Article · · 6 min read
The ten rules for using AI in a warehouse without leaking anything
Somebody above you said “use AI.” Nobody said what is allowed. The first person who pastes a payroll file with names into a chat window ends the program for the whole building.
Here are ten rules, each with what it looks like broken. Print it. Pin it where the exports get pulled.
Before the rules: three questions for IT
The rules assume you have cleared the gate. Most managers have not. Ask these three, in writing.
- Which AI tools are approved, and for what kind of data? Make them name the tool and the tier.
- Is there a data-loss rule on uploads, and does an anonymized operational export fall inside it? Order IDs, timestamps, quantities, costs, no names. Most policies call that operational data, not personal data.
- Who owns the policy, and will they give you one line of written approval for that specific use?
Write the answers at the top of your prompt log. When the first scare comes through, the manager with a dated approval keeps the tool.
Rules 1 to 4: what goes in
Each rule exists because somebody skipped it and paid.
1. Your company's policy wins. If it says no, no. If it says nothing, that is not a yes. Ask. Broken, it looks like: a manager in front of HR and IT in the same meeting, and a building that loses the tool for everyone.
2. Never names, wages, or confidential files. Employee IDs and hours: yes. Names, pay rates, addresses, customer contacts, contracts, anything stamped confidential: no, ever. Delete the column before the file leaves your laptop. If the export cannot be made without names, it does not get uploaded. Broken, it looks like: a labor file with the name column still in it, and a screenshot of the output in a group chat by lunch.
3. Read-only on every system. The tool reads exports. It never logs into the WMS, never writes a record, never sends a message, never moves a schedule. Broken, it looks like: a "helper" allowed to edit the cycle-count queue, and a batch of counts gone with nobody able to say what removed them.
4. Draft, never send. Every email, SOP, summary, and handoff it writes is a draft you read and sign. Your name goes on it, not the tool's. Broken, it looks like: a summary that went up the chain with your name on it and a figure no file ever contained.
Rules 5 to 7: what comes out
5. Gaps stay blank, flagged. If the file cannot support a metric, the page shows a blank and the reason. Not an estimate. Not yesterday's figure carried forward. Not "typical for the industry." A blank with a reason beats a number with no source. Broken, it looks like: a productivity number in a brief built from a file that has no units in it.
6. Verify one number by hand, every time. A different one each day. Thirty seconds in Excel, from the same file. If it is wrong, nothing else on the page is trusted until you know why. Broken, it looks like: a month of briefs where the OT percentage used regular hours as the denominator. Nobody caught it because nobody checked one.
7. One source per brief. Do not mix a WMS export and a payroll export in one prompt until you have joined them by hand once and written down the join: which ID, which date rule. Unstated joins are where invented numbers hide. Broken, it looks like: a labor-per-order number that quietly dropped every shipment whose packer ID was missing from that day's payroll pull.
Rules 8 to 10: how you keep it honest
8. Date and version everything. The report day is the latest timestamp in the data. Never the filename, never today. One file-naming rule, and the date in the header of every page. Broken, it looks like: a Monday huddle run on last Wednesday's numbers because the file was named "daily."
9. Keep a prompt log. One line per run: date, prompt name, input file, what came back, what you changed. Fifteen seconds. It is how you answer "where did that number come from?" six weeks later, and what you show IT when they ask. Broken, it looks like: a brief everyone liked in March that nobody can rebuild in May.
10. If you cannot explain it at the huddle, do not use it. Make the tool show the formula. Understand the formula. Then it is your number. Broken, it looks like: a lead asks "how'd you get that?" and the honest answer is "the AI said so." That is not a number. That is a rumor.
Rule 5, worked on the practice warehouse
The practice warehouse is a fictional building with problems planted on purpose. The Morning Read runs on two of its files, a shipments export and a labor export. Neither has a units column.
Report day against the same weekday, as the tested prompt returns it:
| Metric | Tue 2026-03-31 | Tue 2026-03-24 |
|---|---|---|
| Orders shipped | 907 | 976 |
| Hours worked | 385 | 409 |
| OT % (building) | 12.8% | 17.7% |
| Units per labor hour | [GAP] | [GAP] |
Under the table, the gaps list reads: "Units per labor hour — no units column in either file. [GAP]"
That is the rule working. 907 orders, 385 hours, no units anywhere, so the productivity line stays empty, with the reason.
What broken looks like
Same two files, same prompt without the rule, and a units-per-hour figure shows up in the table, as sure of itself as the rows above it. The tool divided something by hours to be helpful. You post it. Three weeks later your director asks why productivity moved and you cannot say what "units" meant.
The fix is not "please be more careful." It is one line in the prompt's definitions: this metric is not computable from these files, leave it blank. The free 12 prompts carry that line. If your tool gave you a units-per-hour number, it invented one.
What the poster on the wall does for a rollout
The ten rules, one page, the biggest type that fits, no explanation. One copy where the exports get pulled, one by the printer.
Its job is to take you out of the loop. When a lead asks whether they can paste something, point at the wall before you answer. After a couple of weeks the question goes to the wall first. A rule you have to be present to enforce is not a rule.
It also gives you something to count: blanks flagged each week, and numbers verified by hand. A week with zero blanks means somebody is filling gaps. Post both counts next to the poster.
Three people will test it. The enthusiast uploads the payroll file with names because it is faster: rule 2, day one, no exceptions, or the program ends when HR hears. The skeptic will not run it: give them the verification job; nobody finds a wrong number like a skeptic. The believer trusts the output because it sounds sure: rule 10, at the huddle, once.
What not to do
- Never upload a file with a name column, "just this once."
- Never let the tool connect to anything that can write a record.
- Never send what it drafted without reading every line.
- Never run a real export before you can get the answer key's numbers on the fake one.
The close
None of this is about the tool. It is about being able to tell the huddle, and IT, where every number came from and what left the building: IDs, timestamps, quantities, costs. Nothing else.
The rules are built into the twelve free prompts. If IT gives you a paste box and nothing else, the rules still hold; the Copilot workflow shows how.
Your company's AI and data policy wins. Never paste names, wages, or confidential files.
— Floor Read
Sources: morning-read/shipments.csv, morning-read/labor.csv — Copperline practice warehouse, fictional.
Written by a sitting director of operations at a large distribution center — eight years on fulfillment floors, from frontline supervisor to running a multi-hundred-person building's P&L. No software vendor pays for a mention here. The author keeps the day job and the brand separate, which is why there's no headshot; the sample chapter is free so you can judge the work instead.