Let's start with the conclusion: AI data preparation is where most warehouse AI projects are won or lost, and it has almost nothing to do with choosing a model. Across more than 30 WMS go-lives I have led, the same pattern repeats. A team spends its budget on an algorithm demo, then plugs in production data and discovers that half the location codes do not match, inventory counts disagree with the books, and two out of ten historical orders are missing key fields. The model is not wrong. It simply has nothing solid to work with.
In the fall of 2024 I was on site at a 3PL warehouse in eastern China. The client wanted AI replenishment suggestions, and the demo looked great. The moment we connected real data, the whole thing fell apart: the same product carried three different codes in three systems, half the pick records had no location information, and a fifth of the outbound orders from the past six months were missing weight data. We stopped the project for six weeks just to fix data. I have seen this exact failure three times now, and the script is always the same. The business side wants to see results fast, the technical side wants to run a demo fast, and data preparation gets treated as something to do on the side. Side jobs always come back to collect double.
AI Data Preparation Means Three Different Kinds of Data
AI data preparation covers three kinds of data, and mixing them up is how teams waste weeks. The first is master data: products, locations, suppliers, customers. Master data is about uniqueness. One product gets one code and one name. One location gets one coordinate and one status. When master data is dirty, every judgment the AI makes is built on a wrong foundation. The check is simple: export the product table and deduplicate it separately by name, specification, and barcode. Merge the duplicates before doing anything else.
The second kind is transaction data: receiving orders, shipping orders, transfer orders, count sheets. This is the only textbook the AI has for learning how a warehouse actually runs. Transaction data is about completeness. An order without a timestamp never happened. A pick record without a location has no path. My method is monthly sampling: pull 100 orders at random each month and trace each one end to end. If the missing-field rate passes one in ten, that month's data stays out of the training set.
The third kind is status data: inventory snapshots, location occupancy, equipment status. Status data is about freshness. The AI needs the warehouse as it is right now, not as it was last month. Many warehouses refresh inventory snapshots once a night in a batch job, so changes made during the day do not show up until tomorrow. Recommendations built on that kind of data arrive after the goods have already moved.
Each kind of data goes dirty in its own way. Master data gets messy, transaction data goes missing, status data goes stale. Cleaning them with one single process is the same as not cleaning them at all.
The Order Matters, and Getting It Backwards Is the Classic Mistake
The right order is master data first, then transaction data, then status data. Master data is the dictionary. If the dictionary is wrong, every translation after it is wrong. I watched one project spend three weeks cleaning historical orders, only to find the product coding scheme had changed twice and half the codes in those orders were already retired. Three weeks wasted.
What counts as "done" for master data? My go-live checklist has three hard gates: product codes are one hundred percent unique, location codes match physical locations one to one, and supplier and customer names are deduplicated. If those gates are not met, do not move on. Moving on just builds interest on the rework.
For transaction data, there is a shortcut: do not try to clean all of history. Take the most recent continuous stretch of good-quality data, say the last three months, and demand that it be complete and consistent. AI learns patterns, not archaeology. Three months of clean data beats two years of half-dirty history.
Status data is the hardest, because the test is not a one-time cleanup but a lasting capability. Before go-live, ask one question: can an inventory change show up in the system within an hour of happening? If not, the AI's suggestions will always be one step behind. No AI deployment fixes that gap. It only dresses up the lag.
Do the Math: What AI Data Preparation Saves You
Take a mid-size warehouse as an example: 500 orders a day, three lines per order on average. That is 45,000 transaction lines a month. With a missing-field rate of one in ten, that means 4,500 broken records every month, and more than 50,000 in a year. Once the AI is live and starts giving strange suggestions, fixing those 50,000 lines costs several times what it would have cost before go-live, because the business is already running on the dirty data and cannot stop.
So data preparation deserves its own budget line and its own schedule. In my experience it should take at least a third of the whole AI project timeline. Anything less, and the project is likely to come back for rework within two months of going live. Managers often frown at that ratio because data preparation produces nothing visible. That is true, and it is also what decides whether everything after it is real.
Here is a counterintuitive judgment: more data is not better data. Many warehouses open with "we have five years of history," which sounds rich. Open it up and you find two system changes, three coding-rule revisions, and six months of manual ledgers in between. Five years like that is worth less than three months of consistent data. AI is not afraid of small data. It is afraid of data that contradicts itself.
There is also an ownership problem that gets ignored. The business side thinks data is a technical job, the technical side thinks it is a business job, and in the end nobody owns it. At every project kickoff I write it down: the business owns master data, operations owns transaction completeness, and the system team owns status freshness. Three owners, one weekly count check. Unclear ownership always ends in a messy ledger.
Two more failure points I keep seeing. One is a coding-rule change that nobody announces. A product code gains a digit, a barcode scheme changes, and the AI keeps recognizing goods by the old rules. Every rule change goes into the change log and to every stakeholder the same day, not just a message in a small chat group. The other is test data mixed with production data. Test orders, training orders, and drill orders mixed into transactions teach the AI patterns that never happened. Clear them out before go-live and keep a separate test environment. Sharing one environment to save effort is a false economy.
My go-live checklist ends with the same line every time: find a frontline picker, show them one AI suggestion, and ask, "Does this look like our warehouse?" If they nod, data preparation is done. If they shake their head, go back and keep cleaning. The model will never tell you the data is dirty. Only the people on the floor can.
By Lin Xiao, WMS Implementation Consultant








