AI in Operations
How can I ask Claude or ChatGPT questions about my live Shopify inventory and margins without exporting CSVs?
Asking Claude or ChatGPT about your live inventory and margins without exporting spreadsheets.
Last updated September 16, 2026
The constraint is a live, permissioned system of record. The model is not the hard part.
Choose the right data pipe
A CSV is a photograph. It goes stale while you paste. You will ask about available-to-sell and the 3PL will have picked 40 units. Use a CSV only when you are debugging one SKU and you will throw the file away. Do not paste the catalog into a chat.
A nightly BI warehouse is better than a CSV: history, joins, a place for WAC and fees. It is still last night. Fine for "what was contribution margin by channel last week." Wrong for "can I accept this Nordstrom 850" or "what will stock out before Thursday's receipt."
A live connection is the third option. MCP is the pattern: the model calls tools against objects you already permission, instead of eating a file. Expose read-only what an ops lead may see: on-hand by location, reserved, inbound ETA, published Shopify qty, WAC, open PO qty, channel class, variable fees if you have them. Do not expose bank credentials, customer PII, or a write path that posts journals. The prompt is an SOP: columns, the math, what to flag, what not to touch. Human-gated on money. Draft a reorder. Do not send it. Do not let the model "fix" Shopify by typing a number. Inventory flows warehouse to channel.
Pick the pipe by the question. Live for anything that can oversell or chargeback today. BI for close, seasonality, and margin after the fact. CSV only as a one-SKU debug you delete. If the answer cannot name the location, the reservation, and the WAC, it is not live inventory. It is last week's export with a friendlier UI. Data age belongs on every answer. A number with no timestamp is a CSV pretending to be live.
The usual wrong answers
The usual wrong answers are lead time, kits, and FBA Reserved. A parent kit that shows units while a component is at zero is not in stock. Amazon Reserved looks like on-hand and will not ship to Shopify today. A published qty that does not match on-hand minus reserved is a publish bug, not a demand signal. Log the questions that were wrong. That log is how you tighten the tool grants: revoke anything the model should not have seen, and add the field it keeps inventing.
A storefront assistant can draft a PO suggestion. It cannot see 3PL counts or landed cost. Treat that as prompting, not a live ledger. Prefer MCP-style live tools for today. Prefer BI for history. Reject a stale CSV for a drop.
Connect Claude or ChatGPT to your Fulfil data with the Fulfil MCP, then run this prompt on your own numbers.
You are an ops agent. You may read live inventory and margin fields. You may draft. You may not post journals, send POs, or change published ATS. Here is a live (not CSV) extract: sku, location, on_hand, reserved, inbound_eta, shopify_published, wac, open_po_qty, channel, fees, data_age_minutes, pipe (mcp / bi / csv). [paste] Produce: 1. Which questions this pipe can answer (live ATS / today's 850 vs last week's margin). 2. SKUs that stock out before inbound, with the date, using only these fields. 3. Rows where published ATS does not match on-hand minus reserved. 4. Fields that must stay read-only, and any PII or write path that should be revoked. Prefer MCP-style live tools for today. Prefer BI for history. Reject a stale CSV for a drop.
Explore this in Fulfil
More on AI in Operations
How do I actually use AI in ecommerce operations, not marketing?
Practical ways to put AI to work in operations, with a person still in the loop.
What tech stack should an AI-forward Shopify Plus / DTC brand use in 2026?
What an AI-ready tech stack looks like for a growing Shopify brand.
How do I replace ecommerce ops spreadsheets with internal tools my team can actually maintain?
Turning fragile ops spreadsheets into internal tools your team can maintain.
See it run on
your data.
Fulfil runs inventory, fulfillment, purchasing, and accounting for scaling DTC brands in one system.