We had a client — a 3PL in East China running fulfillment for FMCG brands — that spent seven figures last year on an AI demand-forecasting system. The sales demo was gorgeous: feed in historical sales, get an 8-week replenishment recommendation for every SKU, "over 90% accuracy" on the slide. The owner signed on the spot.
Three months after go-live, their warehouse manager called me to vent. The system's replenishment suggestions were psychic on a good day and sabotage on a bad one. Once it recommended stocking three months of a paper-towel SKU — because that SKU had just been repackaged, the old-packaging SKU was still alive in the system, and the AI added both sales histories together and concluded "this product is on fire." The warehouse nearly drowned in paper towels.
I told the owner straight: your AI project didn't die on the algorithm. It died on the data. You bought an imported engine and filled it with gutter oil.
The insider truth no AI vendor volunteers
I've been selling 3PL services long enough to have signed something like a hundred warehouse contracts, and I've watched plenty of AI projects crater. Here's an insider truth from the vendor side of the table: no AI vendor seriously audits your data quality before the contract is signed.
Why would they? An audit might kill the deal. Sales needs the signature; data quality is "an implementation-phase problem." The contract's data clause is usually one sentence: "Client shall provide necessary business data support." What counts as necessary? What does support mean? How clean is clean enough? None of it's defined. Then go-live underperforms, the vendor shrugs: your data is bad — garbage in, garbage out. Technically true. But the bill lands on you, because you didn't negotiate data responsibility when you had leverage.
The worst case I saw: an auto-parts distributor in South China, 50,000+ SKUs, and I'd estimate a 15% duplicate-code rate — the same brake pad had one code in purchasing, another in the warehouse, a third in the e-commerce team. After the AI replenishment system went live, the model treated three codes as three products, each with thin sales history, and every forecast was noise. The vendor sent an implementation consultant to "cleanse the data," billed by the day, spent two months at it, and the cleansing invoice ran into the hundreds of thousands of RMB. The owner was livid: the data cleanup cost more than the software.
Bluntly: data preparation for an AI project is money the vendor will never spend for you. Either you spend a little of your own money cleaning data before you sign, or you spend a lot of money paying the vendor to clean it after. That money is unavoidable — the only choice is whether you spend it proactively or reactively.
The four chronic diseases of warehouse data — see which ones you have
Before any cleanup, give your warehouse a checkup. Four ailments, and nine warehouses out of ten have at least two:
- One product, many codes. The same item lives under multiple SKU codes across purchasing, warehousing, and sales. The AI sees three "products" where there's one.
- Chaotic location codes. The old building uses one scheme, the new building another, and the WMS has a third. One client ran three location systems in parallel — after the AMRs went live, they threw "goods not found" errors daily while robots circled the floor and humans cleaned up behind them.
- Missing historical documents. Forecasting models want 12–24 months of clean sales history. In many warehouses that history lives in sales reps' spreadsheets — and walks out the door with every resignation.
- Excel vs. system: two versions of the truth. The WMS says one thing, the shift lead's spreadsheet says another. Cycle counts run on Excel; the system runs on "roughly." The AI only ever sees the "roughly."
The three-step cleanup: fix the data before you buy the AI
Step 1: Master data governance — one product, one code. Start here.
This is the foundation. Form a three-person team — one each from warehousing, purchasing, and IT (or your system admin) — with a single mission: SKU master-data unification. Concretely: export every SKU, fuzzy-match on name, spec, and barcode, pull suspected duplicates for human review, designate one canonical code, map the rest to it, and forbid new alias codes going forward. Then set the rule: every new SKU goes through an admission process — no duplicate check, no code creation. Whoever creates a rogue code owns the consequences.
This typically takes 4–8 weeks. Don't rush it. A shaky foundation means everything above it is a hazard.
Step 2: Backfill process data — recover your history.
AI eats historical data, so feed it history. Pull 24 months of inbound/outbound records, sales, promotions, and returns — recover whatever you can. Those Excel ledgers aren't trash, they're a goldmine — get an intern or an outsourcing crew to key the critical fields into the system in structured form. For months you genuinely can't recover, mark them "missing" honestly. Never fabricate data to fill gaps. Models tolerate sparse data; they do not tolerate fake data. A model handles declared gaps gracefully; invented numbers lead it straight into a ditch.
Step 3: Data quality gates — keep dirty data out.
Cleansing is a one-time event; gates are permanent. Configure three hard rules in your WMS: new SKU creation must pass a duplicate check (barcode/name similarity warning); receiving must scan and validate — unscannable codes can't be waved through with "manual entry, done"; and run a monthly data-quality report tracking duplicate-code rate, location anomaly rate, and document-missing rate, with red lights and named owners when thresholds break.
Finish these three steps and your warehouse data quality reaches the passing grade for "AI-ready." The process usually takes 3–6 months and costs mostly labor — an order of magnitude cheaper than the AI software itself. But it decides whether your seven-figure software budget bought a money printer or an expensive paperweight.
The counterintuitive call: the worse your data, the less you should buy AI first
One last contrarian point: plenty of owners think "our data is a mess, so we need AI even more to manage it." The logic is backwards. AI is a data amplifier, not a data janitor. Clean data, and AI multiplies your efficiency tenfold; garbage data, and AI multiplies your errors tenfold — while charging you seven figures for the amplification.
The order is never negotiable: get the data to a passing grade first, then let AI in. And when a salesperson tells you "our algorithm handles dirty data automatically," ask him to write that sentence into the contract's performance commitments and watch whether he dares.
Next time you negotiate an AI contract, ask these three questions:
- Who owns data preparation, and what exactly counts as "ready"? Put acceptance criteria in a contract appendix, quantified: duplicate-code rate, document completeness rate, concrete metrics.
- If the model underperforms, how is payment tied to results? Holdback percentages, performance guarantees, pilot periods — every line in writing. Verbal accuracy promises are worth zero ink.
- How long is the pilot, and what data validates it? It must be back-tested on your own real historical data — never the vendor's demo dataset. Demo datasets never contain duplicate SKU codes.
Remember: in AI, the vendor sells shovels and you're the one digging for gold. The sharpest shovel is useless if the mine is full of rocks. Pick up the rocks first, then argue about shovels — get that order right and you may save yourself two years of tuition.
By Susan, 3PL Sales Director
