AI improves inventory planning in a narrow and specific way: it produces better demand estimates at aggregate levels, and better estimates faster than a planner working in a spreadsheet. It does not improve the two things that actually cause stockouts and overbuys in a multichannel business, which are lead time uncertainty and cost data that was never right. A forecast is one input into a reorder decision. It is not the decision.
The research on this is unusually clear, which is rare in any conversation about AI, so it is worth looking at before spending money.
The M5 competition, reported by Makridakis, Spiliotis, and Assimakopoulos in the International Journal of Forecasting in 2022, set the benchmark for retail demand forecasting. Entrants forecast 42,840 hierarchical time series of Walmart unit sales, and the competition required 30,490 point forecasts at the lowest cross-sectional level, meaning individual item and store combinations.
The organizers' companion paper, "The M5 competition: Conclusions," reported that complex global machine learning methods delivered significant accuracy and uncertainty improvements over the statistical benchmarks, and that those improvements diminished at lower hierarchical levels and at the tails of the uncertainty distribution.
Read that carefully, because it is the whole story for a seller. Machine learning wins on the aggregate: total company demand, a category, a region. The advantage shrinks exactly where reorder decisions are made, at one item in one location, and shrinks again in the tail, which is where a stockout or a write-off lives.
A single SKU on a single channel produces a sparse, spiky series. Days with zero units. A day with 380 because a deal ran. A dead week because the listing lost the buy box. Ranking changes, ad spend changes, competitor pricing, a review that went sideways.
Multichannel makes it worse. The same SKU has four demand series with different shapes. Amazon demand responds to advertising and rank. Shopify demand responds to email sends and paid social. Walmart demand often has a different weekly seasonality. eBay demand can be steadier and lower.
Aggregating those into one company-level forecast produces a smooth, forecastable series. Disaggregating back into per-channel reorder quantities loses most of that accuracy, and the disaggregation is what you actually need.
Speed and coverage. A planner can maintain careful forecasts for 40 SKUs. A model maintains them for 4,000. Even mediocre coverage across the long tail beats no coverage, which is what most sellers have below their top 50 items.
Pattern recognition on promotions and events. Models absorb the effect of a price change or a deal event faster than a human updating a spreadsheet.
Uncertainty ranges instead of point estimates. Good models produce a distribution. That is more useful than a single number, because safety stock is a function of variance, not of the mean.
Surfacing the exceptions. The most valuable output is often not the forecast. It is the list of 30 SKUs whose recent demand no longer resembles their history.
Lead time. Your supplier's actual lead time distribution is the largest single driver of how much inventory you have to hold, and no model can observe it from your sales history. If your supplier quotes 45 days and delivers between 38 and 71, that spread does more damage than a forecast error of 15 percent.
Cost. A reorder decision is an allocation of cash. If landed cost is wrong because freight and duty sit in an operating expense account, the model will happily recommend buying more of the SKU that looks profitable and is not.
Physical truth. A forecast against an inventory position that does not match the warehouse is arithmetic on a fiction.
One SKU, sold on Amazon and Walmart. Landed cost $11.40. Current on-hand 4,180 units across two locations, plus 900 in transit.
Trailing 12-week demand averages 1,240 units per week across both channels, with a standard deviation of 310. Supplier lead time is quoted at 45 days and has run between 38 and 71 days on the last six purchase orders.
Scenario one: forecast error. A model cuts forecast error from 22 percent to 15 percent. Applied to a 45-day lead time demand of roughly 7,970 units, the improvement in expected demand accuracy is about 558 units, or $6,361 of inventory at landed cost.
Scenario two: lead time variance. Holding the forecast fixed, the difference between a 45-day lead time and a 71-day lead time is 26 extra days of coverage, or roughly 4,606 units, or $52,508 at landed cost. That is the cash you must either carry or lose to a stockout.
The lead time gap is eight times the forecast improvement. A seller who buys a forecasting tool and does not tighten supplier terms or track actual receipt dates has optimized the smaller number.
This is the reason a restock report should incorporate lead times and inbound stock rather than sales velocity alone. The inventory layer inside ConnectBooks drives its restock report from sales history and velocity with lead times and inbound stock factored in. Seasonality is not part of that calculation today, which matters if your demand is holiday-shaped, and is worth knowing before you rely on it in October.
Forecast at the level you can measure, decide at the level you buy. Most sellers buy at the SKU level and sell across four channels. Build the demand view by channel, aggregate to the purchase decision, and keep the channel split visible so you can see which one moved.
Attach a confidence range to every number. A single-point reorder quantity encourages false precision. A range forces the conversation about how much variance you are willing to carry.
Track forecast bias separately from forecast error. Consistently forecasting high is a different disease from forecasting inconsistently, and only the second is fixed by a better model.
Measure the model against a naive baseline. Last four weeks of demand, extended. If a tool cannot beat that on your data, it is not adding anything, and running that comparison takes an afternoon.
Recalculate after every receipt, not on a schedule. Inbound stock arriving changes the answer more than a weekly refresh does.
None of this works on top of unreliable inventory positions or wrong unit costs. Before any forecasting layer earns its cost, four things have to hold.
Stock quantities by location that match a physical count, including at marketplace fulfillment centers. Landed cost in the unit, so a reorder recommendation reflects the cash actually committed. Transfers between locations tracked as in transit until received, so units are not counted twice or lost. Channel-level demand carried through to the ledger, so profit by channel and SKU is visible and the forecast can be checked against contribution rather than revenue.
For a seller running Walmart alongside other channels, the fee structure differences alone can flip which SKU deserves the next purchase order, which is why Walmart accounting detail belongs in the same conversation as demand planning.
Treat AI demand forecasting as a coverage tool. It gives you a maintained estimate for every SKU instead of careful estimates for the top 40 and guesses for the rest. That is genuinely worth having.
Do not treat it as a solution to inventory risk. The risk lives in lead time variance, cash timing, and cost accuracy, and those are procurement and accounting problems. ConnectBooks has announced Crunch, an AI CFO built on reconciled seller data, with a waitlist open ahead of its release. Whatever analysis layer you eventually use, the numbers it reads have to be right first, and what a plan includes at each tier is on the pricing page.
Running an e-commerce business comes with plenty of challenges, but ConnectBooks is here to make your life easier. With real-time insights, seamless integrations, and detailed tracking of your profitability and inventory, you can stay ahead of the game. Whether you’re selling on Amazon, Shopify, Walmart, TikTok or eBay, ConnectBooks helps you manage your finances with 100% accuracy and confidence, so you can focus on growing your business.
Ready to level up? Start making smarter, data-driven decisions every step of the way. Try ConnectBooks Free Today or Schedule a Demo
