Demand Forecasting With AI for Small Businesses: Where to Start

Demand Forecasting With AI for Small Businesses: Where to Start

Demand forecasting with AI for small businesses works best when it starts simple: a spreadsheet with sales history by product and by week, a basic statistical model or an AI assistant to organize and explain the numbers, and an honest comparison against the method you already use. If your history is short, demand is very irregular or the product is new, no tool makes up for missing data. The goal is not to predict the future. It is to be wrong less often than you are today, in a way you can measure.

Key takeaways

• A spreadsheet with clean sales history is usually enough for most recurring items in a small business.
• The minimum data is sales by period, product identification, a flag for stockouts, and a record of promotions and special dates.
• Compare against your current method (gut feel, last month's average, the supplier's suggested order) using past data the model has not seen.
• AI helps clean data, explain deviations and draft forecasts, but the purchasing decision stays with a person.
• Never present a forecast internally as precise. It is an estimate with a known margin of error.

What demand forecasting is and where AI fits

Forecasting demand means estimating how much customers will buy of each item in a future period, so you can decide how much to buy, produce or staff. If you run a shop, restaurant, distribution business or small factory, you already do this every day, usually from memory, looking at last month and the calendar.

AI enters in three different ways, and it helps to separate them because the risk and cost of each are different.

• Classic statistical models, such as moving averages, exponential smoothing and simple regressions. They run in a spreadsheet or a standard library and are easy to explain.
• Machine learning models trained on your data, useful when you have many items, many variables and a long history.
• Language assistants, meaning general-purpose generative models, which help organize the dataset, build formulas, write the comparison plan and explain in plain language why a number came out different from what you expected.

For small businesses, the third category is the most realistic way in. The assistant does not guess the future. It reduces the work of preparing and interpreting a forecast that, at its core, is basic statistics. This matches the general advice to start with the simplest solution that solves the problem and add complexity only when it shows a measurable gain.

When is a simple spreadsheet forecast enough?

In most cases it is enough when demand has some regularity and the decision that depends on it is reversible or low impact. Use this decision guide.

• Items sell often, nearly every week: a spreadsheet is enough.
• There is a visible pattern by weekday, month or season: a spreadsheet with a seasonal adjustment is enough.
• A forecasting error costs little, for example an order can be topped up within a few days: a spreadsheet is enough.
• Many items sell sporadically, one or two units at a time: group them by product family before forecasting, or use a minimum stock rule instead of a forecast.
• The product is new, or its price and target customer changed: there is no reliable history. Use a manual estimate and review it quickly.
• The decision is expensive and hard to undo, such as a large purchase with a long import lead time: use the forecast as one reference among several, never the only one.
• The business depends on a few large customers: the sales conversation predicts better than any model.

The central point is that the model only needs to beat your current method, not be perfect. Often a weighted average of recent weeks, adjusted for holidays and promotions, already outperforms a hunch. Move to something more sophisticated only after you have measured that the simple version leaves money or stock on the table.

What is the minimum data you need?

The most common mistake is not in the model, it is in the dataset. If recorded sales are only the sales you fulfilled, not real demand, the model learns scarcity as if it were low demand. At a minimum, gather:

• Sales by product (or product family) and by period, preferably weekly. Daily only if your operation decides daily, as a restaurant does.
• A history covering at least one full year, so the model sees your strong and weak seasons. Without it, seasonality becomes guesswork.
• A stockout flag: periods when the item was out of stock or unavailable. Without it, a drop in sales misleads you.
• An event calendar: promotions, holidays, campaigns, price changes, price increases and unusual closures.
• Returns and cancellations kept separate from sales, so you do not double count or mix signals.
• A clear definition of unit: boxes, units, kilos or orders. Mixing units is a frequent source of silent error.

Data that helps but is not required at the start: weather, competitor prices, website traffic, the volume of WhatsApp conversations about a given product. Add one variable at a time and keep only those that improve the comparison.

About where the data comes from: if your sales live in a system such as Bling, Omie or Tiny, the simplest route is usually a periodic spreadsheet export. If you later want to automate the reading, first understand the paths and risks described in integrating AI with Bling, Omie or Tiny.

How to compare against your current method without promising accuracy

The comparison is what turns an idea into a decision. The technique is called a backtest: you pretend you are at a past date, make a forecast using only what was known then, and compare it with what actually happened. Follow this plan.

• Write down your current method. It might be last month's average, the standard supplier order or the manager's estimate. If it is not written down, you cannot compare against it.
• Set aside a recent test period, for example the last few weeks of the dataset, and hide that data from the model.
• Generate the forecast from the current method and from the new one for the same period and the same items.
• Calculate the error of each with a simple measure, such as mean absolute error (the average gap between forecast and actual, ignoring direction) and bias (whether the method tends to over-forecast or under-forecast).
• Also look at the practical impact: how much stock would have been left over, how many stockouts would have happened, how many emergency orders.
• Repeat across more than one test period, including a peak season and a quiet one, so a lucky window does not fool you.
• Record where the new method did worse. Those cases tell you more than the average.

