Implementing inventory software: fix the data before it scales

Here is the uncomfortable truth about every inventory system ever sold: it multiplies whatever you feed it. Give it clean numbers and it scales a good process across every SKU and channel. Give it three spellings of the same item and an on-hand count from last spring, and it produces wrong answers faster than your spreadsheet ever did — now with more confidence and in more places at once. Software does not solve a bad data problem.

What an implementation does give you is leverage you will not get again: for a few weeks, everyone agrees the data is worth looking at. The owner will approve a count. The warehouse will tolerate questions. Nobody defends the duplicate SKUs. Use that window to fix problems while they are still small, because after go-live every one of them scales and multiplies with the system.

Live inventory comes first

Every feature you are buying — forecasts, reorder suggestions, channel sync — sits on one assumption: the on-hand number is true. If it is not, you do not have an inventory system; you have a rumor mill with reports. So before configuring anything, make inventory live: the system’s count matches the shelf, commitments to customers and incoming purchase orders are recorded where the system can see them, adjustments are posted the day they happen, and one named person owns accuracy per location. That is the first milestone of the project, not a cleanup task for later.

Clean the item master before you migrate it

• •

  Merge the duplicates. “BLT-2IN”, “Bolt 2in” and “2" BOLT (old)” are one item holding three partial stock records. Pick a survivor, move the history, kill the rest.

• •

  Pin down units. If you buy by the box and sell by the each, record both and the conversion — half of all “impossible” counts are unit confusion.

• •

  Don’t migrate the dead. Items with no sale and no purchase in 24 months get archived, not imported. Every dead SKU you carry over is a row someone will count, forecast and scroll past forever.

• •

  Take lead times from receipts, not from quotes. The supplier’s “two weeks” is marketing; the gap between your POs and their arrivals is data.

• •

  Close the ghost orders. Purchase orders from last year that never arrived and sales orders that already shipped are still inflating someone’s available-to-promise.

Count what matters, not everything at once

A wall-to-wall freeze count is expensive, morale-sapping and stale within a month. Cycle counting beats it: count your A items (chapter 3 of our playbook explains the split) every month, B items quarterly, C items once a year, a few bins per day. Count an item when its stock is low — fifteen minutes instead of two hours, and the error stands out. Count blind, without showing the counter the expected number, or you will mostly confirm your mistakes. And treat every variance as a clue: a bin that is short keeps a reason in receiving, returns or the damage pile, and the reason outlives the correction.

The new edge: let AI do the tedious half

The honest reason data cleanup gets skipped is that it is brutally boring: thousands of rows, judgment calls every tenth line. This is exactly the work AI assistants are good at now. Give Claude or ChatGPT your exported sheets and have them do the first pass:

• •

  Find the duplicates a filter can’t: “group rows that likely describe the same item despite different SKUs, spellings or abbreviations, and show your confidence per group.”

• •

  Match items across systems: feed it the ERP export and the webstore export and ask for a SKU-to-SKU mapping table with the unmatched leftovers listed separately.

• •

  Normalize the mess: consistent descriptions, one unit convention, supplier names deduplicated and matched to your vendor list.

• •

  Hunt anomalies before they migrate: negative stock, costs ten times their neighbors, lead times of zero or 400 days, POs older than their items’ entire sales history.

• •

  Merge channel exports into one sales history per SKU per month, the input every forecast in your new system will depend on.

Treat the output like a sharp intern’s first draft: spot-check a sample by hand, keep a person on every merge decision, and never let a tool write corrections back into a live system unreviewed. Verified that way, a cleanup that used to take a hired consultant a month fits into the evenings of one implementation. Our playbook has a full chapter of prompts and guardrails for this kind of work.

Playbook chapter: using AI on your inventory data

Then roll out one workflow at a time

With live inventory and a clean item master, the software part is almost anticlimactic. Pick the one weekly job that hurts most — entering orders, deciding what to buy, receiving — and move only that. Test it with real orders, including one messy one. Run old and new in parallel until the new numbers earn trust, review the exceptions, then take the next workflow. Teams that stall are almost never stuck on features; they are stuck relitigating data they could have fixed in week one.