Case Study

From 1.5 Million Photos to Shelf Intelligence

By Jeff Czischke, Co-Founder & CEO, Fulcrum AI Labs · September 3, 2026

A Fortune 500 consumer packaged goods company had a data problem hiding inside a photo problem.

Their field reps walked thousands of stores every month and photographed the shelves: product facings, price tags, competitor placement, out-of-stocks, promotional compliance. The intent was good. Those photos are the ground truth of how a brand actually shows up at the point of sale. The reality was that 1.5 million images a month were landing in a folder that no one could read at that scale.

A small team of analysts sampled what they could. It took more than forty hours a week to review a fraction of the volume, and by the time a pattern surfaced, the shelf had already changed. The company was sitting on the single richest signal it had about its own retail execution and could not turn it into a decision.

The real constraint

The interesting part was not "use AI on images." Everyone in the building already knew computer vision could read a shelf. The constraint was everything around the model: getting 1.5 million images a month through a pipeline reliably, keeping accuracy high enough that merchandising teams would trust the output, and doing it inside the company's own cloud without a year-long platform project first.

In other words, the gap was not capability. It was implementation.

What we built

We embedded a pod with the team that owned the problem and scoped one workflow tightly: shelf photo in, structured shelf intelligence out. The system does four things on every image:

• Detects individual products and counts facings
• Reads price tags and labels
• Measures share of shelf against the category
• Flags compliance issues, like a planogram violation or a missing promotion

It runs in the client's environment, on their cloud, with logging at each step so the merchandising team can see why the system called what it called. That last point mattered more than the model architecture. People adopt a system they can question and route around one they cannot.

Prototype to production in 11 weeks

We built a working proof of concept on real images in the first weeks, not a slide about one. Once the client's team could see it read their own shelves, the conversation moved from "will this work" to "how fast can we widen it." From there it was integration, hardening, and the unglamorous work that decides whether a pilot ever becomes a system: error handling on bad images, throughput at full volume, and a review loop for the edge cases.

Within 90 days the system was processing the full 1.5 million images a month at 94% accuracy, replacing a manual review process that had cost more than forty analyst hours a week.

The analysts did not lose their jobs. They stopped spending their week sampling images and started spending it acting on what the whole set was telling them.

The pattern underneath

This engagement looks like a computer-vision story. It is really a delivery story, and the shape of it repeats across almost everything we ship.

The value was never locked behind a smarter model. It was locked behind execution: moving a capable idea into a governed, owned production system without a platform project in front of it. We scoped one bounded workflow, built the real thing early, kept the client's people at the decision points, and handed over something their team runs.

If your organization is sitting on a signal it cannot yet act on, the question worth asking is not whether AI can read it. It almost certainly can. The question is who is going to get it into production, and whether you will own it when they are done.