AI Camera Inventory Counting: Real-World Lessons

13/11/2026

Context: Warehouses Don't Wait

Periodic inventory audits are always a nightmare for materials accountants and warehouse managers.

We once worked with an electronic component manufacturing plant in Binh Duong. The issue wasn't a lack of stock, but a lack of up-to-date information. The ERP system reported 500 pallets, but only 420 were actually on the floor. 80 pallets vanished over two weeks with no one able to explain why. Warehouse staff had to spend entire weekends manually reconciling records, leading to exhaustion and continued errors.

This is when we proposed applying computer vision. Not to replace humans, but to free them from tasks that machines can perform 100 times faster. The specific goal: use existing AI cameras at warehouse entry/exit points to monitor goods flow in real time. There was no need to replace the entire CCTV system, just add an AI processing layer to the existing video data.

Why Manual Monitoring Fails Repeatedly

Many businesses think that simply adding more staff is the solution. But reality shows the opposite.

• Numbers don't match: A 15% discrepancy between records and reality is common in low-automation warehouses. Each manual count takes 4 to 6 hours for a medium-sized warehouse area.
• Delayed information: Data only updates when goods are checked in or out. If items are stolen or lost during internal movement, the system doesn't detect it until the next inventory audit.
• Hidden costs: Beyond staff salaries, there are production delays caused by sudden material shortages. In the electronics industry, a one-day delay can cause billions of dong in losses due to canceled orders.

The core issue lies in continuity. Manual counting is intermittent. AI cameras are continuous. That is the fundamental difference.

Deploying AI Cameras: What We Did

We chose a solution integrated with the existing IP camera system. No new camera hardware investment was required, only ensuring a minimum resolution of 1080p and a wide viewing angle covering the aisles.

Layer 1: Define Regions of Interest (ROI).

We drew frames around palletized areas and entry/exit points. The AI focuses processing on these regions, ignoring dead zones to save GPU resources.

Layer 2: Motion sensitivity and barcode recognition.

Instead of trying to identify individual products inside pallets (which is very difficult and expensive), we focused on counting the number of pallets passing through and reading barcodes on the outer pallet labels. This is a reasonable trade-off. You don't need to know the exact component type inside pallet #12 if you just need to know that pallet #12 left the warehouse at 14:30.

Layer 3: Real-time alert integration.

When a pallet passes through an entry/exit without a corresponding outbound slip in the system, AIV Camera sends an alert via Telegram to the shift leader and accountant within under 1 second. This is the most critical feature. It turns an end-of-day issue into an immediate action item.

During the pilot phase of this project, we encountered lighting errors in the afternoon when sunlight shone diagonally into the warehouse. Barcode recognition accuracy dropped to about 92%. We had to adjust the camera angle and add a glare filter. Lesson: the physical environment is always more important than the algorithm.

Measuring Effectiveness and Accepting Limitations

After 3 months of stable operation, we measured the following metrics:

• Discrepancy detection time: Reduced from an average of 14 days (monthly inventory) to under 5 minutes.
• Inventory discrepancy: Narrowed from 15% to under 2%. This number is not 0%, but it is enough to eliminate most major risks.
• Inventory staff costs: Reduced daily reconciliation workload by approximately 60%.

However, there are limitations we do not hide.

AI cameras are not magic. They cannot detect if goods are removed slowly, without passing through monitored entry/exit points, or if a pallet is opened and a few products are taken without changing the overall position. For high-value products, it is still necessary to combine with RFID systems or random sample audits.

In addition, maintaining the system requires care. Dirty camera lenses, misaligned angles due to collisions, or changing lighting conditions all affect accuracy. We scheduled weekly lens cleaning. This is a real operational cost, not a virtual one.

We believe AI cameras are a powerful monitoring tool, but they are only effective when placed within a clear workflow. They do not replace humans; they free humans from repetitive and tiring tasks.

A small action you can take this week: Choose the warehouse entry/exit with the highest goods flow, check if the IP camera there has sufficient resolution and a clear viewing angle for pallet barcodes. If so, plan to test a basic recognition module. You will see the difference from the very first alerts about abnormal pallet movements.

AIVISION helps enterprises turn AI into working systems. Explore our enterprise AI solutions, read more on the AIVISION blog, or talk to our team about your own use case.