Partly. Rules can plan a stock count, prepare the lists and compare counted quantities with the system, AI can read count sheets and suggest why numbers differ, but people still count the goods and approve every adjustment. If you are looking at inventory counting software, the useful question is which steps of the count it takes over and which it leaves to you. Below, every step of a count gets an honest label, with the conditions that must be true first and a small pilot you can run in one warehouse zone. The aim is not a count without people. It is a count where the repetitive work is done by software, the differences are explained faster and a named person is responsible for what changes in the books.
What each step needs
01● Standard automation
Choosing what to count and when
Deciding the count schedule is mostly rules. A full count once a year, or cycle counts where fast-moving and high-value items are checked more often than slow ones, can be written down as a calendar and a list of selection criteria. Software can then pick the locations or items due this week, skip anything that was counted recently and spread the work so the warehouse is never short of people. A model is not needed, and a rule is easier to audit: anyone can see why a shelf was on the list. A person only sets the criteria once and reviews them when the range or the layout of the warehouse changes.
⛨Works when every item has a location, a value class and a last counted date in the stock system, and the selection criteria are written down and reviewed twice a year.
02● Standard automation
Preparing count lists and pausing stock movements
Count lists, labels and the cut-off moment are plumbing. The system can export what it believes is on each shelf, print or push the list to handheld scanners and record the exact time the count starts. Movements for the counted area are paused or logged separately, so goods that are picked, received or moved during the count do not look like errors afterwards. Decide whether counters see the expected quantity or count blind; blind counting is slower to set up but catches more mistakes. All of this is fixed logic, and it should behave the same way every time.
⛨Works when receipts and picks have timestamps, the cut-off is announced before the count and counters are told whether the expected quantity is visible.
03● Human review
Physically counting the goods
Someone has to stand at the shelf, open the box and count. Today this stays human for most companies. Handheld scanners, barcodes and RFID make it faster and reduce typing mistakes, and drones or shelf cameras exist for large warehouses, but they still need a person to deal with damaged packaging, mixed pallets, items in the wrong place and products that look the same. Counters also notice things no list contains, such as water damage, expired goods or stock stored in an unofficial corner. Keep counting simple and consistent: one unit of measure, one rule for opened boxes and a clear way to mark what has been counted.
⛨Applies when stock is physical and varied; automation of the count itself only pays off in large, regular, well-labelled storage.
04● AI candidate
Capturing and reading the counts
Getting the numbers into the system is where AI helps most, as long as the count was not made directly on a scanner. A model can read handwritten count sheets, photos of shelf labels and pallet tags, match them to item numbers and flag lines it cannot read instead of guessing. It can also turn a spoken or typed note such as 'two boxes damaged, rest ok' into a structured comment. Every extracted number should keep a link to the original photo or sheet, so a reviewer can check the source in seconds. Where counters scan directly, this step does not exist, and that is the better setup.
⛨Works when sheets or photos are legible, item numbers are printed on the shelf and every extracted line stays linked to its source image.
05● Standard automation
Comparing counts with the system and flagging differences
Comparing the counted quantity with the system quantity is arithmetic. Rules can calculate every difference, apply tolerances per item class, such as a few units on cheap items and zero on high-value ones, and sort the results by value. They can also run a second count automatically for lines above the tolerance. This is the part of the process that should never be left to a model, because the answer must be exact, repeatable and explainable to an auditor. Output a short list of variances with quantity, value and last movement date instead of a long report nobody opens.
⛨Works when the system quantity at the cut-off is saved, item values are current and tolerances are agreed with finance before the count.
06● AI candidate
Investigating the causes of variances
Once a variance is confirmed, the question is why. AI is useful here as an investigator that does the legwork: it can look through recent receipts, picks, returns, transfers and adjustments for the item, notice that a delivery was booked under a similar item number, or that the same variance appears on every item from one supplier. It then writes a short list of likely causes with the evidence it used. It does not decide anything. A person confirms the cause by checking the shelf or the paperwork, because a plausible explanation is not a proof, and a wrong explanation can hide a real process problem such as theft or a broken receiving routine.
⛨Works when movement history is complete, the model can only read and not change stock, and each suggestion shows which records it is based on.
07● Keep human
Approving adjustments and write-offs
Changing the stock quantity changes the books, so a named person approves it. That is usually a warehouse or finance lead who sees the variance, the evidence and the value before it is posted. Write-offs above a limit need a second signature, and repeated variances on the same item or location should lead to a process fix rather than another adjustment. Automation can prepare the journal entry and attach the evidence, but it should not post it. Keep the approval record, because auditors and insurers will ask who accepted the difference and on what basis.
⛨Applies to every count; limits for approval are set by finance and reviewed yearly.
A sensible first experiment
Pick one zone of the warehouse with 100 to 200 items and run two cycle counts, one the usual way and one with the new setup. Let rules select the items and compare the numbers. If counts are written on paper, let an AI tool read the sheets and suggest causes for the variances. Measure four numbers: lines read correctly without correction, variances found, how many suggested causes a person confirmed and the minutes spent from first count to approved adjustment. Keep the pilot only if the readings are correct and the confirmed causes make the investigation faster.
The trap to avoid
The most common mistake is letting automation correct the stock quantity so the numbers match, without anyone asking why they differed. That hides theft, receiving mistakes and bad locations, and it makes the next count look better than the warehouse really is. A second mistake is counting while goods are still moving without a cut-off, which creates variances that are not real. Keep the approval step human, log every adjustment with its reason and look at repeated variances as a signal about the process.
Questions teams ask
Can AI count stock instead of people?
In some warehouses, partly. Scanners, RFID, shelf cameras and drones can count faster and more often than people with clipboards, but they work best in tidy, well-labelled, large storage. In most small and mid-sized companies a person still counts, and AI helps afterwards with reading sheets and explaining differences. Check the technology against your own shelves before buying anything.
Is a full yearly stock count still required?
That depends on your country, your accounting rules and your auditor. Many companies must be able to prove the stock value at year end, and some allow cycle counting or sampling if it is well documented. Ask your accountant or auditor what evidence they accept before you replace a full count with cycle counts.
How often should we count stock?
There is no single right frequency. A common approach is to count fast-moving and high-value items more often and slow, cheap items less often, so every item is checked at least once a year. Let the size of past variances guide you: items with repeated differences deserve more frequent counts.
What is a safe first step with AI in stock counting?
Start with reading and explaining, not with changing stock. Let a model read count sheets or suggest causes of variances for one warehouse zone, while rules and people keep the comparison and the approval. Measure how many lines it reads correctly and how often its suggested cause is confirmed.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.