AI Agents in Inventory Management: Where It Breaks
Reorder decisions get made from a spreadsheet that is already out of date.
The loop, as it actually runs
- Stock levels monitored
- demand forecast
- reorder point hit
- PO raised
- receipt recorded
- counts reconciled
Where it breaks
Not in the steps. In the joins between them:
- Monitoring is periodic rather than continuous
- Forecasting is a formula nobody revisits
- Reorders are triggered by memory
- Cycle counts disagree with the system
What that actually costs
Inventory errors are paid twice and never at the same time. Order too little and you pay in lost sales, which are invisible because nobody records the order that was not placed. Order too much and you pay in working capital and eventually in write-offs, which are visible but arrive a year later attached to a different decision. Because the two costs land in different periods and different reports, the reorder point tends to be set by whoever was most recently embarrassed, which is not a forecasting method.
What we would build first
Continuous monitoring goes first, ahead of forecasting. Forecasting is the intellectually interesting problem and it is the wrong place to start, because a better forecast applied to stale stock levels does not help. Monitoring that runs continuously rather than when someone opens the spreadsheet turns the reorder point from a number someone checks into a trigger that fires. It is also the prerequisite for any forecasting work you do later, so nothing is thrown away.
It runs inside Shopify, NetSuite, Airtable, Google Sheets, Slack — the tools already in use.
How this one goes wrong
An agent that raises reorders automatically will amplify a data error into a purchase. A miscounted cycle count, a receipt recorded twice, a return processed to the wrong SKU — each becomes an order at the vendor rather than a discrepancy in a report. The guard is a sanity band: any reorder that deviates materially from the trailing pattern for that item is proposed rather than placed, and any item whose system count and last physical count disagree is excluded from automatic reordering entirely until the two are reconciled.
The individual jobs inside this loop — what starts each one and where it has to hand back to a person — are covered under finance & bookkeeping.
Where it lands hardest: freight brokerage & 3PL, marketing agencies and construction & trades. Each of those pages carries the sector’s own economics — the three roles that absorb the work, and the season that decides when a build can start.
In most companies this work sits with production planners and operations managers. Each of those pages carries the wage arithmetic for that role, and says plainly where the numbers do not justify the spend.