TL;DR & Quick Summary
Inventory distortion — the combined cost of stockouts and overstocks — reached $1.77 trillion globally in 2025 according to IHL Group, which has tracked the figure for 18 years. Roughly two-thirds of that is lost sales from empty shelves ($1.2 trillion), and one-third is the carrying cost of stock nobody wanted ($572 billion).
Both halves are forecasting failures in opposite directions, and both are addressable with the same toolset.
McKinsey's research on AI-driven forecasting reports:
- 30%–50% reduction in forecast error
- 20%–30% reduction in inventory levels (up to 50% in some operations)
- Up to 65% reduction in lost sales from stockouts
- 5%–10% lower warehousing costs and 25%–40% lower administrative costs
The catch is that these gains come from wiring forecasts into ordering decisions, not from producing better dashboards. Most predictive analytics projects fail at exactly that handoff.
- Key Takeaway: Forecast accuracy is worthless in isolation. Value appears only when the forecast and its uncertainty automatically set reorder points, order quantities, and safety stock.
- Get Started: Want to know whether your SKU mix justifies predictive inventory work? Schedule a Strategy Call with Cogniq AI or explore our predictive analytics services.
Working in a service business rather than retail or distribution? The economics are different — parts consumption, technician utilisation, and job-driven demand behave unlike SKU-level retail sell-through. Our AI demand planning guide for service businesses covers that case directly.
Why Traditional Inventory Management Breaks
Most inventory systems still run on parameters set by a person, reviewed on a calendar, and applied uniformly. Three structural weaknesses follow.
1. Static Reorder Points in a Non-Static World
A reorder point set in January assumes January's demand rate and January's lead time. By June both have moved, but the parameter has not. The result is predictable: overstock on decelerating SKUs and stockouts on accelerating ones, simultaneously, in the same warehouse.
This is why businesses frequently report high total inventory and frequent stockouts. Those are not contradictory symptoms — they are the same symptom, which is that stock is allocated by outdated assumptions rather than current demand.
2. Safety Stock Set by Rule of Thumb
"Two weeks of cover" is the most common safety stock policy and the most expensive. It over-protects stable, predictable SKUs and under-protects volatile ones, because it ignores the only variable that matters: demand variability relative to lead-time variability.
Correct safety stock is a function of forecast error, lead-time variance, and your target service level. Applied uniformly, a flat rule guarantees you are wrong on nearly every SKU — just in different directions.
3. Stockouts Are Invisible in the Data
This is the subtle one, and it corrupts almost every forecasting attempt built on raw sales history.
When an item is out of stock, recorded sales are zero. A model trained on that history learns that demand was zero, when in fact demand was unmet. The forecast comes back low, the reorder point drops, and the next stockout arrives sooner. The model learns to starve exactly the SKUs that sell fastest.
Unless historical stockouts are labelled and demand is censored-corrected, a predictive system will actively make availability worse on your best-moving lines. This single issue explains a large share of inventory AI projects that quietly underperform.
What a Predictive Inventory System Actually Does
A working system has four layers. Skipping any one of them is the usual cause of failure.
| Layer | Function | Common Failure Mode |
|---|---|---|
| Data foundation | SKU-level history, stock positions, lead times, stockout flags | Unlabelled stockouts corrupt demand signal |
| Demand forecast | Per-SKU, per-location prediction with confidence interval | Point forecast only, no uncertainty estimate |
| Inventory policy | Converts forecast + uncertainty into reorder point, order quantity, safety stock | Forecast produced but never wired to ordering |
| Execution | Generates purchase orders and replenishment tasks | Requires manual re-entry, so nobody uses it |
The second column of the third row is where nearly all the value sits. A forecast that a planner reads and then overrides from memory has changed nothing. Our AI integrations practice treats the write-path into the ERP or inventory system as the first requirement, not the last, precisely because a read-only forecast is a report rather than a system.
Segment Before You Forecast
Not every SKU deserves the same treatment. The standard approach is a two-axis classification:
- ABC by revenue or margin contribution — how much the SKU matters.
- XYZ by demand variability — how predictable it is.
| Segment | Characteristics | Appropriate Policy |
|---|---|---|
| AX | High value, stable demand | Tight automated reorder points, low safety stock |
| AY / AZ | High value, volatile demand | Statistical safety stock, frequent review, highest model investment |
| BX / BY | Moderate value | Automated policy, periodic human review |
| CZ | Low value, erratic demand | Simple min/max or make-to-order; do not model |
The last row is the one that saves money. Applying sophisticated forecasting to thousands of low-value, erratic SKUs consumes engineering effort and produces noise. Concentrate modelling where variability is high and the item matters.
How Predictive Analytics Optimises Shelf Replenishment Cycles
Shelf replenishment deserves separate treatment, because it operates on a different clock than warehouse replenishment and is the point where stockouts actually cost sales. IHL Group attributes $690.9 billion of global losses specifically to empty shelves — the single largest line in the inventory distortion breakdown.
Traditional replenishment is interval-based: staff walk the floor on a fixed schedule and restock what looks low. This fails in two directions. Fast-moving facings go empty between passes, while slow-moving ones are checked repeatedly for no reason.
A predictive approach replaces the interval with a depletion forecast per facing.
The Mechanics
Step 1 — Model rate of sale at facing level. Point-of-sale data gives units sold per hour per SKU per store. Layered with day-of-week and hour-of-day patterns, this produces an expected depletion curve rather than a daily average — the distinction that matters when a facing empties at 2pm on a Saturday.
