Demand forecasting is the thing every inventory tool leads with and the thing your business is least ready for. The useful work is quieter, and you can start it this week.
In the M5 forecasting competition, 5,507 teams tried to beat a set of benchmarks on 42,840 series of real Walmart sales data. Only 7.5% beat a simple exponential smoothing method, and fewer than half beat a plain naive forecast. That is the honest ceiling on AI demand forecasting, and it assumes clean history you probably do not have: a landmark retail study found 65% of nearly 370,000 store inventory records did not match the physical count. So the biggest wins for a small ops team are the unglamorous ones. Reconcile your SKU list, find the stock that has not moved in a year, get your reorder logic written down where you can read it, and make the model argue with your assumptions before you spend money on a platform.
Inventory management is two problems wearing one coat. Either you have cash sitting on a shelf that nobody wants, or you have a customer standing in front of an empty one. Every tool in this category, AI or otherwise, is selling you a better trade-off between those two.
The money involved is large enough to feel abstract. IHL Group, which has tracked this measure for 18 years, puts the global cost of inventory distortion (out-of-stocks plus overstocks) at $1.73 trillion a year, equal to 6.5% of global retail sales, even after retailers spent $172 billion on improvements in the previous twelve months.[1] North America accounts for $415 billion of it, and supply chain disruption alone for $301 billion.[1]
None of that tells you anything about your business. This might: across US retail as a whole, the Census Bureau’s inventories-to-sales ratio sat at 1.25 in May 2026, meaning retailers were holding roughly one and a quarter months of merchandise against a month of sales.[2] If your number is meaningfully above that, you have a cash problem before you have a forecasting problem.
The closest thing to a picture of businesses your size comes from Netstock, which combined anonymised platform data from more than 2,400 customers with a survey of over 130 users at companies under $250M in revenue. It found 55% holding at least 20% excess stock, up from 48% the year before, 46% reporting that 5% or more of inventory counts as dead stock, and 17% carrying more than 10% dead stock, up from 12%.[3]
Worth sitting with for a second: every business in that sample had already bought inventory planning software.
I used to open workshops with the demand forecasting demo, because it lands. You paste in two years of sales, the model draws a confident line into next quarter, and the room goes quiet in a good way. I stopped doing it, because I could not answer the one question a sharp ops manager always asked: better than what?
There is a clean public answer to that question and almost nobody in this market quotes it. The M5 competition, run on Kaggle and written up in the International Journal of Forecasting, gave 5,507 teams (7,092 people across 101 countries) five and a half years of real Walmart unit sales, 42,840 series covering 3,049 products across 10 stores, and asked them to forecast 28 days ahead.[4] These were motivated people competing for a $50,000 prize pool with modern machine learning.
The scoreboard against the organisers’ benchmarks:
Of 5,507 teams forecasting 42,840 series of Walmart sales data, the share whose final submission outperformed each benchmark method (M5 Accuracy competition results).[4]
More than half the field could not beat a forecast that says “tomorrow looks like today.” Only 415 teams, 7.5%, beat exponential smoothing applied at the product-store level and summed upward, a method you could implement in a spreadsheet.[4]
The honest other half of that story: the teams at the top did win, and they won properly. Every one of the top 50 improved on that exponential smoothing benchmark by more than 14%, the top five by more than 20%, and the winning team by 22.4%.[4] Machine learning is genuinely better here. The catch is that “genuinely better” means roughly a fifth better than a method with no AI in it, achieved by specialists on immaculate data, and 92.5% of the people who tried could not get there at all.
So when a vendor tells you their AI is “30% more accurate,” the only useful reply is: more accurate than which baseline, at what level of aggregation, over what horizon? Stephan Kolassa, who spent years producing automated forecasts for large European retail chains, wrote a whole piece in Foresight pulling apart the published accuracy surveys people quote at each other, and concluded that “the quest for external forecasting benchmarks is futile.”[5] His own working number is worth knowing: in grocery retail, he saw normal one-week-ahead errors ranging from 20% to 60% MAPE across different companies, depending on how many fast sellers they had, how much fresh produce, and how good their data was.[5]
That range is the real answer to “what accuracy should I expect.” It’s a three-fold spread, and the thing that moves you within it is mostly not the algorithm. If you want the longer version of this argument applied to revenue rather than units, we covered it in our guide to using AI for sales forecasting.
A few months ago I sat with an operations manager at a homeware wholesaler who wanted help choosing between two forecasting tools. Before we looked at either, I asked her to export her product list. It had 1,900 rows. One planter appeared three times: once as “Terracotta Pot LG”, once as “TERRACOTTA-POT-LARGE”, and once as “TP-LG-01”, each with its own stock count and its own sales history.
No model fixes that. Every forecast it produces will be a forecast of one third of the demand for a planter, three times over, and it will look completely plausible.
