Here is the answer up front: if your WMS data cleansing has not reached a go-live standard, do not cut over. The standard is a list below you can check off item by item; any unchecked item means you delay. I have seen this exact failure three times, and all three followed the same script. Inventory numbers in the old system looked plausible enough. The team migrated them as-is. In the first week on the new system, pickers could not find stock, cycle-count variances exploded, customer service phones lit up, and the project team rolled back at midnight. A system switch is a change of ledger, and a dirty ledger stays dirty no matter how good the new software is.
In May 2023, a home-goods e-commerce client outside Foshan, China, came to me ahead of leaving a legacy WMS they had run for six years, with a warehouse of roughly 8,000 square meters and about 12,000 active SKUs. Before migration I asked them to spot-count 200 SKUs. Thirty-eight did not match the books, a variance rate near one in five. The root cause was duplicate codes. One bedside table existed under three different SKU codes because different buyers had created records in different years, so its stock sat split across three codes. Pickers knew from experience which pile was which; the system did not. A batch received in 2021 had also been moved to new locations while the system still carried the old location numbers, so the goods sat in one place and the record pointed somewhere else. Had they cut over on schedule, day-one accuracy in the new system would have dropped below eighty percent.
In September 2024, an auto-parts warehouse near Ningbo, China, showed me the other classic pattern. The building was smaller, around 4,000 square meters, but the history was messier. The books showed stock for more than 300 SKUs that, on the floor, turned out to be partly scrapped goods never written off, partly returns put back on shelves without inspection, and partly locations that showed stock and opened up empty. That last category, phantom inventory, does the most damage. Sales sees stock and takes the order; picking finds nothing to pick, and the customer remembers the cancellation, not your new system. Before cleansing, orders touched by phantom inventory shipped on time about thirty percent less often than clean orders.
Why WMS Data Cleansing Has to Happen Before the Switch
After go-live, every process runs on the new system's numbers, so the smoother the process, the faster an error travels. In the old system, dirty data survived because people compensated for it. A veteran picker knew which code was the duplicate, which location was really empty, which batch must not ship. A new system, a new hire, an order feed — none of them compensate; they read the record and act on it. WMS data cleansing exists to move what used to live in veterans' heads into clean records, so the new system can run on the books from day one.
There is also a cost argument, and it is not close. Fixing a bad record in the old system costs one person and one spreadsheet. Fixing it after go-live costs stopped lines, corrected orders, angry customers, and recounts, several times over. Teams that cleansed seriously before cutover typically logged roughly half as many first-month problem tickets as teams that skipped it — a difference in books, not software.
WMS Data Cleansing in Four Steps: Freeze, Count, Fix, Trial Cut
Step one is freezing the baseline. Pick a freeze date. From that day, nobody creates new SKU codes in the old system, nobody hand-edits stock levels, nobody bulk-imports historical documents. Every adjustment goes through an approval form with a paper trail. A freeze does not stop the warehouse from working; it stops the data from getting dirty again while you clean it, because without one, you clean one round and dirty the next.
Step two is a full count plus sampled re-checks. Do not count only fast movers. Slow-moving stock, the returns area, and the damaged-goods zone are exactly where dirty data concentrates, and they need full coverage. Count in pairs, one person counting and one recording, and resolve every variance on the spot instead of carrying it to the next day. After the count, I require sampled re-checks by zone of at least ten percent, and any zone with a sampled variance above roughly three percent gets fully recounted. One recount is far cheaper than a month of firefighting after go-live.
Step three is fixing each dirty-data type differently, because lumping them together guarantees mistakes. Duplicate SKUs get a master code: merge stock under the master, freeze the duplicates, and keep the old codes in history for past orders. Book-to-physical mismatches get adjusted to the counted quantity, with a reason code and an approver on every adjustment — no unexplained adjustments, ever. Dead locations split two ways: a location that no longer exists physically must have its stock moved out before it is deactivated, while a location the system shows but nobody can find on the floor gets suspended until someone traces its origin. Phantom inventory takes the most work, case by case: scrapped but never written off, returned but never inspected, or simply missed in a count. The cause decides the process fix, so skipping the cause means the same phantom comes back.
| Dirty-data type | How to spot it | How to fix it | What happens if you migrate it anyway | | Duplicate SKU codes | Same product and spec under several codes, stock split | Pick a master code, merge, freeze the rest | Pickers grab the wrong item, stock looks inflated | | Book-to-physical mismatch | Counted quantity differs from system quantity | Adjust to the count, with reason code and approval | Stockouts and overselling after go-live | | Dead locations | System location with no goods, or no matching floor spot | Move stock out, then deactivate, or suspend pending investigation | Tasks sent to the wrong location | | Phantom inventory | Books show stock, shelf is empty | Write off case by case by cause, and patch the process | Orders accepted that cannot be fulfilled |
Step four is a trial cut. Run about ten percent of real orders end to end in the new system first, from order receipt and allocation through picking to shipment. During the trial, reconcile daily across three numbers: new-system stock, old-system stock, and counted stock. Any gap above roughly one percent stops the trial until someone finds the cause. Passing a trial cut does not mean zero variances; it means every variance has an explanation, an owner, and a fix date. When the trial runs clean, then you set a real cutover date.
Five Things That Blow Up on Go-Live Night
One judgment surprises most clients: WMS data cleansing should not aim for one hundred percent clean. Chasing the last two or three percent of untraceable dead stock can stall a project for months. The right call is a go-live line, written down and signed by the boss, finance, and the warehouse together, not agreed verbally. Fast movers, anything with movement in the past year, and every SKU tied to an open order must be fully clean. Dead stock untouched for three years or more, in small amounts, can move into a clearly marked quarantine zone in the new system and be worked through after go-live. Shipping first and clearing the tail second beats polishing forever.
My go-live night list has five items, and I do not sign off with any of them missing. Inventory accuracy: sampled variance within roughly two percent, tighter for fast movers. Duplicates at zero: no unresolved master-duplicate relationships left, every duplicate code frozen. Location validity: every location holding stock confirmed on the floor, dead locations deactivated or suspended. Open orders resolved: old-system pending orders either shipped or given an explicit migration plan, never carried over half-understood. Rollback rehearsed: if the cutover fails, the old system's data and processes must be restorable within half a day, and that has to be practiced once, not just written in a plan.
The Ningbo auto-parts warehouse followed this sequence. Cleansing took about five weeks, two weeks longer than planned, and the owner complained about the delay at the time. In the first month after go-live, stockout cancellations fell roughly forty percent, pickers stopped guessing from memory, and wrong-shipment complaints dropped sharply. The Foshan home-goods warehouse saw a simpler win: merging duplicates cut active SKUs from about 12,000 to just over 9,000, and purchasing and operations could finally read reports without guessing which code was real. Both projects taught the same lesson: the cutover date is not the deadline, clean books are. Everyone pushing you to cut early is just moving the failure to after go-live, where it costs far more.
By Lin Xiao, WMS Implementation Consultant







