The industry's AI failure is not technical — it's governance: POCs that never earn production, personal data leaking into prompts and traces, and platform bets made on demo enthusiasm. This module gives the architect the gates, the data-flow discipline, and the build-vs-buy method that turn AI ambition into systems a regulator and a board can live with.
This closing module belongs to the enterprise architect's altitude: not one system's design but the org's AI stance. The stakes are documented on both sides — Pluralsight's 2026 research puts the share of organizations seeing zero GenAI ROI near 95%, while India's DPDP Rules 2025 phase in penalties reaching ₹250 crore from May 2027. Ungoverned AI fails commercially; governed-by-paperwork AI fails to ship. The architect's job is the third path: governance expressed as architecture gates that things can actually pass.
A business unit arrives at Vikram's ARB with a document-summarization POC for loan operations. It demoed brilliantly to the CEO. The ask is one line: "it works — we just need production access and real customer data."
Vikram runs the readiness screen before the meeting: no data-flow diagram, no evals — quality is a highlight reel of good demos — real customer documents were pasted into the demo prompts, the model endpoint egresses to the public internet, no deletion path, no cost model.
The team argues momentum: the CEO loved it. Vikram doesn't argue back with policy PDFs — he shows two numbers: the ~95% zero-ROI statistic for exactly this pattern, and ₹250 crore, the DPDP penalty ceiling that starts biting in May 2027, then walks the data-flow gaps on one slide.
Gated, not killed. Conditional pilot: synthetic data until the DPIA-style review lands, private-endpoint inference, an eval harness with a golden set before any quality claim, a data-flow diagram covering prompts, embeddings, outputs and traces, deletion paths, and a cost model with bands.
The team returns with the gates passed — and a better product: the eval harness caught a summarization failure mode the demos never surfaced. Production approval follows, with a quarterly platform review attached.
Vikram turns the one-off review into the ARB's standing responsible-AI gate: every AI initiative now meets the same checklist at POC exit, pilot exit and production entry. The next team designs against it from day one — which was the point.
Notice the posture: the ARB that only says no gets routed around — shadow AI is the certain result. The ARB that converts enthusiasm into gates gets passed through, and every gate it enforces is one this module teaches you to design.
Classical data-flow mapping tracked databases, queues and backups. An AI system adds four surfaces the org's existing data map doesn't show — and the DPDP practitioner analysis is explicit that every RAG system is part of the compliance risk surface, because prompts, embeddings and model outputs all carry personal data:
| Surface | How personal data enters | Design obligation |
|---|---|---|
| Prompts | Users paste it; RAG retrieves it into context | Redaction gateway upstream; prompt logs classified as personal data |
| Embeddings | Vectorized source documents encode it | Vectors inherit source classification; deletion must reach the index |
| Outputs | The model repeats or infers it | Output logging under the same retention rules as inputs |
| Traces & logs | Observability captures full prompts/completions | Retention limits, PII policy for traces, access control on the observability stack |
Each surface needs three things designed, not documented after the fact: consent lineage (which lawful basis covers this data reaching this surface), retention (how long it lives there, including caches and eval samples), and a deletion path (given one person's erasure request, enumerate and remove their data from all four surfaces — if the answer involves "rebuild the index," design the tenant- or user-scoped partitioning that makes deletion tractable).
The India specifics make this concrete. The DPDP Rules were notified in November 2025 with phased compliance deadlines running to May 2027, when full enforcement — and penalties up to ₹250 crore — begins. Two design consequences separate DPDP from GDPR thinking: first, DPDP recognizes only consent and legitimate use as lawful bases, against GDPR's six — no legitimate-interest catch-all, which tightens what a RAG pipeline may ingest for India-scoped personal data. Second, there is no blanket localization mandate today, but the government may restrict transfers per destination — so the architecture goal is restrictability: region-pinned deployments and portable vector stores, so that if a restriction lands, re-homing is a migration, not a rebuild.
An observability stack that logs full prompts and completions is a personal-data store under DPDP and GDPR — if your trace retention and access controls weren't designed for that, the compliance gap is already in production.
