Superintelligence-Ready Cycle Counts: Stock Checks Without Blind Writes
An AI agent can prepare cycle count lists and variance reports, while every stock adjustment waits for a named person's approval. Here is how to set it up.

An AI agent can prepare an inventory cycle count by building the count list, pulling expected quantities and writing a variance report once the counts arrive. It should not change the stock record itself. Each proposed adjustment waits in an approval step, where a named person checks the counted value, the evidence and the likely cause before the system posts anything.
Key takeaways
- The agent prepares count sheets, expected quantities and variance reports. It does not post stock adjustments.
- Every adjustment proposal carries the counted value, the expected value, the evidence and a reason code.
- Approval belongs to named roles with limits, and large or unusual variances go to a second reviewer.
- Each approved write carries a unique key tied to the count, so a retry cannot post the same change twice.
- Check the agent's flags against what counters actually find, and keep that review on a fixed schedule.
What a cycle count is and where blind writes happen
A cycle count checks a small slice of inventory on a rotating schedule instead of stopping the whole warehouse for a full physical count. The slice might be one shelf, one supplier's items or the fast-moving parts that cause the most trouble when the record is wrong. The goal is to find the gap between the system's on-hand quantity and the shelf, then decide what to do about it.
A blind write happens when something changes the on-hand quantity to match a counted number without anyone checking why the difference exists. A script that overwrites the record after each count is the classic example. It can be wrong in two directions: a miscount becomes the truth, or a real loss gets hidden under a quiet adjustment that nobody reviewed. An agent with write access to stock can make this mistake faster than a person, which is why its job should stop before the write.
Most variances have ordinary causes: a receipt that was counted but not posted, a pick that went to the wrong bin, a unit of measure mismatch between cases and eaches, damaged goods that were never moved to a quarantine location, or a count entered against the wrong location code. The useful question is not only "how many are missing" but "what changed recently that would explain the difference." That question is where an agent helps, because it can read recent movements quickly and show them to the person who decides.
What is the step-by-step plan for the agent's preparation work?
Start with the work the agent does before anyone approves anything. Keep each step visible in a log so a reviewer can see what the agent read and what it produced.
- Select the count list: choose the locations or items due under your rotation rule, and skip any item with an open count or an unresolved receiving discrepancy.
- Pull the expected position: read the on-hand quantity, open orders, pending receipts and recent adjustments, each with its timestamp.
- Generate the count sheet: sort by walking route so the counter moves once through the location, and hide the expected quantity if your policy requires a blind count.
- Accept counts only through the counting method you defined, such as a handheld entry or a named staff submission, and record the time and location with each count.
- Compute variances in your own units and value bands, and tag a probable cause only when a recent movement clearly matches.
- Draft one adjustment proposal per item, not one per batch, and attach the count, the expected position and the movement history used.
- Route each proposal to the approver for its threshold. Nothing is posted at this stage.
- Write the variance report, which summarizes coverage, open variances and aging, and lists what was proposed and what was decided.
The agent should be able to finish all of this without touching the stock record. If a step needs a write to complete, that step belongs on the approver's side of the line.
How do you design an approval step that stays a real check?
An approval step is only a control if the approver can see enough to disagree. A list of two hundred proposals approved with one click is a rubber stamp, and it gives the appearance of control without the substance. Design the approval so that each decision is small, specific and recorded. For wider routing patterns, see how to route each request to the right approver.
Use a simple decision guide, written down before go-live:
- Below the low threshold you set in units and value, the inventory lead approves in batches after checking a sample against the count sheets.
- Above that threshold, or when the proposed change reverses a receipt posted in the last few days, a second person approves the item individually.
- When the count is outside the rotation window or the location is under dispute, the item goes to a recount rather than an adjustment.
- When the cause is unclear, the proposal shows "unknown" and the approver must enter a reason before the item can be approved.
- Any proposal that changes cost or valuation goes to finance, not only to operations.
Permissions should match the guide. The agent's credential should read stock and counts and create proposals, but not post adjustments. Approvers hold the write permission and nothing broader. The OWASP Authorization Cheat Sheet describes the principle of granting only the access a function needs, and it applies cleanly here. Review these permissions on a schedule, because a role that was narrow at launch tends to widen as people change jobs.
Illustrative scenario: A distribution team counts a bin of fasteners and finds eleven fewer units than the system shows. The agent's proposal notes that a receipt of the same item was posted the previous afternoon and that the receipt quantity matches the gap. The approver checks the receiving record, finds the receipt was entered twice, and rejects the adjustment in favor of a correction to the receipt. Nothing in the stock record changes until a person chooses the fix.
