Problem
A business-support unit of a large retailer (lawyers, procurement, finance) gets about a hundred requests to "make us an AI". Some are not AI tasks at all, and a small team cannot afford months of specification and approval for a case that may fail. The speaker's concern is also adoption: a tool can be built well and still not be used. A long cycle makes both failures expensive.
Example from the article: a request to automate refunds for double card charges (no receipt printed) needed recognition of handwritten applications. A two-day PoC showed 55% accuracy and a one-FTE effect, so the team advised moving to an online form and built nothing.
Solution in brief
Filter requests by hard metrics and effect in FTE before starting. Pass five stages (PoC, demo, test environment, pilot, production) with a go/no-go at each, so cheap bets die in days. Deploy from a ready starter template on a pre-approved architecture. Decide on production by a four-group pilot metric profile: if one group sags, the solution does not go to production.
Questions this case answers
- How to decide quickly and cheaply whether an LLM-agent use case is worth building before committing to a full production process
- How to set go/no-go gates for an AI pilot with a multi-part metric profile where one weak part blocks the move to production
- How to decide quickly whether an AI-agent use case is worth building before investing in it
- How to take an LLM agent from idea to production in weeks with hard go/no-go gates
Tools and methods
- Intake filter — refusal of non-AI cases, hard target metrics, FTE estimate.
- Five-stage funnel — PoC, demo, test environment, pilot, production, each with a go/no-go.
- Pilot metric profile — four groups: convenience, quality, load, adoption.
- Starter template and AI coding assistants — Claude Code, Cursor, Codex.
Google or a one-time email code. No password.
