For a company adopting agentic AI for the first time. Phased — do not try to ship Phase 4 in month 1.
Phase 1 — Foundation (weeks 1–6)
- Confirm readiness gates (§5). Address gaps before going further.
- Appoint CoE Lead with executive sponsorship and block-deployment authority.
- Define approved stack (§16).
- Stand up registry + intake form (§12, §15).
- Write Agent Card template (§14).
- Define risk classification (§10).
- Define responsible-AI checklist (§18).
- Identify 2–3 candidate pilot workflows.
- Target outcome: the scaffolding exists. No agents in production yet.
Phase 2 — First pilots (weeks 7–18)
- Ship 2–3 high-impact agents end-to-end.
- Each agent uses the registry, Agent Card, risk classification, observability from day one.
- Each agent has a real KPI, real ROI numbers, real audit trail.
- Vendor-embedded AI catalog completed (§15.2).
- First procurement-integration review (§26).
- Lessons fed back into the framework.
- Target outcome: 2–3 governed agents in production. Framework validated by real use.
Phase 3 — Internal platform (months 5–9)
- Reusable building blocks (prompt templates, agent blueprints, RAG patterns).
- Centralized monitoring dashboards.
- AI gateway for token caps, rate limits, central policy enforcement (§28).
- Runtime guardrails audited end-to-end (§19, §20).
- Department champions empowered to build with shared blueprints.
- CoE begins shifting from Centralized → Hybrid.
- Target outcome: 5–15 governed agents in production. Self-service path for low-risk agents.
Phase 4 — Scaled ecosystem (months 9–18)
- Cross-functional agents (Sales × Finance × Ops).
- Stage-3 autonomous workflows where ROI and safety justify.
- Executive dashboard for the AI portfolio.
- Formal compliance alignment (NIST AI RMF or ISO 42001 readiness).
- CoE moves toward Advisory (§6.3 Stage 3).
- Target outcome: Agentic AI is a managed enterprise capability, not a series of pilots.