Step 2 — Estimate time-to-empty. Given current shelf quantity and the depletion curve, the model predicts when each facing crosses its minimum presentation threshold.
Step 3 — Incorporate demand shifters. Promotions, local weather, nearby events, and competitor stock positions materially change rate of sale. Weather is unusually predictive in grocery and convenience: a forecast temperature swing reliably moves specific categories, and it is knowable days ahead.
Step 4 — Sequence tasks by cost of not acting. This is the step that converts a forecast into value. Rather than a list sorted by aisle, staff receive a queue sorted by expected lost margin if the facing goes empty before the next pass. A high-velocity, high-margin facing 40 minutes from empty outranks a slow mover that has been low for two days.
Step 5 — Close the loop. Compare predicted versus actual depletion, feed the error back, and retrain. Without this, accuracy decays silently as assortment and customer behaviour drift.
What Changes Operationally
| Dimension | Interval-Based | Demand-Triggered |
|---|---|---|
| Trigger | Fixed schedule | Predicted time-to-empty |
| Task order | Walk route / aisle order | Expected lost margin |
| Labour profile | Flat across the day | Concentrated before predicted depletion |
| Failure mode | Fast movers empty between passes | Forecast error on volatile SKUs |
| Measurable output | Shelves walked | On-shelf availability % |
The last row is the important one. Interval replenishment measures activity. Demand-triggered replenishment measures the outcome customers actually experience — and on-shelf availability is a number you can put in a business case.
Building the Business Case
Predictive inventory work competes for budget against every other automation project. Build the case with four numbers, in this order.
1. Current inventory distortion cost. Estimate both halves. Stockout cost is lost units multiplied by unit margin — recoverable from POS gap analysis or unfulfilled order reports. Overstock cost is excess inventory value multiplied by carrying rate, typically 18–25% annually once capital, storage, insurance, shrinkage, and obsolescence are included.
2. Realistic improvement, discounted. Apply the McKinsey ranges to your baseline, then discount them. Those figures come from organisations with clean data and executive sponsorship. If your stockout history is unlabelled and your lead times live in a spreadsheet, model the low end.
3. Total cost of ownership. Data engineering is almost always the largest line — usually larger than the modelling. Add ongoing model maintenance, monitoring for accuracy decay, and the integration work to write orders back into your ERP.
4. Probability of reaching production. Gartner forecasts that over 40% of agentic AI projects will be cancelled by end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Inventory projects are especially exposed to the second cause when they stop at the dashboard. We work through this discounting method in detail in our guide to AI agent ROI benchmarks for 2026.
A Phased Implementation Path
Phase 1 — Instrument (weeks 1–4). Establish the baseline. Measure current on-shelf availability, stockout frequency by SKU class, inventory turns, and excess stock value. Critically, begin labelling stockouts so future models can correct for censored demand. Nothing here requires AI, and skipping it makes every later claim unprovable.
Phase 2 — Segment (weeks 3–6). Run ABC/XYZ classification. This alone frequently surfaces immediate wins — high-value volatile SKUs carrying rule-of-thumb safety stock, and low-value SKUs consuming disproportionate working capital.
Phase 3 — Forecast in shadow mode (weeks 5–12). Run predictions alongside existing processes without acting on them. Compare forecast to actual weekly. Shadow mode is how you discover data problems before they become ordering problems, and it produces the accuracy evidence needed to justify Phase 4.
Phase 4 — Automate policy (weeks 10–16). Let the system set reorder points and safety stock for the segments where shadow-mode accuracy proved out. Keep humans approving purchase orders above a value threshold.
Phase 5 — Close the execution loop (ongoing). Automatic PO generation for trusted segments, replenishment task sequencing at store level, and continuous retraining. This is where custom AI agents earn their cost, because the value is in acting on the forecast rather than displaying it.
Common Failure Modes
Forecasting without censoring correction. Covered above, and worth repeating: it is the most common and most damaging error in the category.
Optimising the wrong metric. Forecast accuracy ("MAPE went from 32% to 21%") is not a business outcome. Inventory turns, on-shelf availability, and lost margin are. Teams that report accuracy improvements without operational improvements have usually built a model that nobody acts on.
Ignoring lead-time variability. Most systems treat lead time as a fixed number. It is a distribution, and its variance drives safety stock as much as demand variance does. A supplier averaging 14 days with a range of 7 to 28 requires materially more cover than one that reliably delivers on day 16.
Automating a broken process. If stock records are inaccurate at the shelf, no forecast can help — the model is predicting against numbers that do not describe reality. Cycle-count accuracy is a prerequisite, not an enhancement.
Conclusion
The $1.77 trillion inventory distortion figure persists not because forecasting is impossible, but because most organisations stop one step short of the decision. They buy a forecasting tool, improve accuracy, produce a dashboard, and leave the ordering decision exactly where it was — with a planner overriding the model from memory.
The organisations capturing McKinsey's 20–30% inventory reductions did something narrower and harder: they labelled their stockouts, segmented their SKUs, proved accuracy in shadow mode, and then let the system set parameters automatically for the segments where it had earned trust. The modelling was the easy part.
Start with the baseline you can measure today. If you cannot state your current on-shelf availability or your excess stock value, that is the first project — and it is a prerequisite for proving anything that follows.
Book a strategy call with Cogniq AI and we will assess your data readiness before recommending a build. You can also see what our R&D team is prototyping at Cogniq Labs.
