Week 1: RAI Foundations for Architects
Week Overview
Objective: Build the mental model for RAI as an architectural discipline — not a compliance checkbox applied at the end.
Time Estimate: 8–10 hours
Deliverable: RAI Architecture Lens — a 1-page reference card mapping each RAI principle to design-time questions
The Core Shift
Traditional software architecture asks: What should this system do, and how fast?
RAI-aware architecture adds: Who could this system harm, how, and what design decision prevents that?
Framework 1: Microsoft RAI Standard v2 — The 6 Principles as Design Constraints
| Principle | What It Eliminates | Design-Time Question |
|---|---|---|
| Fairness | Systems that produce disparate outcomes across demographic groups | What populations does this system serve? How is disparate impact measured? |
| Reliability & Safety | Systems that fail unpredictably or cause harm when they fail | What happens when the model returns a wrong answer? Is there a fallback? |
| Privacy & Security | Systems that expose personal data beyond its intended use | What data does this system access? What is the minimum necessary scope? |
| Inclusiveness | Systems that exclude users or produce lower quality for edge groups | Which user populations are underrepresented in training data? |
| Transparency | Systems where users cannot understand what is AI-generated | Is it clear to users when they are interacting with AI? |
| Accountability | Systems where no human owns the outcome of an AI decision | Who reviews high-stakes outputs? Who owns the audit trail? |
Framework 2: NIST AI RMF 1.0 — Govern, Map, Measure, Manage
GOVERN → Establish policies, roles, culture
↓
MAP → Identify context, stakeholders, and risks for a specific AI system
↓ ← YOU DESIGN FOR THIS
MEASURE → Analyze and assess identified risks (metrics, evaluations, red teaming)
↓ ← AND THIS
MANAGE → Prioritize, respond, and monitor risks continuously
What "designing for MEASURE" means:
- Build in evaluation hooks — the system must be independently testable
- Choose architectures where component outputs are observable
- Design for A/B testing of safety mitigations
What "designing for MANAGE" means:
- Every AI component needs a monitoring owner in the architecture diagram
- Alerting thresholds must be defined at design time, not after launch
- Audit logs must be in the architecture — incident response requires them
Framework 3: EU AI Act Risk Tiers — Architecture Decisions Change by Tier
UNACCEPTABLE RISK — BANNED (Feb 2025)
├─ Social scoring | Real-time biometric surveillance | Subliminal manipulation
│ → Architecture decision: Do not build.
HIGH RISK — Strict obligations (Aug 2026 deadline for deployed systems)
├─ HR/hiring | Credit scoring | Medical devices | Critical infrastructure
│ → Required: human oversight mechanism, audit logging, accuracy testing,
│ data governance documentation, EU AI database registration
LIMITED RISK — Transparency only
├─ Chatbots | Deepfakes
│ → Required: AI disclosure to users
MINIMAL RISK — No obligations
└─ Spam filters | Recommendation engines
→ Best practice: document anyway for future re-classification
This Week's Resources
| Resource | Type | Estimated Time |
|---|---|---|
| Microsoft RAI Principles | Reading | 1 hour |
| NIST AI RMF 1.0 — Govern + Map sections | Reading | 2 hours |
| Azure Well-Architected Framework — AI Workloads | Reading | 1.5 hours |
| Microsoft Learn: Responsible AI principles in practice | Course | 2 hours |
| OWASP LLM Top 10 | Reference scan | 1 hour |
Hands-On Exercise
Exercise — Gap Analysis Against NIST AI RMF
Take an AI system you have previously designed or reviewed. Map it against the four NIST AI RMF functions:
- Govern: Is there a documented policy for this system? Who owns it?
- Map: Are all stakeholders identified — including adversarial actors and autonomous agents?
- Measure: Can the system's fairness, safety, and accuracy be measured independently? Are evaluation hooks built in?
- Manage: Are alerting thresholds defined? Is there an incident response plan?
For each function, score: Strong / Partial / Missing. Write a 1-paragraph explanation for any Missing score.
Week 1 Deliverable: RAI Architecture Lens
Build a 1-page reference card you will use for every architecture review going forward:
| RAI Principle | Design-Time Question | Architecture Smell (what absence looks like) |
|---|---|---|
| Fairness | What populations does this serve? How is disparate impact measured? | No evaluation dataset; no bias testing |
| Reliability & Safety | What happens when the model is wrong? Is there a fallback? | No fallback; no monitoring owner |
| Privacy & Security | What is the minimum data scope? Is consent modeled? | Overly broad data access; no consent flow |
| Inclusiveness | Who is underrepresented? How is accessibility handled? | No accessibility review; monoculture test data |
| Transparency | Is AI disclosed to users? Can decisions be explained? | No AI disclosure; no explainability hook |
| Accountability | Who owns each AI decision? What is the escalation path? | No human review gate; no audit log |
Knowledge Check
- What is the difference between RAI as a compliance checklist and RAI as an architectural constraint?
- Which NIST AI RMF function is most directly influenced by architecture decisions — and why?
- Under the EU AI Act, what architecture components are legally required for a High-Risk AI system?
- An architect says: "We'll add the safety filters after the MVP ships." Which RAI principle is most at risk?