Forecasting with AI in Business Central: what really works today?

Business Central includes forecasting functions in the standard: sales and inventory forecasts per item, payment forecasts for cash flow and continuously expanded Copilot analyses on your own figures. They work well where history and master data hold: several years of clean transaction data, unambiguous items, maintained payment terms. They do not replace planning for one-off effects, new products or thin history. The honest sequence is therefore: first check the data history, then activate the forecast, then measure against real months.

The three fields of use in the standard

1. Sales and inventory: forecasts per item based on your own sales history, as a basis for ordering and stock decisions. Strength: a uniform fast-moving range. Weakness: seasonal items with short history and anything that lives on individual orders.
2. Payments and cash flow: forecasts of when customers are likely to pay, based on past payment behaviour. Useful for the liquidity view, but honestly only as good as the maintained payment terms and settlement data.
3. Copilot analyses: questions to your own figures in everyday language, with evaluations as the answer. Not a forecasting model in the narrow sense, but the fastest way to see patterns that turn into decisions.

What a usable forecast requires

Forecasts learn from the past, and they learn everything: your duplicates, your renumbered items and your never-maintained payment terms as well. An item created three times under a new number has three short histories for the model instead of one long one. Forecast quality is therefore not a model topic but a master data topic.

As a rule of thumb: at least two full years of transaction data per forecast field, cleaned master data and a person who holds the first three forecast months against reality. After that you know where to trust the model and where not.

Where the limits are

No standard model sees what is not in the data: the new competitor, the trade fair, the broken machine at the customer's site. Forecasts are a baseline, not a truth. They only become dangerous when nobody contradicts them anymore: the workable order is model suggestion plus human correction with reasoning, neither blind acceptance nor blind ignoring.

For project manufacturers and single-order business, additionally: item forecasts carry little there, because hardly any order resembles another. The value then sits one level deeper, with purchased parts, standard parts and consumables, and in the payment forecast, which works in project business too.

The pragmatic entry

One field, three months, two figures: choose the forecast field with the best data history, let it run alongside for three months without hanging decisions on it, and compare forecast against actuals monthly. If the quality holds, the field goes productive and the next one is checked. This builds forecast usage with evidence instead of gut feeling, and the question of when AI in your ERP pays offanswers itself for this field along the way.