When ChatGPT Helps Ops (And When It Hallucinates Locations)

Published 2026-02-24 by Faraz Rahimi

Language models draft, summarize, and translate. They should not be the system of record for a bin.

When ChatGPT Helps Ops (And When It Hallucinates Locations)

Tags: AI, Operations, LLM, Warehouse, unlisted

I will use a language model to turn a messy vendor email into a checklist. I will not ask it where SKU 4412 lives. It will answer anyway, which is the problem.

Locations come from scans. Descriptions can be cleaned with a model if a human sees the diff. Invented bins will be believed because the sentence sounds calm.

Put the model after the query

Retrieve the row, then ask the model to explain it in plainer language. Never the reverse. If retrieval is empty, the model must say empty, not improvise.

Logs for prompts that touch operations

If a model suggested a putaway location from "context," I want that in an audit. That feature should probably not exist.

Original post: https://farazrahimi.com/posts/when-chatgpt-helps-ops-and-when-it-hallucinates-locations