Let me be blunt: in the AI era, the most valuable thing in your warehouse isn't the AMRs or the tall racking. It's the decade of data sitting in your WMS — your data assets, whether you treat them that way or not.
I get asked all the time: "Lao Mi, should we invest in AI?" My first question back is always the same: are your receiving timestamps accurate? Nine times out of ten, silence.
Here's the thing, and it cuts both ways. On one hand, AI genuinely works in warehouses: demand forecasting, wave optimization, labor planning, exception alerts — use cases everywhere you look. On the other hand, every one of those use cases is fed by data. Garbage in, expensive garbage out. More data is not better data. Dirty data costs more than no data — with no data, at least you know you're flying blind; dirty data makes you confidently wrong.
I've watched too many of these crashes. The one I remember best: a 3PL in California, a decent-sized operation, mostly e-commerce clients. Back in 2023 they spent serious money on an AI labor-forecasting system that built staffing schedules automatically from historical volume. Sounded reasonable. Three months in, the schedules got worse instead of better. The week before peak season it fell apart completely — they had to call a temp agency for emergency staffing at a painful premium.
The postmortem was embarrassing. The algorithm was fine; the data was trash. Their receiving team had a habit of batch-entering receipts right before clocking out, so every receiving timestamp in the system clustered between 5 and 6 PM. The AI learned exactly what it was taught: a volume surge hits every day at 5 PM, so staff up for it. In reality the freight arrived in the morning — 5 PM was just when someone typed it in. The model learned the wrong lesson with total confidence.
Worse, their wave release times and pick completion times were all the same kind of backfilled timestamps. The entire time dimension of their WMS was fiction. The AI project died, the money was gone, and management walked away convinced that "AI is a scam." Let me be blunt: AI didn't die. The data did.
It wasn't an isolated case. An apparel warehouse I know ran a single return-reason code for years: "other." When they later wanted AI to predict returns and streamline QC, they pulled the data and found the vast majority of rows said "other." The algorithm had nothing to learn from. Blame the person who designed the reason codes for convenience — and blame everyone who never treated data as an asset.
What Are Your Data Assets Actually Worth
Walk through the table below. Left column: the data. Middle: the AI use cases it feeds. Right: what you're probably missing. Run through it once and you'll know whether you're sitting on assets or liabilities.
| Data domain | AI use cases it feeds | Typical quality problems |
|---|---|---|
| Order line detail (SKU, qty, service promise) | Demand forecasting, wave strategy | Promotions not flagged, history distorted |
| Inventory transactions (receipts, shipments, adjustments, cycle counts) | Safety stock, dead-stock alerts | Adjustments with no reason codes |
| Task timestamps (receive, putaway, pick, pack) | Labor forecasting, bottleneck analysis | Backfilled or batch-entered times |
| Slotting velocity (ABC movement, utilization) | Slotting optimization, replenishment | ABC classes untouched for a year |
| Labor hours (who did what, for how long) | UPH baselines, schedule optimization | Hours not tied to tasks |
| Equipment events (AMR alerts, charging, faults) | Predictive maintenance | Alert logs nobody reads |
Most warehouses finish this exercise with an uncomfortable finding: plenty of data, not much of it usable. That is what a data assets audit really means — not counting how many tables you have, but checking how much of your data can survive an algorithm's interrogation.
Before Fixing Your Data Assets, Do the Math
This cuts both ways. Fixing data costs real money: real-time scan confirmation at receiving adds seconds to every receipt; wave releases and pick completions need hard timestamp discipline; ABC classes need periodic recalculation; labor hours must be tied to tasks. Every one of these adds steps for the floor team, and in the short term, throughput dips. The shift supervisor will be the first to complain: we're already buried in volume, and now you want more scanning?
But not fixing it costs more. Every backfilled timestamp is a landmine you're planting for your future AI project — and it detonates the day you spend real money on the system. That California 3PL's lesson, translated plainly: the interest on data debt is higher than you think. The dead project, the emergency temp labor, the management team's lost faith in AI — all of it added up to far more than the cost of "one extra scan."
Do the ROI math on the whole ledger, too. In an AI project's budget, the algorithm and software licenses are usually the small part; implementation, integration, and rework are the big part. Bad data quality and rework can eat more than half your budget. It's cheaper to pour the data foundation now than to pay for the fire department later. Every owner understands this math — they just don't like spending money on things they can't see. And that is the most expensive kind of saving.
My advice: three steps, and don't try to do it all at once.
First, protect time and reason. Every task timestamp must be captured in real time — no end-of-shift backfilling. Every inventory adjustment needs a reason code. Lowest cost, highest return. Get the foundation straight first.
Second, tie people to tasks. Labor data that isn't linked to specific tasks is a mess. Scan in and out, dispatch tasks to individuals — get to the point where "who did how much" is traceable. Only then do UPH baselines and schedule optimization have data worth talking about.
Third, then talk AI. Accumulate data across several business cycles — at least one full peak season — before talking prediction and optimization. Forecasting next year's peak on three months of data isn't AI. It's gambling.
Don't Buy the Algorithm Before You Pour the Foundation
When an AI vendor's salesperson tells you how great their algorithm is, ask one question: what's your model's tolerance for dirty data? Watch how they answer.
Here's the blunt version: in the AI era, warehouse competition has two layers. The top layer — algorithms and applications — everyone will eventually be able to buy. The bottom layer — data assets — can't be bought. You can only accumulate them yourself. Whoever has clean, continuous, business-meaningful data wins with AI. The reverse is equally true: a warehouse with a data mess will burn money on the most expensive AI.
So stop asking "should we invest in AI?" Ask three questions instead: Are our timestamps accurate? Do our adjustments carry reasons? Have we accumulated data across several peak seasons? If you can't answer even one of them, the answer is clear: audit your data assets first, then talk AI.
Your warehouse data is either an asset or a liability. Have you audited yours?
By Lao Mi, Supply Chain Veteran (20 yrs)
