AI inventory management: what the "AI" actually does, and the questions to ask

Short answer: in inventory tools today, "AI" usually means statistical forecasting — time-series math on your sales history, sometimes with a chat interface bolted on. That can genuinely work. But the label tells you nothing about whether the math is any good, whether it explains itself, or whether it respects the one fact that breaks naive models: your out-of-stock days aren't zero demand. Here's what the AI in these tools actually does, where it fails, and the seven questions that separate useful from branded.

Your plan, attacked: a pre-mortem of buying the AI tool

Before the feature tour, run the purchase as a pre-mortem. It's six months from now and the AI tool hasn't paid for itself. Here are the four most likely autopsies — each one is a real failure pattern visible in public app reviews as of October 2026, not a hypothetical:

Autopsy 1: "We couldn't tell what it was doing." The suggestions arrive, but not the reasoning. A verified 3-star Prediko review from September 2026 puts it precisely: the prediction method is "a black box… I actually built a more useful/customized tool just with spreadsheets and ChatGPT for effectively $0" (public App Store review, accessed Oct 2026). When you can't verify the math, you can't correct it — and an uncorrectable forecast you quietly stop trusting is worse than a spreadsheet you understand.

Autopsy 2: "It treated our bundles like single SKUs." A 1-star Prediko review from a 5-month user, edited May 2026: "this software is not capable of handling our bundle predictions" (public App Store review, accessed Oct 2026). Components and kits are where naive per-SKU forecasting visibly breaks — the model sees component sales, not kit demand, and neither gets ordered right.

Autopsy 3: "It applied the same logic to everything." Assisty's founder reply to a January 2025 review conceded: "The current reorder suggestions use the same min and max inventory levels for all products" (public App Store review reply, accessed Oct 2026). One-size-fits-all parameters on a catalog of 500 wildly different SKUs is statistical malpractice wearing an AI badge.

Autopsy 4: "It was right about everything except when we were out of stock." The subtlest one. Every model trained on your sales history sees zero sales during out-of-stock weeks — and if it reads those zeros as demand data, it under-forecasts exactly the SKUs that most need reordering. This is the failure the entire category struggles with, and no vendor's landing page will tell you whether they handle it. You have to ask.

None of these autopsies means AI forecasting is a scam. It means the label is not the product. The product is the math, the data hygiene, and the explanations — and those you evaluate, not the adjective.

What "AI" actually is in these tools, concretely

Strip the branding and the inventory category has roughly four things that ship under the AI name (all verifiable on vendor sites and listings as of October 2026):

1. Statistical demand forecasting. Time-series models — moving averages, exponential smoothing, and similar — applied to your sales history per SKU. This is decades-old math, done well or badly. The "AI" framing is mostly packaging; sometimes there's genuine model sophistication behind it, sometimes it's a 4-week average with a confidence interval.
2. Conversational interfaces over your data. Shopify's Sidekick is the clearest mainstream example — built into every admin, free with your plan, and able to answer questions and take actions in your store, with its own low-stock-alert prompt among its starter recipes (Shopify Sidekick page, verified October 2026). Prediko ships "Pia," an AI chat front door for reorder questions ("What's at risk of running out?"). These are genuinely useful as query layers — asking your data questions in plain English — but the chat is not the forecast. The forecast underneath is product 1.
3. Anomaly and risk detection. Flagging SKUs whose recent behavior deviates from pattern — the honest core of "AI watches your store." Valuable when the alerts explain themselves ("this SKU's velocity doubled; here's the math"), noise when they don't.
4. Generative text about your inventory. Summaries, digest emails written by a model. Harmless to useful; no forecasting claims should rest on it.

The evaluation trap is treating these as one thing. A tool can have an excellent chat layer over a naive forecast, or a strong forecast behind an unusable interface. Price them separately in your head.

Explainable vs. black box: the line that actually matters

Here's the stance worth holding: you should be able to reproduce any suggestion the tool makes, by hand, in your spreadsheet. Not because the spreadsheet is better — it isn't, at 500 SKUs — but because reproducibility is how you audit, correct, and ultimately trust a forecast.

The explainable pattern looks like: "Order 240 units of SKU-1234 by March 14. Why: 21 units/week in-stock velocity × 6-week supplier lead time × 1.8 safety factor, minus 85 on hand." Every number traceable to your own data. The black-box pattern looks like: "Suggested reorder: 240." If you push "why" and get a shrug, you've bought a horoscope.

The stakes are higher than they sound, because inventory AI fails asymmetrically. A wrong reorder suggestion that's 10% low costs you a small gap; a systematically naive model that reads out-of-stock weeks as low demand will quietly under-order your bestsellers forever — and you'll have no way to see it if the math is hidden. This is the honest case for wanting the math on the table, whether it's AI or not. (For the underlying method the math should approximate, see the reorder point formula — any AI suggestion should reduce to roughly that logic when you interrogate it.)

