ai-agent-poc-and-pilot-scoping

name	ai-agent-poc-and-pilot-scoping
description	Use when scoping an AI-agent proof of concept or pilot with shadow, supervised, and agentic stages, measurable gates, abort conditions, and production-exit criteria.
metadata	{"portable":true,"compatible_with":["claude-code","codex"]}

AI-Agent POC and Pilot Scoping

Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, [phone].

Use When

• The buyer wants a pilot or POC before committing to a full agentic build or rollout.
• The agency proposes a paid agent feasibility study before pricing the full engagement.
• The pilot must produce evidence the buyer's CTO, CISO, head of operations, and regulator can act on.
• The buyer has been burned by agent demos that "looked autonomous" but failed on integration, intervention rate, or irreversible incidents.

Do Not Use When

• The POC is for a non-agent AI feature (use ai-on-saas-poc-and-pilot-scoping).
• The POC is for a copilot with no action-taking (use ai-on-saas-poc-and-pilot-scoping).

Domain Inputs

• The candidate agent use case (named workflow, named outcome, named volume).
• The action catalogue v0 with reversibility classification.
• The autonomy-level target per action class.
• The success-metric definition (task-success rate / intervention rate / deflection / time-to-resolution) with floor and ceiling.
• The golden dataset assembly plan or the live-traffic shadow plan.
• The kill-switch architecture and the human operator on call.
• The time-box the buyer accepts (typically 8–16 weeks for agent pilots).