AI forecasting for business inventory — what works and what is demo-ware

Demand forecasting is the most oversold AI feature in business software, and the reason is simple: it demos beautifully and deploys badly.

A forecasting demo runs on years of clean, high-volume, well-behaved data. Your business has three years of messy data, a product range that changed twice, two outlets that opened mid-period, and a best-seller whose sales spike whenever a particular influencer mentions it.

This is not an argument against forecasting. It is an argument for knowing which techniques survive contact with business reality — because some genuinely do.

The honest constraint: you do not have enough data

Sophisticated models need volume, and volume means both history and frequency.

A model learning weekly seasonality needs many weeks. Learning annual seasonality needs several years — and by "several years" we mean several years of the same business, which most growing businesses have not had. You changed your range. You opened an outlet. You started selling on TikTok. Every one of those is a structural break, and a model trained across a break learns a business that no longer exists.

Then there is the per-product problem. Aggregate sales might be substantial, but forecasting is per SKU, and most SKUs are slow movers. A product selling three units a month has essentially no signal — the variance swamps the trend. No model fixes that, because the information is not there to extract.

This is the thing demos hide. They forecast the top 20 SKUs, where there is real signal, and quietly do not mention the other 400.

What actually works at business scale

The unglamorous answer: simple methods, applied consistently, beat sophisticated methods applied to insufficient data. Not always — but often enough that this is the right default.

Reorder points and safety stock. Not AI at all. Arithmetic: how fast does this sell, how long does the supplier take, how much buffer covers the variability. This solves a large share of stockouts and it has solved them since the 1950s. Most businesses we meet do not do this properly, and are shopping for AI forecasting instead. Fix this first. It is boring and it works.

Seasonality on aggregate, not per SKU. Your category has a Raya spike, a year-end spike, a school-holiday dip. There is enough data at category level to see that even when there is none per product. Forecast the shape at the level where signal exists, then distribute.

Slow-mover classification. Rather than forecasting a three-unit-a-month product — which is noise — classify it. Is it dead stock? A long-tail keeper? Something to make-to-order? That is a decision, not a prediction, and it is more useful than a forecast with error bars wider than the forecast.

Anomaly detection. Genuinely well-suited to AI and undersold. "This SKU is selling four times its usual rate" or "this outlet's shrinkage has doubled" needs no long history and is immediately actionable. Detecting what already changed is much easier than predicting what will.

Human-in-the-loop adjustment. Your buyer knows about the promotion next month. The model does not. A forecast the buyer can override — and where overrides are recorded and reviewed — beats both a pure model and a pure gut call. The recording matters: it is how you find out whether the model or the human is actually better at your business.

Where AI genuinely earns its place

Three areas where it is not hype:

Messy multi-channel consolidation. You sell on Shopee, TikTok Shop and in an outlet. The same product has three names, three SKU conventions and three settlement formats. Reconciling that into one demand picture is a language and matching problem — exactly what modern models are good at. This is more valuable than the forecast itself, because most businesses cannot even see their real demand across channels. We cover the operational side on the Retail & Ecommerce page.

Explaining the forecast. A number is not decision-support. "Up 30% because the last three Julys ran 25–35% above June, and you are running a campaign" is. Language models are good at this and it changes whether anyone trusts the output.

Handling the long tail cheaply. Not by forecasting each slow mover, but by classifying hundreds of them consistently — a job nobody has time to do by hand.

Questions that expose demo-ware

Ask these. The answers are diagnostic.

• "Run it on my data, including my slow movers." Not the demo set. If the answer involves a lengthy data-science engagement, you are buying a project, not a product.
• "What does it do with a product that has three months of history?" Correct answer: something modest and honest. Wrong answer: a confident forecast.
• "What happened to the forecast during the last big promotion?" Promotions break models. A vendor who says otherwise has not run one.
• "Does it beat a naive baseline?" The bar is "same as last month" or "same month last year". Serious vendors know their numbers against a baseline. Many models do not beat it, and nobody volunteers that.
• "Can my buyer override it, and do you track who was right?" If overrides are not first-class, the tool will be ignored within a month.
• "What is the error rate — and against what?" Accuracy quoted without a baseline is marketing.

That fourth question is the one that ends most sales conversations, which tells you something.

Common questions

Do I have enough sales history for AI demand forecasting?

Probably less than you think. Learning weekly seasonality takes many weeks, and annual seasonality takes several years of the same business — but most growing businesses changed their product range, opened an outlet or added a channel, and every one of those is a structural break. A model trained across a break learns a business that no longer exists. Aggregate volume can look healthy while the individual products you want forecast have almost no signal in them.

Can AI forecast a product that sells a few units a month?

Not usefully. The variance swamps the trend, so there is no signal to extract and no model fixes that, because the information is not there. The better move is to classify the slow mover rather than forecast it: dead stock, a long-tail keeper, or something to make to order. That is a decision instead of a prediction, and it is more useful than a forecast with error bars wider than the forecast.

What should I fix before buying a forecasting tool?

Start with accurate stock data, because no forecast survives inventory records that disagree with the shelf. Then do reorder points and safety stock properly — how fast it sells, how long the supplier takes, how much buffer covers the variability. That is arithmetic, it has worked since the 1950s, and most businesses shopping for AI forecasting have not done it. Anomaly detection and consolidating demand across channels come next; forecasting comes last.

How can I tell whether a forecasting demo is real?

Ask for it to be run on your data including your slow movers, not the demo set. Ask what it does with a product that has three months of history — the honest answer is something modest. Ask what happened during the last big promotion, whether it beats a naive baseline such as "same month last year", and whether your buyer can override it and the tool records who turned out to be right.

The uncomfortable summary

For most businesses, in priority order:

1. Get accurate stock data. No forecast survives inventory records that are wrong. This is the actual bottleneck for most businesses and it is not an AI problem.
2. Do reorder points properly. Arithmetic. Solves much of the pain.
3. Add anomaly detection. Cheap, needs little history, immediately useful.
4. Consolidate demand across channels. Where AI genuinely helps.
5. Then, maybe, forecast. Once you have clean data, a stable range and enough history for the model to learn a business that still exists.

Most vendors sell you step 5 while you are stuck on step 1. If your stock records disagree with your shelf, a better forecast changes nothing — it just predicts the wrong number more confidently.

The general principle holds here as everywhere: AI is very good at the parts that are reading, matching and explaining, and much weaker at the parts that require information you do not have. Forecasting is mostly the second kind.

The next step is usually smaller than people expect. Talk to us about one process worth starting with.

Related: what is an AI-native ERP on where AI belongs architecturally, and AI in accounts payable for a function where it works better.