AI in Price Management: What Actually Works Today

Few topics in price management are discussed with as much exaggeration as artificial intelligence — and few are applied as pragmatically in actual projects. Between the expectation that a model will find the optimal price by itself and what genuinely works lies a considerable gap. This piece describes what holds up in practice, what it requires, and where classical statistics remains the more honest answer.

The expectation and the reality

The common notion runs like this: you give a model the sales data and it returns the profit-maximising price for every item. Technically that is not a wrong formulation — practically it fails on a property of the data that cannot be computed away.

To learn how demand reacts to price, a model needs price variation. In most companies there is none: prices went untouched for years, and when they did change, they changed for all items at once. The model then sees no reaction to price at all, but seasonality, economic cycles and assortment changes — and wrongly attributes those effects to price.

Four applications that genuinely work

1. Demand forecasting rather than price optimisation

The most stable application is not price but volume forecasting. Gradient boosting methods predict sales volumes from seasonality, day of week, weather, stock levels, promotions and competitor behaviour with considerably lower error than classical time series methods. That affects price indirectly: if you know an item will be scarce in six weeks, you do not cut its price today.

2. Elasticity estimation with pooling

For a single item, price elasticity can rarely be estimated cleanly — there are too few observations. The way out is hierarchical modelling: items are grouped by similar behaviour, elasticity is estimated at group level and then adjusted for the individual item. For the large mass of items that produces an estimate where none existed before — not perfect, but considerably better than a rule of thumb.

3. Attribute-based pricing for new items

A newly added item has no history. A model that has learned from the existing assortment which price contribution individual attributes carry — material, performance, brand, weight, availability — can propose a consistent entry price for the new item. In assortments with tens of thousands of positions that replaces the manual "take the nearest thing and add ten per cent".

4. Language models for data preparation

The most underrated application. A large share of the effort in pricing projects goes into consolidating messy data: standardising product texts, matching competitor offers to your own items, extracting attributes from free-text descriptions, assigning categories consistently. Large language models do this reliably today and shorten the preparation phase considerably. It is unspectacular and delivers the most immediate benefit.

Where classical methods are the better answer

Not every pricing problem is a learning problem. In three situations a simple, rule-based method is superior:

• When the decision has to be explainable. Justifying prices to customers, sales and management is not a side issue. A rule can be explained in one sentence; a gradient boosting model cannot.
• When the data is thin. With a few hundred transactions per item per year — the norm in B2B — structured value analysis beats any model.
• When the market changes structurally. Models learn from the past. After a competitor enters, an assortment is restructured or a supply chain breaks, the past is a poor guide.

Prerequisites you cannot skip

Before models are discussed, four things have to exist. If one is missing, any project is an expensive exercise:

1. Historical transaction data at line-item level, at least 24 months, including discounts granted — not just list prices.
2. Complete cost data per item, so that every recommendation can be checked against the contribution margin.
3. Price variation in the data, whether historically grown or deliberately created through tests.
4. A person who takes responsibility. Models produce proposals. Somebody has to decide whether they are adopted and carry the consequence.

If price variation is missing, the first sensible step is not a model but a structured testing programme: controlled price changes on defined item groups over eight to twelve weeks. That generates exactly the data you can later learn from — and produces reliable insight while it runs.

A realistic roadmap

What stands out is how late the models appear. That is not an oversight but the order that works: only once data, segmentation and limits are in place does a model have anything to build on. Start the other way round and you build a technically elegant system on a data foundation nobody trusts — and switch it off after the first implausible proposal.

What this means for your company

If you want to start today, the most productive question is not "which AI tool?" but: how much price variation is in our data, and how clean is our cost data per item? The answer determines whether you can work with models in three months, or whether the first step is a testing programme. Both are good starting points — you just need to know which one you are standing on.