AI Display Scoring: Three Approaches and the Real Cost
14/10/2026
Managers often expect that simply installing the software, taking a few photos, and receiving a perfectly accurate scorecard is all it takes. They believe AI Vision will completely replace humans, operating flawlessly and immediately. Reality is often very different. Messy input data, poor lighting, or constantly changing planograms leave the system blind. The result is a report riddled with errors, causing sales teams to lose trust and revert to old manual methods.
Choosing the Wrong Image Data Collection Method
What goes wrong? It starts at the data source. Many businesses allow employees to freely take photos from personal phones. Angles, sharpness, and lighting vary from person to person. Why is this difficult? AI requires consistency for accurate recognition. If photos are taken at skewed angles, the system struggles to distinguish one SKU from another. How to do it right? Standardize the shooting process. Mandate the use of a mobile app with alignment frames, automatic brightness checks, and GPS location confirmation. How to measure it? The percentage of images rejected due to poor quality. If this number exceeds 10%, the process is failing.
Misjudging Display Scoring Criteria
Many people think display scoring is simply about counting products. But reality is more complex. A fully stocked shelf that is in the wrong position according to the planogram is still a violation. The issue lies in: how do you define 'compliance'? If you only look at quantity, you ignore visibility and prime positioning. The correct approach must separate the layers: presence, correct positioning, and visibility. How to measure it? Compare AI scoring results with manual audit results on a small sample. If the deviation is large, the recognition algorithm needs recalibration.
Comparing Three AI Visibility Approaches
When facing the challenge of display scoring, businesses are often confused between three directions. Each approach has a different price tag in terms of cost, time, and accuracy.
| Criteria | Manual Photos + Excel | Off-the-Shelf AI (SaaS) | Industry-Specific Customization |
|---|---|---|---|
| Accuracy | Low, depends on the photographer | Medium to High | Highest |
| Initial Cost | Very Low | Medium | High |
| Deployment Time | Immediate | Days to Weeks | Weeks to Months |
| Scalability | Very Limited | Good | Infrastructure-dependent |
Approach 1, using Excel, is suitable for small scales, under 50 stores. But when scaling to thousands of points of sale, you will be overwhelmed by data. Approach 2, using SaaS solutions like VTraks, is the most balanced choice for most F&B and FMCG businesses in Vietnam, Thailand, and the Philippines. According to vtraks.com, accuracy reaches 99.97% with processing under 1 second per image, serving over 1 million stores. Approach 3, deep customization, should only be used when you have a strong in-house technical team and specific data that the market cannot meet. The price to pay is waiting time and continuous maintenance costs.
Many choose wrong by chasing technology instead of process. They buy AI without standardizing input data. Or they expect AI to automate everything without human oversight of the feedback loop. In projects I have consulted on, from Masan to retail chains in Mexico, the biggest lesson is: AI is only as good as the data you feed it. If you don't have a standardized photo-taking process and a team for periodic sample checks, no matter how expensive the AI Visibility solution is, the result will just be a mess of unreliable data.
Measuring and Fixing Blind Spots
What is missing? The ability to detect fake and duplicate images. Many employees re-take old photos to trick the system. Why is this difficult? Detecting fake images requires more complex algorithms than standard face recognition. How to solve it? Combine AI content recognition with metadata checks and image hash comparison. How to measure it? The rate of abnormal image detection. If the system lacks this feature, you are leaving a major loophole in management. Additionally, integrating instant alerts via channels like Zalo or Telegram helps shorten the feedback cycle from days to hours. This is the deciding factor in whether sales teams actually act on the data.
We have not yet solved the cost problem for remote areas, where network connections are weak and taking high-quality photos is difficult. More flexibility is needed in data collection methods, such as offline-first with later synchronization. Furthermore, training employees to understand why they need to take photos correctly remains a corporate culture challenge rather than a technological one. AI does not run itself. It needs humans to nurture the data. If you want to implement effective AI display scoring, start with the process, not the software. Clearly define what you are measuring and why, before asking what AI can do.
If you are weighing up a similar project, our team can help you scope it before you spend anything. See what we build or book a conversation.