AI Inventory Cycle Counts: Review Variances Before Stock Is Adjusted
Route inventory cycle count variances to a reviewer before any stock adjustment posts. Set thresholds, evidence rules and approval steps for each count.

To route inventory count variances to a reviewer, an AI agent compares each counted quantity with the system record, classifies the gap by size, value, location and count status, and sends anything outside the approved range to a named person as a proposed adjustment. Small, explained gaps can follow a rule the business has approved in advance. Nothing changes the stock ledger until that rule or a reviewer covers the exact case, so the system never records a count error as fact.
Key takeaways
- A count variance is a claim about stock until someone confirms it with evidence.
- Set tolerances by item category, value and location before the first count runs.
- The agent queues every proposed adjustment and posts nothing unless an approved rule covers that exact case.
- Each reviewer decision needs a named owner, a reason code and a time stamp.
- Review the rules on a fixed schedule using the decisions reviewers actually made.
Why a count variance is a question, not a correction
A gap between the counted quantity and the system record can come from a miscount, a receipt not yet put away, a transfer still in transit, a unit of measure mismatch, or damage and loss that were never recorded. The system cannot tell these causes apart, and a person standing at the shelf often can. An adjustment overwrites the record, and that record may feed reorder logic, cost reports and customer promises, so posting one is a decision with consequences.
An AI agent is useful here because it can gather the evidence quickly: recent movements, the previous count, open receiving documents and the reason the count was started. What it should not do is treat its own comparison as the final answer. The agent prepares the case, and a named person accepts the number.
Set the variance rules before the first count
The rules are the part most teams skip, and they decide whether the review queue is useful or flooded. Write them down with the operations lead, then have the agent apply them literally. A useful rule set separates the question of how large the gap is from the question of how important the bin is.
For each variance, the decision guide should answer one of four routes:
- Within tolerance, no open movements and an approved rule in place: the adjustment may post under that rule, and the rule identifier is recorded with it.
- Out of tolerance on a low-value item in a bin that is quick to recount: send the line to a recount before any reviewer time is spent.
- Out of tolerance on a high-value item, after a recent receipt or transfer, or with repeated gaps in the same bin: send the proposal to the reviewer queue with the evidence attached.
- Pattern suggests theft, damage or a faulty count process: escalate to the stock controller and hold the bin open.
Tolerances are business decisions, so each one needs an owner and a date of last review. Start conservative. It is easier to loosen a tolerance after reviewers confirm it works than to explain a posted error later.
A step-by-step routing plan for each variance
Run every variance through the same sequence so that reviewers can trust the output:
- Lock the count line so later edits cannot change the comparison silently.
- Pull the system quantity at the moment the count started, plus every movement recorded since.
- Compare the gap with the tolerance for that category and location.
- Search the record for explanations such as open receipts, transfers in transit or unit conversions.
- Choose the route from the decision guide and record why that route was chosen.
- Build the proposal with the counted quantity, system quantity, gap, evidence links, suggested action and deadline.
- Post only what an approved rule or a reviewer has accepted, and record the decision with a reason code.
If a step cannot be completed, for example because the movement history is missing, the variance goes to a person with that gap named. It should not be skipped or quietly accepted. For how recounts and holds fit into the wider count process, see cycle counts that avoid blind writes.
What the reviewer sees in the queue
A reviewer should not need five screens to make one decision. A good proposal shows the item and bin, the counted and system quantities, the gap, the value impact in your reporting unit, the last three movements, the previous two counts for that bin, the rule that triggered the route and the suggested action. Reviewers then choose from a short set of outcomes: accept the adjustment, request a recount, hold the line for investigation, or reject with a reason.
The fixed reason list matters more than it looks. Free text such as looks off cannot be counted, while codes such as receipt pending, damage found or unit mismatch show which rules need work. A rejection without a reason should not be allowed to close the proposal.
Illustrative scenario: a receipt that was never put away
Illustrative scenario: a counter finds fewer units of one fastener in a bin than the system shows. The agent checks the record and finds a receiving document for the same item, marked as received at the dock but never put away, with a quantity that matches the gap. Because the explanation is already in the record and the item is low value, the agent does not propose an adjustment. It opens a put-away task for the receiving lead and keeps the count line open. If the put-away is not confirmed by the end of the shift, the line moves to the reviewer queue with the receiving document attached, so the person deciding sees the whole story.
Permissions and stop points for the agent
The agent should read counts, movements and receiving records, and write only to the review queue and the task list. It should not write directly to the stock ledger, change unit costs, close a count with open variances, or edit a count sheet after submission. Each permission should map to a named role, following the least privilege principle described in the OWASP authorization guidance. Agent design guidance from Anthropic also favors explicit checkpoints before consequential actions, so stop points belong in the workflow itself, not in the model's judgment.
In Be Human's managed approach, the first engagement defines the variance rules, the evidence each proposal must carry, the reviewer who owns each category and the agent's read and write boundaries. Ongoing operations cover monitoring of the queue, recurring evaluations on a sample of proposals, permission review when the warehouse system changes, and incident response if the agent proposes a change the rules should have stopped. Rule changes go through the same owner who approved the original rules, and the agent's working memory of rules and past decisions stays under your control.
Keeping the variance rules accurate over time
Rules drift. A new supplier packaging, a relocated bin or a change in counting method can make an old tolerance wrong without anyone noticing. Review the queue on a fixed cadence: which routes were most common, which proposals reviewers rejected, which reason codes appeared most, and whether any adjustment was later reversed. When reversals cluster in one category, change the rule or the count method, and record the change with an owner. The NIST AI Risk Management Framework describes this ongoing cycle of measuring and managing AI risk, which fits a count workflow that keeps changing after launch.
If you are not sure which part of your count process should go to an agent first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
How often should the variance thresholds change? Review them on a fixed schedule, and also after any change to packaging, bin layout or counting method. Changes should come from reviewer decisions and reversals in the log, not from pressure to clear the queue faster. Each change needs an owner and a record of what it replaced.
Can the agent post a small adjustment without a reviewer? Yes, but only under a written rule that the operations lead has approved for that exact category, location and gap range. The adjustment must carry the rule identifier, and anything outside the rule goes to the queue. Many teams begin with proposals only and allow automatic posting after reviewers confirm the rule works.
What if the count sheet is wrong rather than the system? Treat a pattern of count sheet errors as a process finding, not as stock loss. Route the line to a recount by a different person, and record whether the error came from the sheet design, the counter or the timing of the count. The agent can flag the pattern, but a stock controller should decide whether the method needs to change.
Who should review high-value variances? Assign the reviewer by category and location, not by who happens to be available that day. The reviewer needs authority to accept an adjustment and access to the receiving and transfer records needed to check the explanation. If one person cannot cover the queue, name a backup in the rule set so that no proposal waits without an owner.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


