Problem
A team that builds ML for a client or another business unit often gets a ready-made solution as the request, not a problem. Success is measured differently by the customer, the manager and the team. The model then reaches no production, or ships and nothing happens, because nobody agreed on the pain, the metric or the owner's motives.
Example from the article: a customer asked for "ChatGPT that sorts all incoming complaints and writes the answers". Tens of thousands of complaints turned out to fall into about 10 types with 10 typical answers, so a simple classifier and standard replies were enough.
Solution in brief
Treat the first weeks as discovery: find the real pain, agree in writing how success is measured, and learn the sponsor's own KPI. Then build trust with visible intermediate results, start from a working baseline, and show the model through a small playground instead of a raw test build.
Questions this case answers
- How to turn a stakeholder's requested solution into a problem with agreed success metrics before any model is built
- How to build trust in a model that a non-expert client cannot evaluate, using a working baseline and a hands-on demo
- How to turn a customer's request for an ML solution into a problem with agreed success metrics
- How to run a client ML project from first request to MVP so the customer adopts it
Tools and methods
- Problem discovery with the customer — pain, success metric and manager KPI written down before development
- Baseline-first modeling — popular items or heuristics first, then compare newer models to them
- Demo playground — a simple web service where the client sets inputs and sees model output
- Miro roadmap — parallel tracks with sprint goals and tangible results
- PoC → MVP → scaling — the author's staged path, with an explicit decision whether to scale
Google or a one-time email code. No password.