AGENTIC ORG CHART STARTER KIT Goal Make AI-enabled work dependable by assigning decision rights, verification levels, and required evidence. Use this for any workflow where an agent drafts, recommends, or takes actions. 1) WORKFLOW INVENTORY (list 10–20 workflows) For each workflow, capture: - Name (example: “Agent-drafted PRs for Billing service”) - Output type: (a) production code, (b) customer-facing text, (c) finance/reporting, (d) legal/policy, (e) internal decision memo, (f) ops/runbook - Frequency/week - Blast radius if wrong: Low / Medium / High / Regulated - Baseline metric: cycle time, CSAT, rollback rate, escalation rate, defect rate, etc. 2) ASSIGN A VERIFICATION LEVEL (L0–L4) L0 Draft-only: brainstorming; nothing shipped. L1 Human spot-check: internal docs, low-risk internal comms. L2 Test + review: production code, runbooks; requires automated checks + codeowner. L3 Policy + audit trail: customer comms, finance; requires citations + policy checks + retained logs. L4 Regulated approval: legal terms, regulated data workflows; requires legal/compliance sign-off + retention controls. 3) DECISION RIGHTS (ownership follows blast radius) For each workflow, define: - Approver of record (name + role) - Backup approver - Stop-ship authority (who can block a merge, send, or release) - Escalation path (who gets paged/notified) 4) QUALITY GATES (pick a small set per high-risk workflow) Examples: - Citation requirement: key claims must link to a policy doc, ticket, dashboard, or source-of-truth. - CI gate: tests + static analysis must pass before merge. - Sensitive-topic routing: pricing/refunds/SLA/security language triggers approval. - Tool allowlist: the agent may only call approved tools/APIs. - Change windows: model, prompt, or tool updates happen on a schedule with rollback. 5) AUDIT & LOGGING (minimum for L2–L4) Log fields: - Timestamp, agent name/version, requester, workflow name - Inputs (ticket ID, repo, doc IDs), tool calls, retrieved sources - Output hash or stored output, verification level, approver identity Retention guideline: - L2: short retention that still supports investigations - L3: longer retention aligned to customer and finance needs - L4: retention aligned to legal/compliance requirements 6) METRICS (weekly review) Pick 4–6 metrics per workflow: - Speed: cycle time, time-to-first-draft - Quality: defect rate, rollback rate, policy violations - Customer impact: CSAT/NPS movement, complaint rate - Risk: security findings, escalation rate Define thresholds that trigger action (example: a noticeable week-over-week spike in escalations). 7) 90-DAY ROLLOUT CHECKLIST Weeks 1–2: Pick 2 workflows; set baselines; name approvers; define gates. Weeks 3–6: Implement logging + gates; run a limited pilot; review metrics weekly. Weeks 7–10: Add a few more workflows; create a small golden task suite to detect regressions. Weeks 11–13: Run an incident drill; publish an “AI decision rights” doc; plan next quarter’s expansion. Definition of Done For any AI-enabled workflow, anyone can answer quickly: 1) Who approves it? 2) What evidence is required? 3) Where is the audit trail? 4) What metric tells us it’s working—or drifting?