The global retail industry loses $1.7 trillion a year to inventory distortion, which is the combined cost of out-of-stocks and overstocks. The figure gets cited in conference slides and quoted in vendor decks, and then it moves to the abstract file in the reader's head. The math behind it is sound. Global retail sales run at roughly $26 trillion, and distortion accounts for 6.5% of that. But the shape of the number doesn't land until it's translated into the P&L of an individual retailer, where the result tends to be harder to escape.
That translation is also the basis for any serious conversation about enterprise AI ROI in retail. Most ROI cases are built in the other direction. A vendor projects savings against an unspecified baseline, the buying team builds a model that assumes the savings materialize, and the business case makes it through procurement on the strength of the projection rather than the strength of the baseline.
The problem with that sequence shows up eighteen months later, when the implementation has run longer than planned and nobody has the current-state numbers to compare against. The ROI can't be proven because the starting point was never properly measured.
The argument here is that retail inventory distortion is one of the cleanest baselines available for an enterprise AI ROI case, because the cost is already sitting in systems the organization controls. You don't have to project it. You have to measure it. And the number that comes back tends to be large enough to reshape the conversation about what the architecture is worth.
What inventory distortion actually costs a single retailer
Work the math on a specific example. A $500 million omnichannel retailer operating at the industry-average net margin of 3% earns roughly $15 million in profit a year. If inventory distortion runs at the industry rate of 6.5% of sales, the distortion cost on that business is $32.5 million. The cost of the problem is more than double the net profit. That isn't a cost line. It's a structural condition.
The two forms of distortion show up differently on the income statement, which is part of why the total picture rarely gets captured. Out-of-stocks account for $1.2 trillion of the $1.7 trillion global figure. They're an invisible revenue problem because there's no line item for a sale that didn't happen. A customer walked in ready to buy, didn't find the item, and left. That moment produces no signal, no exception report, and no aggregate number the planning team can look at the next day. The damage is real and the accounting is silent.
Overstocks account for the remaining $554 billion, and they're a margin problem rather than a revenue problem. Excess inventory doesn't sit at cost. It sits at cost plus the carrying expenses that accumulate daily (warehouse space, handling labor, insurance, and the capital tied up in goods that aren't moving).
When overstocked items eventually sell, they sell at a markdown. It's the consequence of a forecasting and replenishment process that consistently misjudges what to buy, in what quantities, for which locations.
The two forms aren't independent. A retailer chronically understocked on its best-performing SKUs is typically overstocked on slower movers at the same time, because the demand signal feeding both decisions is the same fragmented, delayed, inaccurate picture. The data problem produces both symptoms at once, which is why tackling them in isolation tends to move the numbers in offsetting directions.
Why most enterprise AI ROI calculations are wrong
The way enterprise AI ROI gets calculated in retail usually inverts the logic that would make the number defensible. A vendor presents a capability. The capability is associated with an improvement range, often drawn from case studies in other organizations. A number comes out. That number becomes the ROI projection.
The problem isn't the arithmetic. The problem is that the baseline the improvement is being applied to is usually a best guess. It's an estimate of current stockout rates, a rough figure on markdown exposure, and a planner-productivity assumption that nobody has actually measured. When the implementation goes live and someone asks whether the ROI is being realized, there's no clean way to answer.
Retail is one of the few environments where this is fixable before the program begins. The baseline cost of inventory distortion is measurable at the SKU, location, and time level, using data the organization already has. Markdown rates come out of the merchandising systems. Stockout frequency comes out of the POS and inventory systems. Working capital tied up in slow-moving inventory is visible in the finance stack. The enterprise AI ROI case that survives procurement review and still holds up at the one-year mark is the one anchored to those measurements, not to projections applied to estimates. Learn about how AI agents can fix inventory imbalance.
Baselining the current state properly
A serious baseline for a retail inventory AI program has four components, and organizations that do this work before engaging vendors have a more productive conversation about what the technology is actually worth.
- Direct distortion cost. Which is the revenue lost to stockouts plus the margin lost to markdowns, measured against the categories and locations where the program will focus. A well-instrumented baseline uses transaction-level data to assign lost-sales probabilities to observed stockout events, and markdown analysis to distinguish planned promotional markdowns from distortion-driven markdowns. The difference between those two categories is where the addressable cost actually lives.
- Working capital cost. Inventory that isn't moving is capital that isn't productive. And the financial cost of carrying it includes the opportunity cost of the capital itself alongside the direct warehousing and handling costs. A baseline that doesn't include working capital tends to underestimate the distortion cost by a meaningful margin.
- Planner productivity cost. Planning teams at retailers with fragmented data spend a significant share of their time reconciling sources, validating numbers before acting on them, and overriding system recommendations they know don't reflect reality. That time has a direct cost, because the work being displaced is the analytical work that would actually improve outcomes if the reconciliation weren't necessary.
- Execution speed cost. Decisions that sit in queues while they wait for approval, data validation, or cross-functional alignment lose value as they wait. Organizations that baseline execution speed against the cost of delayed decisions end up with a more complete picture of what the current state is costing them.
The four components together are what makes the enterprise AI ROI conversation specific. A buyer who can state the distortion cost, the working capital cost, the productivity cost, and the execution speed cost as current-state numbers has a basis for evaluating vendor claims that a buyer working from estimates doesn't.
Enterprise AI ROI Benchmark Report
Why $172 billion in investment hasn't moved the number
The global retail industry spent $172 billion on inventory-related improvements over the past year, according to IHL Group. And despite that hefty investment, the $1.7 trillion distortion figure didn't move meaningfully. That gap is worth understanding, because it describes the limitation of the approaches most retailers have taken and the reason most AI ROI projections fall short of what actually shows up.
Most of the investment has gone into better systems within existing functional categories. A more capable WMS. A more sophisticated demand planning tool. A BI layer that produces cleaner dashboards. Each of these investments improves an individual function. None of them address the cross-functional data problem that generates distortion in the first place.
What hasn't been funded at anything close to the same scale is the horizontal layer that lets those functional tools reason together from the same operational view. Until that horizontal layer is in place, the ROI on additional functional investment is capped by the accuracy of the inputs, and the enterprise AI ROI conversation inherits the same ceiling.
The ROI is already in the P&L
The $1.7 trillion figure describes a problem that is, on a per-retailer basis, already sitting in the systems the organization owns:
- Markdown data
- Stockout frequency
- Working capital
- Planner time
- Execution delays
These are measurable today. And the measurement is what turns an enterprise AI ROI projection into an enterprise AI ROI case. The retailers closing the gap between what their data knows and what their planning teams can act on aren't doing something the rest of the industry can't. They're doing the baselining work first, so the ROI math has somewhere to anchor.
Once it does, the conversation about architecture, deployment timeline, and vendor selection becomes grounded in the specific cost the program is meant to address. That's the conversation that produces durable returns. The speculative version produces pilots.
If you’re ready for durable returns, schedule a demo with our team.
.png)
