How do I actually use AI in ecommerce operations, not marketing?

Start with one live system of record. Then give the model a job with a gate.

The model is not the hard part. Permissioned live data is. A CSV goes stale while you paste it. A nightly BI warehouse is better than a CSV and still late for a drop: fine for last week's contribution margin, wrong for whether you can accept this afternoon's 850. MCP or an equivalent live connection is the difference between asking and acting. If the model cannot name the warehouse, the lot, the clearing account, or the inbound ETA, it is not ops. Marketing AI (copy, creative, audiences) does not count.

Do not start with "write me a forecast." A forecast without receipts, reservations, and channel buffers is fan fiction. Do not let the model post a journal, send a PO, or change published ATS without a person. The gate is the job: draft, flag, ask. A human clicks send.

Useful ops jobs to start with

Useful ops jobs, in that order.

Exception lists, not dashboards. Which SKUs will stock out before the next receipt, using actual lead time, not the supplier's promise. Include reserved, inbound ETA, and published channel qty. A dashboard that is green while a kit component is dry is decoration.

Draft purchase requests from velocity, inbound, and MOQ. Round to the case. Mark the qty DRAFT. A human sends the PO. The model does not talk to the factory.

Cycle-count and 3PL variance: yesterday's count file versus on-hand, only the SKUs that moved. Ignore the thousands that matched. The list is the work.

Close queues: clearing accounts that are not near zero, unmapped GLs, open internal shipments, receipts with no bill. Month-end is a pile of exceptions. The model can rank them. It cannot lock the period.

Warehouse Q&A: where is this SKU, which lot expires first, why did this order split. That only works if location, lot, and reservation are live fields, not a Slack guess.

Write prompts and log the misses

Write the prompt like an SOP: columns, the math, what to flag, what not to touch. The Claude blocks at the end of these posts are the pattern. Review the draft the way you review a junior buyer: check the math, then click send yourself. Log every accepted draft so you can see where the model is consistently wrong. Lead time, kits, and FBA Reserved are the usual misses. A kit that looks in stock because the parent has units while a component is at zero will sail through a naive stockout list. Amazon Reserved looks like on-hand and is not sellable. That log is the training set. Prompts without a feedback loop rot.

Grunt work that used to live in twenty-five spreadsheets becomes twenty-five small tools only after the ledger is one place. Tools on top of a fight between Shopify and the 3PL are still a fight.