reuse rights, or the vendor will not explain exceptions.
Retention and deletion	Separate schedules for prompts, uploads, outputs, logs, backups and support tickets; deletion procedure and legal-retention exceptions.	Run a synthetic deletion request; check active storage and backup expiry separately.	IT + privacy lead. “Deleted” is undefined or backups and legal holds have no documented limits.
Locations and subprocessors	Named provider chain, processing/support locations, access purposes and change process.	Compare the deployment region with support and subprocessor access; record transfer safeguards.	Privacy lead + Procurement. A selected hosting region is treated as proof that all access stays there.
Security and isolation	Access model, MFA, tenant isolation, encryption, relevant assurance scope and incident procedure.	Check report dates/exclusions and test permissions with invented records; certificates alone do not show your configuration is safe.	IT. Unrestricted connectors, shared accounts or missing evidence for material controls.
Incident cooperation	Reachable security contact; notification trigger, response commitments and sample incident information.	Exercise an incident using fictional details and record escalation contacts.	IT + privacy lead. The vendor waits for a final investigation before notifying you.
Input/output rights	Definitions, licences, training rights, output-use terms and any infringement protection.	Compare order form, online terms and data-processing agreement; list exclusions and caps.	Legal + Procurement. “You own outputs” hides broad reuse rights or says nothing about third-party claims.
Rights requests and audit	Assistance process, record-search capability, deletion support and evidence/audit arrangements.	Try a mock access/deletion request using a case token and request missing evidence.	Privacy lead. The provider cannot locate records or refuses meaningful compliance evidence.
Exit and material changes	Export format, assistance, costs, deletion confirmation, change notice and termination remedies.	Test an export and record what is lost; verify whether the purchased plan provides the promised controls.	Business owner + Procurement. Data is locked in, or material data-use changes can occur without an effective response.

Copyable Vendor Decision Record

Complete one record per use case; approval of a drafting tool is not approval of profiling or autonomous actions.

Use case / vendor / product / plan / region / date: [complete]
Data, users, integrations and permitted actions: [complete]
PIA reference and material risks: [link]
Evidence register: [topic; document/version/date; supplier; reviewer; finding; verification result]
Contract gaps: [clause; proposed change; vendor response; unresolved issue]
Action owner and due date: [one entry per gap]
Decision: [hold / synthetic-data test / approved for named use with conditions]
Approver, rationale and permitted data: [complete]
Residual risk, review date and change triggers: [complete]

Illustrative decision: the fictional retailer in the worked AI PIA has a no-training statement but no verified backup-deletion terms. Record the statement as received and deletion as unresolved; Procurement owns the follow-up. Live customer-data use remains on hold. Do not convert missing evidence into a low risk score.

Before approval: reconcile the signed terms with actual settings, resolve material legal and technical gaps, and attach evidence to the AI approval record. Management acceptance cannot make unlawful processing lawful.

Legal context: DPA IRR, Sections 43–45 and 50. This matrix is an editorial procurement aid, not a prescribed government form.