Your ERP Cannot Track a Batch It Never Saw
A pallet arrives with one batch number on the label, another in the receiving note, and no temperature record attached. The ERP says available. Quality says hold. Production says it needs the material now. Congratulations, the spreadsheet has become management.
For a small food manufacturer, this is not abstract data hygiene. It is delayed production, confused inventory, and people walking the floor to reconstruct facts that should already exist.
Record the Event, Then Use AI
A forward deployed AI engineer starts at receiving and follows the batch through quality, production, and shipping. The ERP remains the system of record. The agent reads messy notes, finds the likely batch, and prepares an exception. Deterministic rules verify quantities, status, and permissions before anyone posts a transaction.
The event layer needs simple answers: what moved, when, where, and why. The current GS1 EPCIS standard defines that shared event model. You do not need a grand transformation program. You need stable identifiers and fewer mystery fields.
Reuse the Boring Parts
The reusable agentic asset is not a clever prompt. It is the adapter, event schema, validation rules, approval queue, test examples, and audit log. That package can move from receiving discrepancies to quality holds and shipment investigations.
The useful progress over the past eight months is mostly this: reusable connectors, stricter structured data, and better evaluations are easier to assemble. Not sentience. Plumbing.
Technically, every event gets an idempotency key so retries do not create duplicate inventory. The model proposes. Code checks. A person approves risky changes. Current NIST guidance treats AI risk management as continuous work across governance, mapping, measurement, and management. That is sensible system design, not ceremonial paperwork.
Measure the Damn Queue
Illustrative calculation: 60 batch exceptions each week at 12 minutes require 12 hours. At four minutes each, they require four hours. At an assumed loaded labor cost of $35 per hour, that is $280 of weekly capacity, not guaranteed profit or payroll savings.
Track bad matches, duplicate events, approval time, and reopened cases. Savings from avoided scrap are assumptions until real plant data proves them.
If three spreadsheets disagree, you do not have three sources of truth. You have a small civil war.