When delivery photos need address reading, not more reviewers

The job is: a driver takes a phone photo at the door, and ops needs to know whether the parcel is at the address on the ticket.

People often ask for “vision AI.” The useful question is narrower. Can any single reader survive the photos you already have — night, angle, compression, a sticker over half the plate — and can the output be checked against your own address list?

What we look at before we talk models

We ask for a sample of real captures, not a brochure shot. We ask what the reviewer currently does when the text is wrong: rewrite, call the driver, mark delivered anyway. We ask whether you already store the expected address in a form you can query.

If you cannot match a string to a record, a better reader only produces a prettier unread string. That is an ops and data job first.

If matching exists and reviewers are the bottleneck, an ensemble can be worth it. The FleetOptics address detection case study is that job: more than one reader, a vote, then a check against company records. We published the accuracy we actually got. It is not a promise that every fleet will hit the same number.

Kuwait delivery work has its own address shape (area, block, street). That does not change the consultancy sequence. It does change what “match” means. The Manzil delivery platform is a different job — dispatch and proof-of-delivery as a product — and we will not collapse the two on a first call.

What a useful next step looks like

A two-week slice is usually: score your current photos with one reader, measure miss types, then decide if a second reader plus a match API is cheaper than the VA hours. Sometimes the answer is keep the reviewers and fix capture (lighting, a crop box in the driver app).

That decision is AI consultancy, not a model shopping list.

If this is the loop you are in, book thirty minutes or write hello@jamilglobal.com. Bring twenty photos and the address fields you already store.