Inventory, WMS and storerooms: agents for exception handling

A well-defined operational question

Example: which sales orders promised for tomorrow are at risk because available material is missing? The agent combines orders, reservations and current inventory, identifies the source of each number and presents the risk queue. Total stock alone is insufficient: blocked, reserved or remote-warehouse material may not fulfill the order.

Define availability and freshness

Agree on the meanings of physical, available, reserved, blocked and in-transit inventory. Record warehouse, location, lot, serial and unit where applicable. Answers should indicate query time and source system. If the integration is unavailable, the agent must disclose the lack of current data rather than present the last balance as guaranteed availability.

Reconcile before recommending an adjustment

Define a cutoff time and compare records in the same unit, warehouse, location and lot. Synthetic example: ERP shows 120 pieces, the count finds 108 and an unposted withdrawal accounts for 12. The discrepancy does not authorize writing off 12 pieces: first confirm the withdrawal and its accounting period. Otherwise, posting it after the adjustment would reduce inventory twice.

The agent gathers movements, evidence, reservations and counts, presents hypotheses and routes the discrepancy to its owner. Physical confirmation and approval remain recorded. For an authorized action, use an operation key, check the current balance version and query the outcome before retrying a timed-out call.

Preserve links between item, lot and movement

GS1 describes traceability through identification, capture, recording and sharing of product, party and location information, with quality controls between processes. Apply that principle in the pilot by requiring stable identifiers and source evidence; this does not imply GS1 certification for Zild.

A trade unit, case and pallet are not interchangeable. Preserve the original unit, approved conversion factor and destination unit. Matching similar descriptions should be a reviewable suggestion, never silent creation of a new material code. Separate returns awaiting inspection from material released for consumption.

From storeroom to reverse logistics

In the storeroom, organize requisitions, owners and pending materials by service or production order. In WMS, track receipt, picking and shipping exceptions. For reverse logistics, gather the reason, original order, evidence, item condition and authorized decision. Returning an item to stock depends on inspection and correct system recording, not merely confirmation that the return arrived.

Inventory planning and MRP

Use the agent to explain projected shortages, gather lead times and highlight incomplete master data. Material requirements depend on demand, bills of materials, inventory and planning rules. Keep MRP calculation in the configured system and compare documented scenarios. Creating purchase orders, changing replenishment parameters or substituting materials requires approval and constraint validation.

Pilot with difficult cases

Include blocked lots, unit conversions, competing reservations, partial receipts and inventory discrepancies. Measure prioritization accuracy, exception-resolution time, false alerts and accuracy after checking. Do not attribute fewer shortages to the agent without considering demand, seasonality and operational changes. The benchmark should use the same period and availability criteria.

Measure accuracy with clear denominators

Before testing, define whether a record means item-warehouse or item-location-lot, and set an accepted tolerance. Record accuracy = records within tolerance ÷ checked records. Also report absolute unit and value discrepancies without adding incompatible units. Do not offset an excess of one item against a shortage of another to hide errors.

For the agent queue, measure confirmed alerts ÷ issued alerts and detected discrepancies ÷ known discrepancies in the reviewed sample. Record false positives, query failures, checking time and cost per correctly resolved exception, including human work. A smaller queue does not prove that inventory is correct.

NIST warns that cherry-picked easy examples and data unrepresentative of operations can produce misleading evaluations. Include shifts, warehouses, items with different turnover and rare exceptions, retaining a sample separate from instruction tuning. This guidance informs test design; it does not provide a gain rate for Zild.

Connect the real process

Start read-only with one warehouse and one exception queue. WinThor documentation describes credentials and permissions for inventory and order resources; this does not mean universal coverage of every WMS. Check the available ERP connector, data granularity and exposed actions. Use Zild to coordinate queries and owners before enabling inventory movements.