Over the past year I have sat through more than a dozen AI vendor pitches. The sentence I heard most often was not about accuracy — it was 'just give us your historical data.' Almost nobody asks the follow-up, and no vendor volunteers an answer: how much data does warehouse AI need, exactly, before it can do useful work? Based on public deployment cases and industry research I have been tracking, I would argue the question is pointed in the wrong direction. There is no single number. There is a threshold table that differs by scenario — and the real cliff is not at the pilot stage, but on the road from pilot to go-live.
How Much Data Does a Pilot Actually Need? Three Appetites
Warehouse AI falls into roughly three families, and each has a completely different definition of 'enough.'
The first family is visual recognition — inbound quality inspection, parcel label reading, unsafe-behavior detection. The pilot threshold here is the most concrete of the three, measured in labeled images. A safety-vision study published in November 2025 disclosed its full data ledger: one week of surveillance video from a logistics facility, sampled at one frame per thirty with blurred and occluded segments discarded, yielded 20,000 usable images — 12,000 for training, 8,000 for testing — enough to support a dangerous-action detection model. In other words, a single-point vision pilot sits in the range of several thousand to twenty thousand labeled images, collectable in as little as one to two weeks. The constraint is rarely collection time; it is whether someone has labeled what each image actually shows.
The second family is prediction and optimization — demand forecasting, replenishment, slotting. This family consumes time span, not image counts. IBM's forecasting documentation states a hard rule: history should cover at least twice the forecast horizon, and for yearly seasonality you need a minimum of 24 months before the model can detect the cycle. Industry practice converges on the same numbers: one year of clean transaction history is the floor, 18 to 24 months is comfortable, and SKU-level machine learning forecasts typically want two to three years of order-level history — plus usable metadata: promotion calendars, return flags, stockout markers. A distributor with eight months of ERP history has a forecasting model that has never seen a full peak season. No amount of hyperparameter tuning fixes that.
The third family is robotics and foundation models — robotic picking, AMR fleet orchestration. Here the data threshold has quietly moved off the customer's plate and onto the vendor's. Covariant's RFM-1, released in March 2024, is an 8-billion-parameter robotics foundation model trained, per IEEE Spectrum's reporting, on tens of millions of grasp trajectories from robots working in real warehouses — with fleets across 15 countries feeding back roughly a million new trajectories every few weeks. For comparison, the largest public academic dataset, RT-X, holds about one million trajectories. Plus One Robotics announced its billionth pick in February 2024, holding over 99% pick accuracy through human-in-the-loop operations — remote supervisors handling the long tail of exceptions, every correction becoming new training data. Locus Robotics tells the same story at fleet scale: 5 billion picks by April 2025, the last billion in just 24 weeks, across 350-plus sites, with 15,000-plus robots generating on the order of 30 GB of data per second. The implication for customers is straightforward: you do not need your own ocean of data; you need a vendor that has already drunk from one.
The three thresholds compress into a single table:
| Scenario | Pilot threshold (order of magnitude) | Go-live threshold (the real variable) |
|---|---|---|
| Visual recognition | Thousands to ~20,000 labeled images | Coverage of lighting, shift, and SKU variation across seasons |
| Prediction & optimization | 12–24 months of clean history | Samples of out-of-distribution events: new SKUs, promos, peak surges |
| Robotic picking | Vendor foundation model + tens of thousands of in-house adaptation picks | Whether the human-in-the-loop exception pipeline keeps turning |
The Real Data Threshold: Three Gaps Between Pilot and Go-Live
Once the pilot works, the harder test begins. Across the cases I follow, three gaps sit between pilot and go-live — and none of them is about raw data volume.
The first gap is distribution shift. A pilot typically covers one zone and one shift, a narrow data distribution; go-live means the whole building, around the clock. A 2026 IndustryToday analysis documented a textbook case: a manufacturer replicating an AI solution from its first plant to its second discovered that inconsistent data definitions and file structures turned a planned copy-paste into a custom project, roughly doubling the rollout timeline. Warehouses rhyme: change the rack height, the lighting, or the SKU mix in building two, and the pilot model can quietly fall apart.
The second gap is the missing answer key. Many warehouses are not short on data — years of WMS transaction logs, cameras recording around the clock. What they lack is labeled ground truth. Gartner's survey of 1,203 data management leaders found 63% lacking confidence in their data practices for AI; in manufacturing, only 27% of companies have a data warehouse or data lake at all (versus 60% of mid-market companies overall), and none of the manufacturers surveyed had adopted machine learning platforms. Data sitting in storage and data a model can learn from are two different things. A third-party benchmark makes the point sharply: the same frontier model answering business questions scored about 20% accuracy against raw tables, and 92.5% once a governed semantic layer sat in between. Different domain, same lesson — governed structure is worth more than raw tonnage.
The third gap is whether the data flywheel spins. Go-live, at its core, is the establishment of a loop: production data flows back, the model iterates. Plus One's Yonder model is worth studying — exceptions the robot cannot handle go to remote humans, and every human correction becomes a training sample. Once that loop turns, the data threshold changes from 'stockpile everything before launch' to 'keep growing after launch.' Without a backflow design, the model starts to rot within months — SKUs change, packaging changes, processes get retuned, and the model is frozen on launch day.
The Counterintuitive Part: Data Volume Is the Wrong Scoreboard
This is where I will venture a counterintuitive judgment: pilot-stage data needs are systematically overestimated, while go-live data needs are systematically underestimated.
The overestimation first. Many vendor presales teams ask for 'at least two years of history,' but validating a single hypothesis rarely requires it. One freight-audit AI provider has written publicly that a first pilot needs just 50 to 200 historical invoices, paired with TMS records and exception notes, to expose the main patterns. Vision pilots follow the same logic: twenty thousand labeled images and a week or two of collection can answer 'does this work in our building.' A pilot needs the minimum dataset that can validate the hypothesis — not the entire family archive. This is also where 'how much data does warehouse AI need' gets most misread: people keep asking about volume, while the answer hides in structure.
The underestimation is where the real money goes: cross-site adaptation, long-tail labeling, continuous retraining after launch. Locus went from its billionth pick to its five-billionth in under three years — the exponential growth in data happened after go-live, not before. Put differently, the data threshold is not a wall you climb once; it is a runway you keep paying for. The pilot asks where the starting line is; go-live asks whether you can keep running.
There is also a structural shift worth watching on a 2026–2027 timeline: foundation models are moving the threshold from 'bring your own data' to 'vendor brings the base, customer adapts.' RFM-1 can claim to grasp unseen items precisely because it trains on trajectories pooled across customers and countries — no single site starts from zero. Gartner's numbers corroborate the direction: companies were never short on raw data; they were short on the capability to turn data into something a model can use — and that capability is increasingly being productized by vendors.
So, back to the opening question. How much data does warehouse AI need? Perhaps the better formulation is three questions: which of the three families is your scenario in? Is your answer key — the labeling — ready? Is your data flywheel designed? If those three have answers, data volume itself is rarely the binding constraint.
And a longer-term question worth sitting with: when foundation models arrive with their own data and 'how much data you have' stops being a moat, what becomes a warehouse's truly irreplaceable data asset? The human-in-the-loop pipeline that turns exceptions into samples? The multi-season record of real operations? Something else entirely? I have no answer yet — which is exactly why I will keep watching.
By Dr. Chen, Supply Chain Researcher