How do you prevent double posting and stale data?
Two failures appear again and again in inventory integrations: the same change is posted twice after a timeout, and an adjustment is approved against a quantity that has already moved. Both are preventable with plain habits.
The first is idempotency. Stripe's documentation on idempotent requests describes how a provider can use a key so that a retried request does not repeat its effect. Apply the same idea inside your own inventory integration. Give each approved adjustment a unique key built from the count identifier, the item and the location, and have the posting step refuse a key it has already used. A retry then returns the original result instead of creating a second adjustment.
The second is staleness. Between proposal and approval, a receipt or a pick may change the on-hand quantity. Before an approved change posts, the system should compare the expected quantity the proposal was built on with the current value. If they differ beyond your tolerance, the proposal returns to the queue with the new figures, and the approver decides again. This check is cheap, and it stops a correct decision from being applied to yesterday's stock.
What belongs in a variance report people will act on?
A variance report is useful when someone can read it in a few minutes and know what needs a decision. Long tables of every counted item tend to be skimmed and then ignored. The agent should produce a short summary with a fixed structure, and link to detail only when a reader asks for it.
A practical structure includes count coverage, meaning which locations were due and which were counted; open variances grouped by value band; unresolved counts with their age; proposals waiting for approval; and decisions made since the last report, with reason codes. Keep the language literal. Write "eleven units short against the expected quantity" rather than a vague label about risk. The report should describe what happened and what is waiting, and leave the judgment to the people who own the stock.
Tie the report to the review you already run. If your team meets weekly about operations, the variance report should be the first item, and every open proposal should have an owner and a date. Lessons from one review should change a rule or a threshold, and that change should be written into the agent's instructions with the date it took effect.
How do you monitor the agent after it goes live?
Before the agent has any effect on approvals, run it in a shadow period. It prepares count lists and variance flags as it would in production, but the team performs the real counts and decisions as before. Afterward, compare the two. Note which real variances the agent flagged, which it missed, and which flags turned out to be noise. Our guide on how to watch agents before granting autonomy covers this step in more detail.
Once live, review a sample of proposals each month: what the agent read, what it proposed, what the approver changed and why. Track rejections by reason. A rising number of rejections for an unknown cause often means the movement data is incomplete, and the fix belongs in the data, not the prompt. The NIST AI Risk Management Framework is a useful reference for structuring this kind of measurement and review, because it treats oversight as an ongoing activity rather than a one-time test.
Keep a trace of each tool call the agent makes, including what it read and what it wrote, so an audit can reconstruct any proposal. The OpenTelemetry observability primer explains why this kind of trace matters for diagnosing behavior in production systems, and the same logic applies to stock work.
How does Be Human run this in practice?
Be Human designs and operates this kind of agent as a managed service. In a first engagement, the team defines one useful role, such as the cycle count preparer, along with the memory and tools it needs, where approvals sit and what success looks like, for example how many real variances the team catches before a physical audit. Managed operations then covers monitoring, recurring evaluations of the agent's flags, permission reviews, maintenance when the inventory system changes, and incident response if a bad proposal reaches an approver. Each change to the agent's rules is recorded so the owner can read what changed and when.
If you are not sure which stock task should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent post stock adjustments on its own? It can, technically, if it has write access, but that is the design this article advises against. Give the agent read access and the ability to create proposals, and keep posting with a named approver. The check is the point of the step, so a faster posting path removes the control rather than improving it.
How often should cycle counts run? The right frequency depends on how much an item is worth, how often it moves and how often past counts found errors. Start with the items where a wrong record would cause the most harm, such as high-value or fast-moving stock, and widen the rotation as the variance history becomes clear. The agent can propose a schedule, but the inventory lead should set it.
What happens when the counter and the system disagree on a count? The agent should not decide which record is right. It should mark the item as disputed, attach both records with their timestamps, and send it for a recount or a supervisor check. Only after that check does an adjustment become a proposal for approval.
Does this require a particular inventory system? The approach needs three things: read access to stock, counts and movement history; a place to hold pending proposals outside the live record; and a posting step that a person controls. Many systems can support this with a review queue or staging table, and if yours cannot, the first project is closing that gap, not building the agent. The note on exception queues describes the owner, reason and deadline fields that make such a queue usable.
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.


