Subodh KCSubodh KC/systems
let's talk →
Interactive Architecture Tool

Architecture Decision
Master Sheet

An interactive 25-layer architecture decision sheet for software development PMs, architects, and tech leads — with AI development risks, CSM pillar mapping, documents per step, and Definition of Done for every layer.

July 17, 2026By Subodh KC

Educational notice: This master sheet provides a comprehensive framework for architecture decisions. Production deployments should undergo organization-specific architecture review with qualified architects and security professionals. Decisions should be tailored to your specific tech stack, team composition, and regulatory environment.

Synopsis

The Architecture Decision Master Sheet covers 25 architecture layers across 7 groups — from product scope to governance. Each layer includes purpose, criticality, phase, decision options, RACI (Responsible, Accountable, Consulted, Informed), risks, AI development risks, documents for architects and PMs, and a Definition of Done. Aligned with the Cognitive Systems Management (CSM) framework, it bridges strategy, execution, governance, and operations. Use it alongside the AI Risk Register, AI Security Tools, and the Incident Evidence Checklist.

What Is the Architecture Decision Master Sheet?

The Architecture Decision Master Sheet is an interactive reference covering 25 architecture layers that every software development project must consider — from product scope to governance. It is designed for PMs, architects, and tech leads who need a structured way to track decisions, risks, and deliverables across the full lifecycle of a software project, whether or not AI is part of the product.

Each layer answers seven questions: Why does it matter? Who is responsible? How is it implemented? When should it be addressed? What are the risks? What does bad look like? And what documents should architects and PMs produce?

The sheet also includes AI development risks for every layer — not just the AI-specific layers (Group E). If your team uses AI tools to write code, or if your product includes AI features, these risks apply to you.

CSM-Aligned: Every layer maps to one of the four pillars of Cognitive Systems Management (CSM)Strategic Alignment, Technical Implementation, Governance & Risk, and Operational Excellence. Use the CSM Pillar View in the interactive tool to see how the sheet operationalizes the framework.

Quick Start — 5 Steps

1

Pick a scenario

Select the card that matches your project: web app, AI product, AI tools, audit, or pilot recovery.

2

Filter & explore

Filter by phase, CSM pillar, criticality, or keyword. Click any layer for 6 tabs of detail.

3

Assign owners

Open the PM Tracking tab per layer. Set status, sprint, owner, and ADR reference. Edits save to your browser.

4

Review AI risks

Switch to AI Risk View. Check AI pitfalls and quality gates for every layer — especially if using AI code assistants.

5

Export & share

Export CSV or JSON with all 25 layers, risks, DoD, and tracking data. Import into your PM tool or spreadsheet.

Who Is This For?

Project Managers

Track which architecture decisions are pending, who owns them, and what documents are needed per sprint. Use the PM Tracking tab to assign owners and statuses. Export the sheet to share with stakeholders.

Architects

Review all 25 layers before design begins. Document ADRs for each decision. Use the Decision Details tab to capture options, tools, and reversibility. Identify one-way doors before they become expensive.

Tech Leads

Use the Phase Matrix to sequence work across sprints. Check dependencies and blockers before committing to a sprint plan. Review anti-patterns to avoid common mistakes.

AI Leads

Focus on Layers 15–17 (AI Pipeline, Agents, Validation). Use the AI Risk View to see AI-specific risks across all layers. Ensure quality gates are enforced in CI before AI features ship.

DevOps & SRE

Focus on Layers 22–25 (Observability, DevOps, Performance, Governance). Ensure rollback strategies are tested. Set up AI-specific monitoring (drift, cost, quality) before launch.

Security & Compliance

Focus on Layers 18–21 (Security, Validation, Testing, Code Review). Use the CSM Pillar View to filter by Governance & Risk. Map controls to NIST AI RMF, ISO 42001, and regulatory requirements.

Interactive Decision Sheet

Filter by group, criticality, phase, CSM pillar, or keyword. Click any layer to see full details across 6 tabs. Use the floating ? button at the bottom-right to open the glossary. Your tracking edits save automatically to your browser.

Architecture Decision Master Sheet

A practitioner's guide for software development PMs, architects, and tech leads — with or without AI.

