ICMD Agent Readiness & Rollout Framework (2026)

Use this framework to move from an agent demo to a production workflow with measurable ROI and controlled risk.

1) Define the Unit of Work (UoW)
- Name the unit: ticket resolved, invoice discrepancy closed, vendor onboarded, lead packet generated.
- Confirm volume: target at least 5,000 UoW/month for meaningful automation ROI.
- Baseline economics: current cost per UoW (labor + tools + outsourcing). Write it down.
- Success metrics: containment rate, assist rate, escalation rate, rework rate, and cycle time.

2) Map the Workflow Boundary (Execution Envelope)
- List allowed tools (read + write) and explicitly ban everything else.
- Set budgets: max steps, max tool calls, and max $ compute per UoW.
- Define risk tiers:
  Tier A (safe): read-only, drafting, classification.
  Tier B (moderate): creating records, updating CRM fields with validation.
  Tier C (high): sending emails, refunds, payment changes—requires approval.

3) Build Governance First (So You Can Automate Later)
- Identity: SSO/SAML, RBAC, service accounts per connector, SCIM if enterprise.
- Audit: log prompt, retrieval sources, tool calls, outputs, and approvals.
- Data: retention policy, redaction in logs, tenant isolation for vector stores.
- Security: prompt-injection rules (treat all external text as untrusted input).

4) Create an Evaluation Suite
- Collect 100–300 “golden tasks” from real production history.
- Define pass/fail criteria per task (structured output checks, policy compliance, citations).
- Add regression gates in CI: no release if containment or policy compliance drops.

5) Pilot Plan (2–4 weeks)
- Start read-only mode for 1 week: draft outputs + recommendations.
- Add human approvals for writes in week 2.
- Track daily: containment, assist, escalation, rework, time-to-resolution, and reviewer time.
- Set go/no-go thresholds (example): rework < 3%, policy violations = 0, containment > 30%.

6) Production Rollout
- Expand volume gradually: 5% → 25% → 50% → 80% of queue.
- Introduce sampling-based QA (e.g., review 10% of Tier A, 30% of Tier B, 100% of Tier C).
- Implement incident response: severity levels, rollback plan, and customer comms templates.

7) Pricing & Unit Economics Check
- Price to the UoW (per resolved ticket, per invoice) with minimum commitments.
- Target gross margin: >70% by keeping compute + tool COGS bounded.
- Monitor weekly: $ compute per UoW, tool failure rates, and review labor minutes per UoW.

8) Expansion Strategy
- Expand only after you can prove before/after metrics (cycle time, backlog, SLA adherence).
- Move to adjacent workflows that share the same systems of record and policies.
- Graduate autonomy: unlock Tier B writes after quality thresholds, then Tier C with approvals.

If you can’t measure containment, rework, and cost per unit of work by week two, you’re not running an agent product—you’re running a demo.