What do you tell the team and leadership? That the forecast is an estimate with a margin of error measured on the past, and that there is no guarantee the future will behave the same way. Avoid goals like getting everything right. Prefer criteria such as: the new method is wrong less often than the current one on most important items and does not raise the risk of stockouts on critical items. To structure that acceptance criterion, the article on how to measure an AI automation pilot offers metrics and decision gates that apply here.

Illustrative scenario: a cleaning products shop

Illustrative scenario, no real data. A small cleaning products shop buys from two suppliers and decides its weekly order based on what the owner remembers from the previous month. Sometimes it runs out of detergent at the end of the month, other times it is left with too much disinfectant.

The owner exports weekly sales for each item from the past year and a half, flags the weeks when an item ran out, and lists the promotions and holidays. They ask an AI assistant to help build a spreadsheet with two forecast columns: one from the current rule (the average of the last four weeks) and one from an average adjusted for time of year and promotions.

Then they hide the most recent weeks, generate both forecasts and compare them with what was actually sold. They find that for steady sellers the adjusted rule errs less. For sporadic items neither works well, so they decide to keep a fixed minimum stock for those. They also notice that the history of a reformulated product is not usable, and handle it by hand.

The practical result is not a magic forecast. It is a clear split: where to use the spreadsheet, where to keep the minimum stock rule, and where to decide from knowledge of the business.

Where AI helps and where it gets in the way

AI helps with supporting tasks: cleaning duplicate product names, standardizing units, suggesting formulas, summarizing the week's biggest deviations and drafting a short explanation for the purchasing team. It is also useful for proposing hypotheses, such as a drop coinciding with a stockout or the end of a promotion, which you then check in the data.

It gets in the way when it is treated as an oracle. Language models can produce plausible numbers with no real calculation behind them, and can get arithmetic wrong without warning. So keep the forecast calculation in formulas or code you can inspect, and use the assistant to prepare, explain and review. In practice the rule is: numbers calculated by a reproducible method, text and interpretation with AI support, the decision with a person. To define what stays out of AI's reach in your daily operations, read what not to delegate to AI in business.

There is also the question of data. Sales spreadsheets by product usually contain no personal data, which simplifies compliance with Brazil's LGPD. But if your dataset includes names, phone numbers or purchase history by customer, remove those columns before sending anything to an external tool. Decide which tools are allowed and what kind of data may go into them; a company policy for employee ChatGPT use handles this simply. For official guidance on data protection in Brazil, consult the ANPD, the national data protection authority.

To treat the forecast as a system that can be wrong and needs monitoring, the NIST AI Risk Management Framework is a useful reference on mapping, measuring and managing risk. And if you later move to an AI automation, OpenAI's guidance on evals shows how to build repeatable tests to know whether a change improved or worsened the result.

How to apply map, limit, pilot and transfer

The method we use in AI projects for business processes fits this topic well, without requiring anything heavy.

• Map: describe how the forecast is made today, who decides, with what data, and what it costs to be wrong in each direction (too much or too little). Keep this snapshot as your baseline.
• Limit: choose a few relevant items, define which data goes in, which tools may be used, and that the AI only suggests. No purchase order is issued automatically.
• Pilot: run the new method in parallel with the current one for an agreed period, record forecasts and outcomes, and compare them with the success metric defined before you began.
• Transfer: document the step-by-step so the team can repeat it without depending on anyone outside, and decide with evidence whether to expand, adjust or stop.

Stopping is a legitimate outcome. If the simple method does not beat the current one, you saved yourself a larger investment and learned where your data problems are.

Frequently asked questions

Does demand forecasting with AI work for a company with little history? In a limited way. With less than one full sales cycle, the model cannot see seasonality and tends to miss in peak periods. In that case, combine a simple average with the knowledge of the people running the business, and keep enriching the dataset.

Do I need an expensive system or a data scientist to start? Not for the first step. A well organized spreadsheet, a simple statistical method and an AI assistant for support are enough to run the backtest and compare against your current method. More robust tools only make sense after the simple version has proven its value and shown its limits.

Can AI place the purchase order automatically? Technically yes, but it is not advisable at the start. Keep AI as a generator of suggestions and have a person approve each order, with a record of what was suggested and what was decided. Only discuss more autonomy once you have a measured track record and clear limits on value and risk.

How do I know if the forecast is good enough? Compare the error of the new method with that of the current one using past data, and look at the practical effect on stockouts and overstock. It is good enough when it is wrong less often than the current method on the items that matter and the worst cases are known and handled. There is no universal accuracy number, because it depends on the type of product and the cost of being wrong.

Sources

• NIST: AI Risk Management Framework
• Anthropic: Building effective agents
• OpenAI: Evals guide
• ANPD: Autoridade Nacional de Proteção de Dados