The seven questions to ask any "AI-powered" inventory tool

Ask before the demo ends. Vendors answer these well or they dodge — either answer is information:

1. How do you treat out-of-stock days in the forecast? The correct answer involves excluding or interpolating them, not averaging them in. If the answer is "our model handles that," ask for the mechanism.
2. Can I see why a suggestion was made — the actual inputs? Per-SKU, reproducible, with the formula visible. Not "confidence score: 87%."
3. How does it handle bundles and kits? Component demand vs. kit demand are different problems; "we don't support bundles yet" is an acceptable answer, silence is not.
4. Is the forecast per-SKU or per-category? Category-level parameters applied to every product is the Assisty failure pattern above.
5. What happens on day one, before I have history in your system? Every model is dumb at cold start; honest vendors say so and show their cold-start behavior.
6. Can I override a suggestion, and does my override persist? You know things the model doesn't — the trade show, the TikTok spike, the supplier's planned shutdown.
7. Does anything execute without me? Auto-reordering is a policy decision, not a feature decision. A tool that drafts suggestions for your approval is auditable; one that fires POs while you sleep is a different risk class entirely. Read-only is a legitimate design choice — some tools are deliberately built to suggest and never execute, which keeps every decision yours.

Where AI honestly fits a small store today

The realistic picture for a 100–5,000-SKU Shopify store, as of October 2026:

• Sidekick is already in your admin, free — it's a legitimate way to query your store's data in plain English, and its inventory-adjacent value is answering questions, not making buying decisions (Shopify, verified Oct 2026).
• The spreadsheet remains a serious baseline. A disciplined spreadsheet computing velocity and cover per SKU is explainable by construction and free — the honest competitor every AI tool must beat, and the reason the DIY path in the Prediko review above is credible. If you haven't built that baseline, it's in the Shopify inventory spreadsheet guide.
• Paid AI forecasting earns its price when your catalog is too big to maintain by hand, your demand has real complexity (promos, seasonality), and the vendor passes the seven questions — with question 1 non-negotiable. The landscape of who charges what is covered in the Prediko alternatives comparison; the method ladder underneath any forecast — whatever powers it — is in the forecasting guide.

One disclosure in the interest of the same standard: we are building InventoryX, a read-only buying-list tool that is in development and not available today, and it does not use AI and doesn't market itself as AI — its velocity and cover math is deliberately simple enough to check by hand, for exactly the reasons in this article. We'd rather you hold every tool, including ours, to the explainability test than buy an adjective.

Common questions

What does AI actually do in inventory management software? Mostly statistical demand forecasting (time-series math on your sales history per SKU), plus conversational query layers over your data, anomaly detection, and generated summaries. The "AI" label spans tools from naive averages with good packaging to genuinely sophisticated models — so evaluate the math and its explanations, not the adjective.

Is AI inventory forecasting accurate? It's as accurate as its treatment of your data's failure modes. The decisive factor is how the model handles out-of-stock days: models that read those zero-sales weeks as low demand systematically under-forecast your bestsellers. Ask any vendor exactly how they handle it — and whether you can reproduce a suggestion by hand before trusting the forecast at scale.

What is the difference between AI forecasting and a reorder point formula? The reorder point formula is transparent math you compute and control: velocity × lead time + safety stock. AI forecasting automates and complicates the estimation — sometimes capturing seasonality and promos the formula can't — but it must still reduce to roughly that logic when you interrogate it. If a tool's suggestion can't be explained in those terms, treat that as a red flag.

Does Shopify have built-in AI for inventory? Yes — Sidekick, Shopify's AI assistant, is included with Shopify plans and can answer questions and take actions in your admin, including inventory-adjacent prompts like low-stock alerts (Shopify Sidekick page, as of October 2026). It's a query and assistance layer; deep per-SKU forecasting for reordering is where third-party tools compete.

Should I let an AI tool automatically reorder inventory? Treat auto-execution as a policy decision, not a feature upgrade. Tools that draft suggestions for your approval are auditable and reversible; tools that fire purchase orders autonomously put your cash flow behind a model you may not be able to inspect. Many operators deliberately choose read-only tools for exactly this reason.

Where to go next

Before evaluating any AI tool, build or refresh the baseline it has to beat: a spreadsheet that computes in-stock-only velocity and days of cover for your top 50 SKUs — the method is in the inventory spreadsheet guide. Then take the seven questions above into the demo. If a vendor can't answer question 1 cleanly, you've saved yourself the money regardless of how good the dashboard looks.