Scoping a proof of concept that leads to production
One use case, one success metric agreed up front, a real dataset, and a written path to production before you start.
OpenAI's Applied AI post is blunt: success is "production systems, sustained adoption, and meaningful customer impact—not simply activity or successful demonstrations" (official, Sources O7). A POC is a step toward that, not an end.
The one-page POC charter
- Use case and users: one workflow, named users.
- Success criteria: a metric, a baseline, a threshold, measured on an agreed eval set.
- Data: which data, how it is accessed, its classification and who approved it.
- Out of scope: written down, so scope creep is visible.
- Timeline: two to six weeks, with a mid-point review.
- Path to production: who owns it after, what reviews remain, how it will be run and paid for.
- Decision meeting: date and attendees, booked on day one.
Common traps
- A POC on synthetic data that says nothing about the real data.
- No baseline, so "it seems good" is the only result.
- Building the general platform instead of the one use case.
- No owner on the customer side for after the POC.
Practice
Written for this site to drill the lesson. Not reported interview questions.
- Write the charter for a POC that uses an LLM to draft answers to inbound RFP questions.
- Halfway through, the customer asks to add a second use case. What do you say?
Check yourself
What three things must success criteria include?
A metric, a baseline and a threshold, measured on an agreed dataset.
When do you book the decision meeting?
On day one.
Sources
- Official OpenAI Applied AI Engineer post
Saved in this browser.