It can help gather the evidence. It can’t count the shelf. AI may be able to pull together the receipts, picks, transfers and adjustments around a mismatch and suggest where to look. A person checks the stock, and the approved process handles any adjustment.
We corrected the count. Why did it drift again?
Fixing a balance and fixing its cause are different jobs. Follow the evidence for the particular item and location, without assuming the cause.
1. Give the investigator more than a flag.
The item. Identifier, location, unit, recorded balance and the latest physical count, with its time.
The movements. Receipts, picks, transfers, returns and adjustments around that count.
The gaps. Missing or duplicate entries, mismatched units and unexplained timing differences, with the records behind them.
Practical check: Treat a suggested cause as a hypothesis.
2. Prepare the evidence. Don’t invent the stock.
Test whether AI can organize the records and explain why an item needs investigating. A person verifies the physical and transaction evidence, and the approved process handles any adjustment. Keep a way to correct errors.
Practical check: Compare it with an ordinary exception report first. It shouldn’t make the balance look right just by changing it.
3. One small test.
Use known past cases or approved test records. Check whether it finds the relevant evidence, and whether the cause it proposes is supported. Track investigation effort, false alarms and missed discrepancies, and record the actual cause when the case closes.
Practical check: Count corrections are not proof of lasting accuracy.
Read more: Did Automation Reduce Supplier Chasing, or Just Add Alerts? · How to Run a Small AI Agent Pilot
This is general guidance, not advice for your situation.