Software Dev PMArchitectTech LeadAI Development Risks25 LayersDocuments per StepCSM-Aligned

Web app without AI

Layers 1–14, 18–25. Standard delivery decisions.

AI-powered product

All 25 layers. Emphasis on AI Pipeline, Agents, Validation.

Using AI tools to develop

All layers. Focus on AI Development Risks per layer.

Auditing existing system

8 key layers: Observability, Governance, Security, Data, Services, API, Testing, Caching.

Recovering stalled AI pilot

7 layers: AI Pipeline, Agents, Validation, Data, Security, Observability, Governance.

25Layers
14Critical
8Important
0Decided
0Blocked
25Not Started
3AI Layers
Group
Criticality
Phase
Status
Applicability
CSM Pillar
#LayerPurposeCriticalityPhaseStatusAI Risk
Foundation & Scope
1Product & Scope
What are we building and why
Strategic Alignment
Aligns technical work to business outcomes; prevents scope creepCriticalDiscoveryNot Started
2Domain Modeling
What are the core business entities and rules
Strategic Alignment
Prevents table-driven chaos; ensures consistent business rules across the systemCriticalDiscoveryNot Started
3Tech Stack & Framework
What technologies are we using
Strategic Alignment
Defines the tools, languages, and frameworks; sets team composition and constraintsCriticalDiscoveryNot Started
4Build vs Buy Strategy
What do we build in-house vs buy as SaaS
Strategic Alignment
Optimizes cost, speed, and risk; prevents reinventing wheels; isolates vendor dependenciesImportantDiscoveryNot Started
Interface & Interaction
5Client / UI Layer
How users interact with the system
Technical Implementation
Directly impacts adoption, usability, and accessibilityImportantDesignNot Started
6UI State Management
How the interface remembers things
Technical Implementation
Keeps interface consistent and predictable across user sessionsStandardDesignNot Started
7API & Gateway
How the system controls access
Technical Implementation
Protects system and standardizes access; defines the contract between frontend and backendCriticalDesignNot Started
Logic & Data
8Service Layer
Where business logic lives
Technical Implementation
Centralizes business logic; ensures consistent rule enforcementCriticalBuildNot Started
9Controllers / Routing
How requests are routed and responses are structured
Technical Implementation
Keep endpoints thin and consistent; standardize request/response handlingStandardBuildNot Started
10Data & Persistence
How data is stored and organized
Technical Implementation
Performance and correctness depend on it; schema decisions are hard to reverseCriticalDesignNot Started
11Caching & Performance
How we make things fast
Technical Implementation
Protects UX and budget; prevents unnecessary load on databases and APIsImportantBuildNot Started
Integration & Async
12Integration / Adapters
How we connect to external services
Technical Implementation
Isolates vendor risk; standardizes external communication; enables swapping vendorsImportantBuildNot Started
13Events & Async
How background work gets done
Technical Implementation
Keeps UX fast and reliable; handles background processing without blocking usersImportantBuildNot Started
14Middleware
Rules that run before every request
Technical Implementation
Enforces cross-cutting concerns early; standardizes pre-processingStandardBuildNot Started
AI Layer (Conditional)
15AI / RAG PipelineAI
How AI finds and uses information
Technical Implementation
Delivers grounded, cost-effective AI; prevents hallucinations; controls costCriticalDesignNot StartedHigh
16AI Agent ArchitectureAI
How AI agents work and escalate
Technical Implementation
Orchestrates multi-agent workflows; defines tool use, autonomy boundaries, and human escalationCriticalDesignNot StartedHigh
17AI Output ValidationAI
How we check AI outputs are correct
Technical Implementation
Prevents hallucinated, biased, or harmful outputs from reaching users; ensures qualityCriticalBuildNot StartedHigh
Quality & Security
18Security & Policy
How we protect the system and data
Governance & Risk
Prevents abuse; meets compliance; protects user data; enables auditCriticalDiscoveryNot Started
19Validation & Integrity
How we prevent bad data from entering
Governance & Risk
Prevents bad data and exploits; ensures data consistency at system boundariesImportantBuildNot Started
20Testing Strategy
How we verify the system works
Governance & Risk
Ensures correctness; prevents regressions; enables confident deploymentCriticalDesignNot Started
21Code Review & Quality Gates
How we review code before it ships
Governance & Risk
Prevents bad code from merging; ensures consistency; catches AI-introduced issuesImportantBuildNot Started
Operations & Delivery
22Observability
How we see what is happening in the system
Operational Excellence
You cannot fix what you cannot see; enables fast debugging and proactive issue detectionCriticalLaunchNot Started
23DevOps & Delivery
How we safely ship changes
Operational Excellence
Safely ships changes; enables rollback; prevents broken deploys; manages environmentsCriticalDiscoveryNot Started
24Performance & Cost
How we keep things fast and affordable
Operational Excellence
Protects UX and budget; prevents cost overruns; ensures system scalesImportantOperateNot Started
25Ownership & Governance
Who owns what and how decisions are tracked
Operational Excellence
Clarifies accountability; ensures decisions are documented, reviewed, and updated; enables auditCriticalDiscoveryNot Started

