ICMD 2026 AGENT READINESS FRAMEWORK (Wedge → Safety → ROI)

Use this framework to validate an AI agent startup idea, ship a production-grade v1, and scale without losing trust or margins.

STAGE 1 — WEDGE DEFINITION (1–3 days)
1) Workflow: Name one queue you will reduce (e.g., “Tier-1 password reset tickets in Jira”).
2) Persona: Identify the operational owner (e.g., Head of IT Ops) and the budget owner.
3) System of Record: Pick ONE SoR to go deep on first (Salesforce, ServiceNow, NetSuite, Jira, HubSpot, Zendesk).
4) KPI target: Write a 30-day measurable goal (e.g., “reduce time-to-first-action by 40%” or “deflect 25% of Tier-1 tickets”).
5) Risk boundary: List actions you will NOT automate in v1 (e.g., payments, production deploys, account deletions).

STAGE 2 — SAFETY + GOVERNANCE (week 1–2)
1) Permissions: Implement least-privilege scopes per tool; no long-lived high-priv tokens.
2) Human controls: Define approval rules by action type (draft vs execute). Add a “pause automation” switch.
3) Audit logs: Every run must capture inputs, outputs, tool calls, policy checks, and who approved.
4) Data handling: Define retention (0–365 days), redaction for PII, and export/delete capabilities.
5) Failure modes: Document top 10 failures and fallbacks (retry, alternate tool, escalate to human).

STAGE 3 — QUALITY SYSTEM (week 2–6)
1) Golden set: Build 200–1,000 labeled historical examples from the customer SoR.
2) Eval harness: Run offline regression weekly. Track success rate, policy violations, and tool failure rate.
3) Observability: Add trace IDs, structured logs, and dashboards for success/fail/needs-human.
4) Launch sequence: 2–4 weeks in recommendation mode, then graduate to limited autonomy.
5) Release gate: No production changes unless the golden-set score stays above your threshold (e.g., 95%).

STAGE 4 — PRICING + ROI PROOF (week 4–8)
1) Baseline fee: Set a platform fee that covers support/security (typical: $1.5k–$10k/month).
2) Usage unit: Price per workflow unit (ticket resolved, invoice exception processed), not per token.
3) Cost controls: Add quotas, alerts, and caps; optimize with caching and small-model-first routing.
4) ROI dashboard: Report hours saved, queue reduction, cycle time, and error/override rates.
5) Expansion plan: Define 2 adjacent workflows inside the same SoR to upsell after trust is earned.

If you can’t clearly pass Stage 1 (wedge + KPI + SoR), don’t build. If you can’t pass Stage 2 (governance), don’t scale. If you can’t pass Stage 3 (eval discipline), don’t promise autonomy. If you can pass Stage 4 (ROI proof + margins), you have a real business—not a demo.