Scoping an AI listing optimizer for a short-term rental company. The first good decision was a smaller v1.
The client's analysts kept listings competitive by hand: photo order, titles, descriptions, all checked against how each listing ranked in search. It was weekly work that scaled with headcount. The client wanted AI to take on the audit while their people kept the final say.
The analysts who own each listing come first. They would approve every change, so the tool had to make their week shorter, and it could not hand them a new queue to babysit. Then the founder, who was buying it and knew his operation better than any spec could.
I started with a broader v1 that included pricing changes. The client corrected it by email. I rewrote the spec around his version, and it was better for it.
| Decision | What I chose | What it cost |
|---|---|---|
| v1 scope | Listing content only: weekly audits plus approved updates to photo order, title, and description. | Pricing, the flashier feature, waits for v2. Listing edits are easy to review and undo. Price changes hit revenue the same day. |
| Photo ranking | A vision model scores each photo. | No API exposes per-photo click data, so the score is a proxy until real engagement data exists. |
| Write accessHardest call | Approve first. An analyst signs off before anything is pushed to a live listing. | Slower than auto-push. The client can turn on auto-push later, once the approvals earn it. |
| Pricing data | Read-only signal in v1. | Pricing context can shape a suggestion but cannot change a price. |
Here, the analyst's approval is the test. The design question was how to make that review fast enough that analysts would actually do it.
In two weeks on a fixed fee, the sprint turned a broad brief into a v1 the client could approve: the scope, the three data sources it reads, an approve-first workflow, and the access list the build needed.
Measure approval rate by change type. When analysts approve one category of edit almost every time, that category has earned auto-push, and only that one. In parallel I'd find a way to capture real engagement data, so the photo score can graduate from a proxy to a measured result.
The spec says in plain words that the photo score is a guess. I'd rather a client read that in the spec than find it out in month two.