The 25 Layers at a Glance

Group A — Foundation & ScopeLayers 1–4 · Strategic Alignment
Group B — Interface & InteractionLayers 5–7 · Technical
Group C — Logic & DataLayers 8–11 · Technical
Group D — Integration & AsyncLayers 12–14 · Technical
Group E — AI Layer (Conditional)Layers 15–17 · Technical
Group F — Quality & SecurityLayers 18–21 · Governance & Risk
Group G — Operations & DeliveryLayers 22–25 · Operational Excellence

Each group maps to a CSM pillar. Group E (red) is conditional — only if your product includes AI features.

The 7 Layer Groups Explained

Group A — Foundation & Scope (Layers 1–4)

These layers define what you are building and why. They are the highest-leverage decisions in the entire sheet. Getting these wrong means every downstream layer inherits the wrong assumptions.

CSM Pillar: Strategic Alignment

  • Product & Scope — What problem are you solving?
  • Domain Modeling — Core business entities and rules
  • Tech Stack & Framework — Technologies and constraints
  • Build vs Buy Strategy — In-house vs SaaS decisions

Group B — Interface & Interaction (Layers 5–7)

These layers define how users interact with the system. Key decisions include rendering strategy, state management, and API gateway design.

CSM Pillar: Technical Implementation

  • Client / UI Layer — Rendering, design system, accessibility
  • UI State Management — Global vs server vs local state
  • API & Gateway — Contract between frontend and backend

Group C — Logic & Data (Layers 8–11)

These layers define where business logic lives and how data is stored. This is where most one-way doors live — schema migrations and domain model changes are expensive to reverse.

CSM Pillar: Technical Implementation

  • Service Layer — Centralized business logic
  • Controllers / Routing — Thin controllers
  • Data & Persistence — Schema, indexes, migrations
  • Caching & Performance — Invalidation strategy, TTL

Group D — Integration & Async (Layers 12–14)

These layers define how external services connect and how background work gets done. The adapter pattern isolates vendor risk.

CSM Pillar: Technical Implementation

  • Integration / Adapters — Isolate vendor calls
  • Events & Async — Queue topology, idempotency, DLQ (Dead Letter Queue)
  • Middleware — Cross-cutting concerns

Group E — AI Layer (Layers 15–17, Conditional)

These layers are only applicable if your product includes AI features. They carry the highest risk in the entire sheet — this is where most AI pilots fail.

CSM Pillar: Technical Implementation

  • AI / RAG Pipeline — Retrieval, chunking, embeddings, guardrails
  • AI Agent Architecture — Tool use, autonomy, escalation
  • AI Output Validation — Multi-layer validation, bias detection

Group F — Quality & Security (Layers 18–21)

These layers define how you protect the system and verify correctness. AI introduces new attack surfaces that traditional security does not cover.

CSM Pillar: Governance & Risk

  • Security & Policy — Auth, encryption, RLS (Row-Level Security), audit logging
  • Validation & Integrity — Schema validation, sanitization
  • Testing Strategy — Unit, integration, e2e, AI evaluation
  • Code Review & Quality Gates — PR review, CI gates, AI code policy

Group G — Operations & Delivery (Layers 22–25)

These layers define how you ship, monitor, and govern the system in production. AI features need specialized monitoring that standard observability does not provide.

