PDPA-Aware Data Product Canvas

Architecting Consent-Led AI Ingestion

Consent is not a checkbox — it is runtime metadata. Governance accelerates AI when it is designed as infrastructure.

Consent coverage
92%
+4.1% QoQ
AI-ready data products
38
+6 this quarter
Retrieval approval rate
76%
+2.8%
Blocked risky retrievals
184
Last 30 days
Audit-ready AI answers
98%
SLA: 99%
Immature AI Flow

Risky AI Ingestion: Fast, but Unprovable

Pre-PDPA
Raw Enterprise Data
Data Lake
Embeddings
LLM
User Answer
Unknown consent status
No lineage
Stale data
Cross-border ambiguity
No audit trail
Data leakage risk
Compliance bottleneck
PDPA-Aware Flow

PDPA-Aware AI Ingestion: Governed, Traceable, Scalable

Production-Grade
Raw Data
Governed Data Product
Consent + Lineage + Quality
Approved Retrieval
AI System
Audit Trail
Runtime consent metadata
Full source lineage
Freshness SLAs
Jurisdiction-aware
Immutable audit trail
Tenant-scoped access
Revocation within SLA
Architecture Flow Map

From data lake to governed AI substrate

Stage 1
Raw Data Sources
Origin systems
CRMTransactionsSupport ticketsApp eventsDocumentsLab recordsMerchant data
Stage 2
Governed Data Products
Contract-first datasets
PurposeOwnerLineageConsent scopeQualityJurisdictionExpiry
Stage 3
Consent + Lineage + Quality Gate
Runtime policy engine
Consent valid?Purpose allowed?Fresh?Lineage known?Cross-border?Tenant allowed?
Stage 4
Approved Retrieval Layer
Permissioned substrates
Vector DBRAG pipelineFeature storeKnowledge graph
Stage 5
AI System
Enterprise assistant
Permissioned data onlyTenant-scopedReason w/ citations
Stage 6
Audit Trail
Immutable provenance
WhatWhy allowedVersionJurisdictionApprover
Governed Data Product

Customer Transaction Data Product

dp.payments.customer_txn · v2.8

Consent ValidRisk: Medium
Data Owner
Payments Platform Team
Purpose
Fraud monitoring & customer support
AI Usage Allowed
RetrievalModel Training
Residency Rule
Singapore + approved ASEAN markets
Quality Score
94%
Lineage
Core banking → Lakehouse → Approved vector index
Revocation Handling
Removed from retrieval index within 4h SLA
Consent Scope
Risk assessment, fraud prevention, customer support
Policy Engine

Consent + Lineage + Quality Gate

7 checks · evaluated in 38ms
  • Consent exists
    Customer consent on file (2025-03-12)
    Pass
  • Purpose match
    Fraud monitoring ∈ allowed purposes
    Pass
  • Data freshness
    Updated 6m ago · SLA < 24h
    Pass
  • Cross-border transfer
    SG → ID requires DPO review
    Review
  • Sensitive attribute detected
    PAN fragment in 0.4% of rows
    Escalate
  • Retrieval permission
    Tenant: grab-fin-ops · Approved
    Pass
  • Training permission
    Personal data — training prohibited
    Blocked
Thesis

Enterprise AI becomes defensible at the data layer, not the model layer.

It becomes defensible when the data entering the model is consent-aware, lineage-rich, quality-scored, jurisdiction-aware, and audit-ready. Smarter models without governed substrates create faster liabilities.

—Consent is not a checkbox. It is runtime metadata.
—Data products before embeddings.
—Governance accelerates AI when designed as infrastructure.
—Every AI answer should prove what it was allowed to know.
Reasoning & Retrieval Trace

AI Answer Provenance

Human review
Query
“Can this customer be offered a credit limit increase?”
Retrieved Data Productdp.payments.customer_txn
Consent BasisCustomer support + risk assessment
Data Versionv2.8
JurisdictionSingapore (PDPA §13)
Retrieval Timestamp2026-06-02 09:41:22 SGT
Confidence87%
Final Answer StatusDraft recommendation only
Every AI answer proves what it was allowed to know.