This is not a small-business embarrassment, it’s the normal condition of retail data. The reference study here is DeHoratius and Raman in Management Science, who examined nearly 370,000 inventory records across 37 stores of a single retailer and found 65% of them inaccurate, with an average absolute deviation of almost five units per SKU, about 35% of the average actual stock on the shelf.[6]
What makes that worth acting on rather than just despairing about is the sales evidence. An ECR Retail Loss study ran matched pairs of stores across several retailers, tracked sales and inventory records for 12 weeks, then performed a stocktake and corrected the records in the test stores only, leaving the control stores alone, and tracked another 12 weeks. Across roughly 233,000 SKUs, correcting inventory records grew sales by 4% to 8%.[7]
Here’s the reframe I now use in every session. Data cleaning isn’t the boring prerequisite to the AI work, it’s the AI work with the highest confirmed payoff attached to it, and it’s the part a language model handles well. Pattern-matching three spellings of “terracotta pot” is exactly what these tools do without complaint at two in the afternoon on a Thursday.
These are the five uses I’d defend to a sceptical finance director. None of them involve the model predicting the future.
Paste your product export into a general-purpose assistant and ask it to group probable duplicates, showing its reasoning for each group. The prompt phrasing that works: “Group rows that are likely the same physical product under different names or SKU formats. For each group, list every row you included and say what evidence you used. Flag anything you are unsure about separately.” That last sentence matters more than it looks, because without it you get a confident merge list with no way to audit it. Our walkthrough on using AI to analyse a spreadsheet covers the mechanics, and the same technique transfers straight across to product records.
Dead stock hides because nobody has a report for it. Give the model your SKU list with last sale date, units on hand and unit cost, and ask for every item with no sale in 180 days, sorted by capital tied up, with a running total. You’ll get a number for how much cash is sitting still. In the Netstock survey, 11% of businesses admitted they had no strategy at all for reducing excess, and 69% relied on promotions while only 29% redistributed stock between locations.[3] Knowing which items to promote is the whole game, and it’s a sorting problem.
Ask for a reorder point calculation per SKU, written as a spreadsheet formula, using your average daily sales, your supplier lead time and a safety stock buffer you choose. The output should be a formula in a cell, visible and editable. The moment your reorder logic lives inside a model rather than inside your spreadsheet, you’ve lost the ability to explain a purchase order to your accountant.
Forward it your last twenty supplier confirmation emails and ask for a table of supplier, product, quantity, promised date, actual date, and days late. Netstock’s respondents named lead time variability their top supplier problem at 68%, ahead of long lead times at 58% and cost at 48%.[3] You cannot buffer for variability you have never measured, and most small teams have the data for it sitting unread in an inbox.
This is the one people skip and the one I’d keep if I could only keep one. Paste your draft purchase order alongside last year’s sales for those items and ask: what is this order assuming about demand that isn’t written down? Which of these lines would I regret if sales came in 20% under plan? Which am I buying because of a minimum order quantity rather than because I need it? On that last point, 45% of surveyed businesses said minimum order quantities force them to buy more than they need.[3]
Something I notice in almost every session: people expect inventory to be harder than it is. Then we write the reorder point on a whiteboard and there’s a small silence.
Average daily demand, multiplied by supplier lead time in days, plus a safety buffer. That’s the reorder point. When stock falls to that number, you order. The safety buffer is where judgement lives, and judgement is the part that should stay yours.
Lead time is the input that quietly destroys people, because most businesses use the number the supplier quoted rather than the number the supplier delivers. The spread is enormous. In Netstock’s benchmark data, the top quartile of businesses held average lead times of 20 days or less, while the bottom quartile faced delays stretching beyond 80 days.[3] Same industries, same regions, four times the wait. If you’re buffering against a quoted 20 days and actually receiving 60, no forecasting model on earth will save your service level.
So the sequence that works is: measure your real lead times from your own emails and receipts, feed those into the reorder point, and only then worry about whether your demand estimate is clever. The supplier conversation matters more than the algorithm, and the receipts that prove your real lead times are usually sitting in accounts payable, which is where our guide to using AI for invoice processing picks the thread up.
My honest take after years of teaching this to non-technical teams: use AI to write the formula, explain the formula, and pressure-test the inputs. Let the spreadsheet do the arithmetic. Language models are excellent at reasoning about structure and unreliable at long chains of calculation, and there is no reason to make them do the one thing a spreadsheet has done perfectly since 1985.
Most operations managers who ask me which inventory platform to buy need a clean spreadsheet and two good prompts. A minority genuinely need software, and the tell is that they’ve already run the manual process twice and keep hitting the same wall. Here’s what the market looks like if you’re in the second group, with every price taken from the vendor’s own pricing page in August 2026.
