AI Visibility Checklist: Verify Out-of-Stock Before Deployment

07/10/2026

The Warehouse Reality and the Out-of-Stock Challenge

I once sat in a conference room at a major FMCG company in Ho Chi Minh City. The field sales manager was presenting the weekly report. On the screen were hundreds of photos taken from convenience stores. The issue wasn't a lack of data; it was that the data didn't reflect the truth. He admitted that about one-third of the "out-of-stock" reports were actually due to employees taking photos from the wrong angle or shelves being obscured by other products.

That was when the concept of AI Visibility began to be discussed seriously. Many people think that simply installing cameras or using a photo app is enough. But reality shows that the camera is just the input. Artificial intelligence is what turns pixels into business decisions. If the system isn't "sharp" enough to distinguish a tilted milk carton from an empty shelf slot, all efforts in display scoring are meaningless.

What Everyone Says vs. What Reality Shows

There is a major contradiction in how businesses approach this technology. Everyone claims AI accuracy is 99%. But operational reality tells a different story.

The most common focus is on product detection. Everyone wants to know "Is product X still available?" But what reality shows is that the biggest challenge lies in determining "Does that position exist in the planogram?" An empty shelf slot could be due to out-of-stock, but it could also be because the store hasn't been allocated that product, or it's a special display position not in the standard plan.

If AI simply counts the number of products and compares it to the expected quantity, you will receive thousands of false alerts. This causes alert fatigue for the monitoring team. They will start ignoring the alerts, and the system becomes useless. Therefore, AI Visibility is not just about seeing, but about understanding the context of the point of sale.

5-Step Checklist Before Deployment

Before signing a contract with any provider, use this checklist to audit your current platform. These are not theoretical questions, but practical bottlenecks I often encounter when consulting partners in Vietnam and neighboring countries like Thailand and the Philippines.

1. Identify the input data source: Where does the system receive images from? From fixed IP cameras or from employees' mobile phones? Each source has its own set of errors. Fixed cameras provide a consistent view but are hard to adjust for seasonality. Phones are flexible but shooting angles and lighting change constantly. You need to know if the AI system can compensate for these deviations.
2. Handling Occlusion: In reality, shelves are rarely empty. A water bottle might obscure half a biscuit box. If the AI doesn't have a reasoning mechanism to estimate obscured products, it will report inaccurate inventory. Request a demo with messy shelf scenarios, not perfect model shelves.
3. Detecting Fake and Duplicate Images: This is a common "cheating" issue in traditional sales channels. Employees might re-upload old photos or use images from the internet. A reliable AI Visibility system must have mechanisms to detect metadata, screenshots, or visual duplicates. Otherwise, the output data will be poisoned at the source.
4. Latency and Throughput: How many points of sale do you have? If it's thousands of stores, real-time or near-real-time image processing is mandatory. A system that takes 10 minutes to process one image won't help with same-day replenishment. Check the average processing time per image.
5. Score Transparency (Explainability): When AI scores a shelf as "failing," it must point out exactly which position and why. A general number is not enough. You need a dashboard displaying each SKU and each deviation. This helps the sales team take specific actions, rather than just knowing they are doing something wrong.

Quick Answers

Can AI Visibility completely replace field inspectors?

No. AI Visibility is a monitoring and anomaly detection tool, helping employees focus on problematic points of sale. It cannot replace humans in rearranging goods or negotiating with store owners. It is a third eye, not a hand.

How often does the AI model need to be updated?

It depends on the rate of change in the product catalog and shelf design. If you launch new products frequently, the model needs periodic retraining to recognize new packaging. A good system should have a continuous learning mechanism that doesn't require stopping the system.

What factors determine the deployment cost of AI Visibility?

Mainly the number of points of sale, photo frequency, and product catalog complexity. A system processing millions of images daily with thousands of SKUs will have a different infrastructure and cost structure compared to a system monitoring only a few hundred key stores.

There is more here than one article can hold. Keep reading on the AIVISION blog, look at our display scoring solution, or get in touch.