CSM Pillar: Operational Excellence

  • Observability — Logs, metrics, traces, AI monitoring
  • DevOps & Delivery — CI/CD, environments, rollback
  • Performance & Cost — Load testing, auto-scaling, AI cost limits
  • Ownership & Governance — ADRs, decision log, ownership matrix

Real-World Example: Building an AI-Powered Customer Support Tool

Here is how a 5-person team would use this sheet to build an AI-powered customer support tool from scratch:

Sprint 0 — Discovery

The PM selects the "AI-powered product" scenario. The team reviews Layers 1–4 (Foundation & Scope): they write a PRD with success metrics (deflection rate, CSAT, cost per ticket), document the domain model (tickets, users, knowledge base articles), choose Next.js + Postgres + pgvector, and decide to use OpenAI API (build-vs-buy decision). The architect writes ADR-001 (tech stack) and ADR-002 (AI vendor). The PM creates a sprint plan and staffing matrix.

Sprint 1 — Design

The team filters by Phase: Design. They address Layers 5–7 (UI, State, API), Layer 10 (Data — including vector schema for embeddings), Layer 15 (AI/RAG Pipeline — chunking strategy, embedding model, retrieval pipeline), and Layer 18 (Security — RLS for multi-tenant, audit logging). The AI Lead writes the AI pipeline design document. The architect reviews the reversibility column: domain model and schema are one-way doors — they get these right before building.

Sprint 2 — Build

The team filters by Phase: Build. They address Layers 8–9 (Services, Controllers), Layer 11 (Caching — with AI response caching + validation), Layer 12–13 (Integration with OpenAI via adapter pattern, async queue for embeddings), Layer 16 (Agent Architecture — tool use for ticket actions, escalation to human), Layer 17 (Output Validation — LLM-as-judge for response quality, content filtering), Layer 19 (Input validation), Layer 22 (Observability — AI-specific monitoring for drift, cost, hallucination rate). The team checks the AI Risk View to verify they have mitigations for every AI pitfall.

Sprint 3 — Launch

The team filters by Phase: Launch. They address Layer 20 (Testing — including AI evaluation suite for response quality), Layer 21 (Code Review — AI code policy), Layer 23 (DevOps — CI/CD with rollback, feature flags for gradual rollout), Layer 24 (Performance — load testing, AI cost limits with alerts). They run the Definition of Done checklist for every layer before launch. The PM exports the final sheet as CSV and shares it with stakeholders.

Operate — Ongoing

The team uses the CSM Pillar View in quarterly reviews. They check Operational Excellence (Layers 22–25) for drift, cost trends, and incident response readiness. They update the sheet when new tools are added, when the AI vendor changes models, or when new compliance requirements emerge. The sheet becomes the living record of their architecture decisions.

Key takeaway: The sheet is not a one-time exercise. It evolves with the project. The team uses it in sprint planning, architecture reviews, and quarterly audits. Every decision is documented, every risk has a mitigation, and every layer has a Definition of Done.

AI Development Risks: What Is Different

Building software with AI — whether AI is in the product or AI tools write the code — introduces risks that traditional architecture decision frameworks do not cover. This sheet addresses both dimensions:

AI in the Product (Layers 15–17)

If your product includes AI features, three additional layers apply. These are conditional — skip them only if your product has no AI components. The risks here are existential: hallucinations in production, runaway agents, compliance violations, and cost spirals.

AI Tools in the Development Process (All Layers)

Even if your product has no AI features, if your team uses AI code assistants (Copilot, Cursor, Claude), every layer carries additional risk. AI-generated code:

  • Looks correct but contains subtle bugs — handles the happy path but fails on edge cases, race conditions, and transaction boundaries.
  • Does not know your architecture decisions — generates code that violates existing ADRs because it has no context about your decisions.
  • Introduces security vulnerabilities — missing auth, over-permissive CORS, SQL injection, prompt injection, insecure crypto.
  • Generates tautological tests — coverage goes up while real validation stays flat. Tests pass but do not validate behavior.
  • Recommends deprecated or incompatible packages — libraries that are no longer maintained or that conflict with your existing stack.