For BFSI, government and client-MSA-restricted data classes, "the model is an API on the internet" is disqualifying. The pattern library is now first-party:
The platform decision is the largest reversible-looking irreversible decision in the AI stack, and it deserves the same weighted-ADR treatment as Module 5's RAG decision — with dimensions specific to this market:
The output is a weighted decision record — and because platform churn is structural in this market, the record carries a quarterly revisit date. A platform ADR without a review cadence is how orgs end up strategically committed to last year's product line.
Write the exit-cost section of the platform ADR first — it forces the portability disciplines (prompts, evals, tool contracts in your repo, provider-neutral) that make every later platform decision cheaper, whichever way this one goes.
The zero-ROI statistic has a mechanism: demos are selected evidence, and production is unselected reality. POCs skip exactly the things production is made of — security, reliability, evals, cost controls, compliance — so the organization accumulates impressive demos and no returns. The fix is not to ban POCs; it is to stage them through gates with published exit criteria, so "it works" has a definition:
Synthetic or public data only. Explicitly labeled "POC ≠ production." The deliverable is a findings memo — what was learned, what breaks, what production would require — not a deployment.
Limited users, bounded data classes under a completed data-flow review, private endpoints where required, eval harness gating changes, cost model with bands filed.
Full controls: responsible-AI gate passed, audit trails live, kill switches tested, unit economics metered, ownership and rollback paths named.
Two enablers make the stages workable rather than bureaucratic. Synthetic data (schema-faithful generation with LLM structured outputs or SDV-class tools) lets teams demo and pilot without touching regulated data — for services firms, it's often the only client-MSA-compatible way to show anything. And the readiness assessment itself is AI-accelerated: walking a POC through the Well-Architected GenAI-lens question set with an assistant drafting the gap register turns a week of assessment into a day — with the architect owning every risk rating, per this track's standing division of labor.
The scored deliverable of a POC is the findings memo and the gap register, not the demo — a POC that produces only a demo has produced marketing, and the ~95% statistic is what happens when organizations ship marketing.
The frameworks are now table stakes in JDs — NIST AI RMF, ISO 42001, and AWS's Responsible AI Lens (shipped at re:Invent 2025 alongside the updated GenAI and new Agentic lenses). The architect's job is translation: from framework prose to gates a review can pass or fail. A workable ARB checklist covers:
Run the gate as review-by-checklist with artifacts attached, AI-assisted where it helps (drafting the pre-read, checking a design doc against the checklist) and human-owned where it matters (every rating, every waiver, every sign-off). That's the same trust discipline this track opened with, applied at governance altitude.
Assessment scenarios in this domain reward the gated-not-killed posture — the correct ARB answer converts an ungovernable request into staged exit criteria, and the correct compliance answer names the specific surface (prompts, embeddings, outputs, traces) where the personal data actually lives.
AI governance that works is architecture — four data surfaces with deletion paths, sovereignty as network design, platform bets as weighted ADRs with revisit dates, POCs staged through published exit gates, and responsible-AI frameworks compiled into checklists an ARB can run — because a control that isn't a design artifact is a policy PDF, and policy PDFs don't stop ₹250 crore penalties or a 95% failure rate.
Take this module into Lab 14 (DPDP-compliant RAG data-flow design) — for an India-scoped HR-docs copilot, you'll produce the data-flow diagram across prompts, embeddings, outputs and traces, the consent-lineage and deletion-path design, the residency ADR, and the redaction-gateway placement — the exact artifact set Vikram's gate demands, and the capstone input for the track's full AI solution architecture package.
Companion playbooks: pb-48 (POC → Production Readiness Gap Assessment), pb-49 (Compliance-Aware AI Data-Flow Design — DPDP/GDPR), pb-50 (Build vs Buy — Bedrock, Azure AI Foundry, Vertex, or Self-Hosted), pb-51 (Sovereign & Regulated Deployment Patterns), and pb-52 (Responsible-AI Architecture Gates for the ARB).