AI pilots that don't ship to production are just expensive demos. Here's the 6-week playbook we use to keep them honest.
Week 1 — KPI discovery
Pin down the metric. Pin down the owner. If you can't name both, you're not ready to start.
Week 2 — Feasibility
Data audit, model selection, guardrails design, cost envelope. By the end of week 2 you should know whether this is a 6-week pilot or a longer programme.
Week 3 — Prototype
Working prototype against real data, tested with a small user set. Not a slide. Not a Loom. A thing they can use.
Week 4 — Evaluation
Accuracy, safety, cost evals against acceptance criteria. If it doesn't pass, kill it now — not after another four weeks of polishing.
Week 5 — Productionise
Integrate into Fusion / OIC / OAC, wire logging and observability, roll out to target users.
Week 6 — Measure and iterate
Track the KPI weekly, tune prompts and policies, compound wins. AI is never 'shipped and done'.