AI SUPPLY CHAIN LEADERSHIP CHECKLIST (2026)

Use this to turn “we’re using AI” into accountable operations. Fill it out for ONE AI feature first.

1) AI BILL OF MATERIALS (AI BOM) — one page per feature
- Feature name + customer surface:
- Owner (single accountable person):
- Model provider + endpoint (OpenAI / Anthropic / Azure OpenAI / Bedrock / Vertex AI / self-host):
- Model name + versioning approach (pinned version/date? rolling?):
- Routing rules (single model, fallback model, cost/latency routing):
- Tools/function calls used (search, DB query, ticket creation, code execution, etc.):
- Prompt inputs (list concrete fields):
- Data classes allowed (customer content, PII, secrets, regulated data). Include explicit “NOT allowed”:
- Logging: what is stored (prompts/outputs/traces), where, retention, and access control:
- Vendor data handling: training usage yes/no (per provider setting/contract) and retention settings:
- Human override path: escalation triggers, manual review queue, rollback switch:

2) MODEL CHANGE CONTROL — release gate (minimum viable)
- What changes require a gate? (model swap, routing change, prompt/system change, tool schema change)
- Evaluation suite location (repo/path) + golden set definition:
- Safety/abuse checks included (policy violations, PII leakage, prompt injection attempts):
- Deployment method (feature flag/canary) + rollback procedure:
- Approval: who signs off (engineering + product; security/legal if data classes change):

3) COST & RELIABILITY GUARDRAILS
- Spend limit and alerting owner:
- Rate limits and retry policy:
- Timeout policy and fallback behavior (degraded mode):
- Monitoring signals (customer complaints, eval regressions, tool errors, anomaly detection):

4) FIRST TABLETOP INCIDENT DRILL (60 minutes)
Scenario: AI output causes customer harm OR data exposure OR a high-profile wrong answer.
- Who declares an AI incident?
- Who is on-call / primary + backup?
- What’s the immediate stop button? (disable feature flag, force fallback, block tool calls)
- What evidence do you collect? (prompt/output logs, traces, model version, routing decision)
- Who communicates to customers? (name the role)
- Post-incident: what gets written down? (root cause, contributing factors, permanent fixes)

Definition you can paste into your on-call docs:
An AI incident is any model-driven behavior that risks customer harm, legal exposure, brand damage, security compromise, or uncontrolled cost. If it’s customer-facing, it’s production.