Quality Gate Approach: Every layer includes an AI Quality Gate — a specific, testable checkpoint that must pass before AI-generated code or AI features are deployed. These gates are enforced in CI, not in code review alone.

Documents Every PM and Architect Needs

Each of the 25 layers specifies two document lists: Architect Documents and PM Documents. These are the concrete deliverables that prove a decision was made, reviewed, and approved.

Architect Documents

Technical artifacts: ADRs, design documents, diagrams, specifications. They capture what was decided and why. Examples: technology decision ADR, domain model document, security architecture, AI pipeline design.

PM Documents

Project management artifacts: sprint plans, review schedules, staffing plans, compliance timelines. They capture when and by whom. Examples: PRD with metrics, team skill matrix, pen test schedule, AI cost dashboard.

Rule of thumb: If a layer has no documents, it has not been decided. The documents are the evidence that the decision was made, reviewed, and is being maintained.

CSM Alignment: How This Sheet Maps to the Four Pillars

The Architecture Decision Master Sheet operationalizes the Cognitive Systems Management (CSM) framework — the methodology underlying the HAIEC platform. CSM bridges AI strategy, technical execution, and governance into a single operating model with four pillars:

Strategic Alignment

1–4 (Group A)

Business objectives, risk-reward, stakeholder alignment. What are we building and why?

Technical Implementation

5–17 (Groups B–E)

Architecture decisions, infrastructure, integration patterns. How are we building it?

Governance & Risk

18–21 (Group F)

Policy frameworks, compliance, continuous monitoring. How do we protect it?

Operational Excellence

22–25 (Group G)

Performance, incident response, continuous improvement. How do we run it?

Why CSM matters: Most architecture frameworks stop at technical decisions. CSM ensures that every technical decision is connected to business strategy, governance, and operations. The Architecture Decision Master Sheet is the practical tool that makes this connection explicit — every layer has a CSM pillar, and every pillar has concrete deliverables.

Definition of Done: Why It Matters

Each of the 25 layers includes a Definition of Done (DoD) — a specific, testable criterion that must be met before the layer is considered complete. The DoD is not a checklist item; it is a gate. Examples:

  • Product & Scope: PRD signed by Product Owner; success metrics defined and measurable; constraints documented; team briefed.
  • Security & Policy: Security architecture documented; RLS implemented; encryption verified; audit logging active; security scan passing in CI.
  • AI / RAG Pipeline: Pipeline tested end-to-end; evaluation suite passing; guardrails implemented; cost model validated; compliance reviewed.
  • Ownership & Governance: Governance framework documented; ADR process active; ownership matrix published; decision log maintained; review cadence scheduled.

The DoD prevents the most common failure mode in software projects: teams believe a layer is "done" when it is merely "started." The DoD makes the difference explicit and testable.

Common Anti-Patterns

Every layer includes an anti-pattern field — a description of what bad looks like. Here are the top anti-patterns across all 25 layers:

Layer 1: "Add AI" as a requirement

No success metrics, no failure criteria. Cannot tell if AI feature is working or hallucinating.

Layer 2: No documented invariants

Business logic scattered across layers. Silent data corruption.

Layer 3: Choosing tech because it is trendy

No rationale, no skill gap analysis. Hiring becomes impossible.

Layer 7: No API versioning

Breaking changes in production. Frontend and backend disagree.

Layer 15: No retrieval for knowledge tasks

Prompt-only approaches for tasks that require RAG. Hallucinations in production.

Layer 16: No escalation path for agents

Agents can call any tool, take unauthorized actions, loop infinitely.

Layer 17: No output validation

AI outputs go directly to users. Hallucinations, bias, harmful content reach users.

Layer 18: No RLS in multi-tenant

Shared credentials, no audit log. Data breaches, regulatory penalties.

Layer 25: No ADRs; decisions made in Slack

Nobody knows why decisions were made. New team members repeat mistakes.

