AI Architecture Review

# AI Architecture Review

Sample Deliverable

Executive Summary

This architecture review turns an AI product design into a set of reviewable security boundaries. It covers the model boundary, retrieval boundary, tool boundary, approval boundary, trace boundary, provider boundary, and release boundary.

The review does not ask whether the system “uses AI responsibly.” It asks the questions that matter before launch: what data moves, where authority expands, what can act, what is approved, what is logged, and what proof exists.

Heads up

Public sample notice

This is a shortened, synthetic excerpt prepared as a public sample. A client version would include system-specific evidence, implementation references, architecture screenshots, control test results, owner sign-offs, and full supporting documentation. This sample uses Northstar Support Cloud / Customer Support Copilot as the synthetic reference system. This sample is not legal advice, not a compliance certification, not an audit opinion, not a warranty, and not proof that any unreviewed system is secure.

Decision · conditional

Architecture review decision

Proceed with constrained pilot use, but do not expand retrieval sources, tool authority, or enterprise claims until the retrieval boundary, tool boundary, and provider boundary have stronger evidence.

Metrics

Architecture Review Snapshot

Boundaries reviewed

Critical boundary gaps

High-risk boundary gaps

Evidence-backed controls

Required companion artifacts

Note

Architecture review is where AI risk becomes concrete

AI security gets real when the diagram shows where data crosses trust zones, where generated text becomes action, where humans approve, where evidence is stored, and where buyers will ask for proof.

System in scope

Field	Value
System	Northstar Support Cloud / Customer Support Copilot
Status	Production pilot
Risk tier	Tier 4: agentic or state-changing
Business owner	VP Product
Technical owner	AI Platform Engineering
Security owner	Product Security
Primary use case	retrieve support context, draft responses, prepare workflow actions

Boundary review

Trust boundary map

Architecture Boundary Review

The architecture review maps the major AI product boundaries and identifies where implementation evidence is still partial.

Synthetic public-safe architecture review for AI product launch, covering model boundary, retrieval boundary, tool boundary, approval boundary, trace boundary, provider boundary, and release boundary.

Nodes

Boundaries

Flows

Controls

Architecture boundary review

Boundary	Risk	Owner	Status	Evidence
Model boundary	High	AI Platform Engineering	Partial	model-provider-boundary-statement
Retrieval boundary	Critical	Search Platform	Partial	rag-authorization-review
Tool boundary	Critical	AI Platform Engineering	Partial	agent-tool-inventory
Approval boundary	High	Product Operations	Partial	approval-context-bundle
Trace boundary	High	Security Engineering	Partial	ai-trace-schema
Release boundary	High	Product Security	Planned	ai-release-gate-checklist

Findings

Architecture Findings

Finding · critical

Retrieval boundary is designed but not proven

Evidence: rag-authorization-review

The architecture relies on authorization-preserving retrieval, but evidence does not yet prove the full path from source ACL to generated answer.

Heads up

Impact

A generated answer can become the place where access-control failure appears.

Finding · critical

Tool boundary needs action-class enforcement

Evidence: agent-tool-inventory

Tool access is not fully enforced as distinct action classes. This makes blast radius and approval requirements harder to reason about.

Finding · high

Provider boundary is not ready for enterprise review

Evidence: model-provider-boundary-statement

Provider route, training-use, retention, subprocessors, and data minimization need buyer-ready language tied to evidence.

Finding · high

Trace boundary has a sensitive evidence policy gap

Evidence: ai-trace-schema

AI traces are useful for audit and incident response, but they can also become a sensitive data store without retention and access controls.

Review questions

Architecture review questions

Boundary	Question	Status
Model	Are all model calls routed through the AI gateway?	Partial
Retrieval	Does authorization survive indexing, chunking, reranking, and prompt assembly?	Partial
Tool	Are actions separated by read, suggest, draft, queue, approve, and execute?	Partial
Approval	Can a reviewer make a meaningful approval decision?	Partial
Trace	Can AI behavior be reconstructed after an incident?	Partial

Required companion artifacts

Checklist

Companion artifacts needed for a complete review

✓AI Trust Boundary Map.

✓RAG Authorization Review.

✓Agent Tool Inventory.

✓Model Provider Boundary Statement.

✓AI Release Gate Checklist.

✓AI Risk Register.

Decision · conditional

Launch readiness decision

Launch readiness should depend on boundary evidence, not architecture confidence. Retrieval, tool, provider, trace, and release boundaries all need a named owner and evidence before expansion.

Related artifacts

Artifact

Related artifact: AI Trust Boundary Map

The trust boundary map is the diagrammatic companion to this architecture review.

/deliverables/ai-trust-boundary-map

Artifact

Related artifact: RAG Authorization Review

The RAG authorization review deepens the retrieval boundary evidence.

/deliverables/rag-authorization-review

Artifact

Related artifact: Agent Tool Inventory

The tool inventory deepens the tool boundary evidence.

/deliverables/agent-tool-inventory