| Tool | Published price | Worth knowing |
|---|---|---|
| Zoho Inventory | Free plan (50 orders, 1 user, 1 location); Standard $29/organisation/month billed annually, up to Enterprise at $249 | Priced per organisation rather than per user, which suits small teams |
| Katana | Free plan; Core from $299/month | No per-user fees; built around manufacturing and assembly |
| Inventory Planner by Sage | Free to install, then quote only | 4.4 out of 5 across 130 reviews on Shopify’s app store |
| Stocky by Shopify | Included with a Shopify POS Pro subscription | Rated 3.0 out of 5 across 188 reviews on Shopify’s own app store |
| Netstock | From $900/month, annual subscription | Implementation typically takes 6 to 10 weeks and requires an ERP to plug into |
| Microsoft 365 Copilot | $30/user/month, paid yearly, on top of a qualifying Microsoft 365 licence | Puts Copilot inside Excel, where your inventory data probably already lives |
Prices as published on each vendor’s own pricing or listing page, checked August 2026. Ratings are the vendors’ own app store listings.[8,9,10,11,12,13]
Two things stand out to me in that table. The first is the gap between $29 and $900, which is not a gap in quality so much as a gap in what the product assumes about you: Netstock is built to sit on top of an ERP and needs six to ten weeks of implementation before it does anything.[12] If you don’t have an ERP, that price isn’t for you yet.
The second is the Stocky rating. Shopify’s own inventory app, free with POS Pro, sits at 3.0 out of 5 across 188 reviews on Shopify’s own store.[11] Free bundled tooling is worth trying precisely because it costs nothing, and it’s also worth checking the reviews before you build a process around it.
No new software, no integration, no project plan. Four hours and an export.
Two rules while you run it. Don’t paste supplier contract terms, pricing agreements or anything commercially sensitive into a consumer chatbot; use product codes and aggregate figures, or a tool where your organisation has a contract and training on your data is switched off. And write the date next to every assumption, because in six months you’ll want to know whether your plan was wrong or whether the world moved, and those need different fixes.
Run this twice, a quarter apart. If the second run surfaces the same problems as the first and you’re still fixing them by hand, that’s the moment a platform starts making sense.
Four failure patterns, in the order I see them.
Forecasting on top of unreconciled data. Covered above, and still the most common. The output looks fine, which is the entire problem. A forecast built on three spellings of one product is wrong in a way that no accuracy metric will reveal to you.
Buying the platform to avoid making the decision. Software will tell you what your data says. It will not tell you whether to carry a slow-moving line because a good customer expects it. In Netstock’s survey, 62% of businesses fell into what it labelled insufficient forward planning, sitting on slow stock or replenishing excess without cutting purchase orders, and that share went up, not down, year over year.[3] Those are businesses that already own the software.
Letting the model do arithmetic out of sight. If you cannot point at the cell where a reorder quantity came from, you cannot defend the purchase order, and eventually somebody will ask you to.
Chasing accuracy when the constraint is cash. Reliance on credit for inventory fell from 53% to 47% between 2024 and 2025 in that same survey, cash usage from 54% to 44%, and 27% now report no defined financing strategy at all.[3] If capital is your binding constraint, a 5% better forecast is worth far less than clearing the dead stock you already know about.
The single thing I’d have you take from all of this: AI is genuinely useful here, but the usefulness sits in getting your own numbers into a state where a decision is possible rather than in the demand curve. That work is unglamorous and checkable, and a small team can do it in an afternoon. The teams getting value from AI for inventory management are rarely the ones with the best model, they’re the ones who finally looked at what they own.
Less well than the marketing suggests. In the M5 competition, 5,507 teams forecast 42,840 series of real Walmart sales and only 7.5% beat a simple exponential smoothing benchmark, while fewer than half beat a plain naive forecast. The best teams did beat it, by around 22%, but they were specialists working on clean data. For a small business with patchy history, expect a modest improvement over a sensible spreadsheet method, not a transformation.
Two full years is the practical minimum, because you need at least two cycles of your seasonality to distinguish a pattern from a one-off. More important than length is consistency: one product per row, consistent units of measure, discontinued lines marked as such, and physical counts done within the last quarter. A short clean history beats a long messy one every time.
Forecasting on top of data you have not reconciled. A study of nearly 370,000 store inventory records at one retailer found 65% did not match the physical count, with an average error of about 35% of the stock actually on the shelf. Duplicate SKUs, inconsistent units and stale counts all produce forecasts that look completely plausible and are quietly wrong.
Most small teams get further with a clean spreadsheet and a general-purpose assistant than with a platform they have not defined a process for. Published entry prices range widely: Zoho Inventory has a free plan and a $29 per organisation per month tier, Katana starts at $299 a month, and Netstock starts at $900 a month with a six to ten week implementation. Run your manual process twice, then buy if you keep hitting the same wall.
Product codes, quantities and dates are usually low risk. Supplier contract terms, negotiated pricing and customer-identifiable order data are not, and should not go into a consumer chatbot. Use aggregate figures and internal codes where you can, and reserve anything commercially sensitive for a tool where your organisation holds a contract and model training on your data is disabled.
Every number here was checked against its original source in August 2026, and prices come only from each vendor's own pricing or listing page. Several widely circulated figures were left out because they could not be traced to a primary source, including the “carrying cost is 20 to 30% of inventory value” rule of thumb, which appears everywhere and originates nowhere I could verify. I also excluded the widely repeated claim that Shopify is retiring Stocky on a specific date, because the only source for it is a Shopify help page that would not load for me. The Netstock figures come from a vendor's own customer base and are labelled as such in the text. This is general operational guidance, not financial advice.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.