How This Compares to Other Architecture Frameworks

Several architecture frameworks exist. This sheet is not a replacement for any of them — it is a complementary tool that operationalizes their principles into a trackable, layer-by-layer checklist with AI risks and PM deliverables.

FrameworkBest ForAI RisksPM DeliverablesDoD per Layer
This SheetAI-native projects, PMs + architectsYes (5 fields per layer)Yes (arch + PM docs)Yes
TOGAFEnterprise architecture, large orgsNoPartialNo
C4 ModelVisualizing architecture at 4 levelsNoNoNo
AWS Well-ArchitectedCloud infrastructure reviewPartial (ML lens)NoNo
arc42Documenting architecture decisionsNoNoNo

Key difference: This sheet is the only framework that pairs every layer with AI development risks (both AI-in-product and AI-tools-in-process), PM-specific deliverables, and a testable Definition of Done. It is designed to be used alongside TOGAF (for enterprise context), C4 (for visualization), and AWS Well-Architected (for cloud review) — not to replace them.

"We used this sheet to audit a stalled AI pilot and found 3 missing layers in 20 minutes — output validation, observability, and governance. The Definition of Done alone saved us from shipping another hallucination."

— Architecture review team, AI customer support pilot

Free CSM Framework Guide

Get the Cognitive Systems Management (CSM) framework guide — the four-pillar methodology underlying the HAIEC platform that bridges AI strategy, technical execution, and governance into a single operating model.

No spam. Unsubscribe anytime.

Frequently Asked Questions

What is an Architecture Decision Master Sheet?

An interactive reference covering 25 architecture layers that every software development project must consider — from product scope to governance. Each layer answers seven questions: Why does it matter? Who is responsible? How is it implemented? When should it be addressed? What are the risks? What does bad look like? And what documents should architects and PMs produce?

How does this sheet handle AI development risks?

Every layer includes AI development risks in two dimensions: AI in the product (Layers 15–17 cover RAG pipeline, agent architecture, and output validation) and AI tools in the development process (all 25 layers include AI pitfalls, risks of using AI to build, mitigation approaches, and quality gates). This means even if your product has no AI features, the sheet covers risks from using AI code assistants like Copilot, Cursor, or Claude.

What is the CSM pillar mapping and why does it matter?

Every layer maps to one of the four pillars of Cognitive Systems Management (CSM): Strategic Alignment (Layers 1–4), Technical Implementation (Layers 5–17), Governance & Risk (Layers 18–21), and Operational Excellence (Layers 22–25). CSM is the framework underlying the HAIEC platform. The mapping ensures every technical decision is connected to business strategy, governance, and operations.

What documents does each layer specify?

Each layer specifies two document lists: Architect Documents (ADRs, design documents, diagrams, specifications) and PM Documents (sprint plans, review schedules, staffing plans, compliance timelines). These are the concrete deliverables that prove a decision was made, reviewed, and approved. If a layer has no documents, it has not been decided.

Can I use this sheet for projects without AI features?

Yes. The sheet includes a "Web app without AI" scenario that filters to Layers 1–14 and 18–25, skipping the conditional AI layers (Group E). However, if your team uses AI code assistants (Copilot, Cursor, Claude), the AI development risks on all layers still apply — they cover risks from using AI tools to write code, not just AI features in the product.

How is the Definition of Done (DoD) different from a checklist item?

The DoD is a gate, not a checkbox. It is a specific, testable criterion that must be met before a layer is considered complete. For example, the DoD for Security & Policy is: "Security architecture documented; RLS implemented; encryption verified; audit logging active; security scan passing in CI." This prevents the common failure mode where teams believe a layer is "done" when it is merely "started."

Does the tool save my progress?

Yes. When you assign owners, statuses, sprints, or ADR references in the PM Tracking tab, your edits are saved automatically to your browser via localStorage. Your data persists across sessions on the same browser. Use the Export CSV or JSON buttons to share your tracking data with your team, or the Reset button to start fresh.

Need Help With Your Architecture Decisions?

Get a tailored architecture review and decision framework — from Subodh KC, co-founder of the HAIEC AI security and compliance engine.

Let's Talk →