How bonprix uses AI for assortment, product ranking and pricing
The German online fashion retailer has published how it uses machine learning to cull weak styles, rank products daily and advise on size. Here is what is documented, and what is not.
KEY TAKEAWAYS Summary by the editors
- bonprix's Learning Collection system, announced in September 2020, predicts demand for individual styles so that items with poor forecasts can be dropped or adjusted before the January 2021 collection.
- bonprix's product ranking model, described in December 2021, uses a convolutional neural network on seven days of clicks, wish-list additions, basket additions and purchases to forecast next-day sales.
- bonprix states that humans keep the final say on the range: its Head of Supply Chain Strategy said that AI can offer valuable input for creating styles but humans have the final say.
- bonprix has not published quantified results for its range or ranking models, so claims about uplift in sales or margin cannot be verified from its own statements.
- Blue Yonder lists bonprix as a customer for AI price optimisation software, but the public summary gives no figures, so pricing outcomes remain undisclosed.
bonprix, the Hamburg-based online fashion retailer, uses machine learning at three points in the commercial cycle: deciding which styles stay in the range, ranking products in its online shop each day, and, according to its software partner, optimising prices. The company has published how the first two work but no quantified results, which makes it a useful example of practical AI with unproven financial impact.
How does bonprix use AI to decide what to sell?
In September 2020 bonprix announced a self-developed system called Learning Collection, built with AI specialists from Otto Group data.works. Its article-specific success analysis tool predicts demand and sales success for items using all of bonprix's product data. It generates lists of predicted flops. Items with poor forecasts are dropped, and others are adjusted, for example in colour.
bonprix said the system uses a machine learning algorithm that keeps improving as it receives new information. It was tested in summer 2020 and used for the January 2021 collection. Jessica Külper, Head of Supply Chain Strategy, said that AI can offer valuable input for creating styles but that humans still have the final say. The announcement gave no accuracy figures or sales impact.
How does bonprix rank products in its online shop?
In December 2021 bonprix described a product potential forecasting model that drives daily-updated rankings in every category of bonprix.de. It uses a convolutional neural network, a technique from image recognition. The inputs are time series of each product's daily clicks, wish-list additions, basket additions and purchases over the previous seven days, arranged like an image so that the network can forecast the next day's sales.
Products whose stock falls below a set threshold are ranked lower, which links the ranking to availability. Registered users receive rankings refined by five age groups. The model replaced the previous application in Germany and was implemented in Austria, Switzerland and ten other countries after extensive A/B testing. bonprix says its predictions are far more reliable than those of its previous machine learning model, and expects a positive effect on conversion rates, but discloses no percentages.
| Application | Purpose | Inputs described | Published result |
|---|---|---|---|
| Learning Collection (ASA) | Predict style success and remove or adjust weak items | All product data | None quantified |
| Product ranking model | Daily ranking of products in each category | Seven days of clicks, wish-list adds, basket adds and purchases; stock level; age group | Described as more reliable than the previous model; no figures |
| Fit Finder | Size recommendations | Customer input plus purchase and product data from Fit Analytics | Stated aims: fewer returns and fewer multi-size orders; usage up from April 2020 |
| Price optimisation | Improve margin | Not published | Software partner states improved bottom line, no figures |
What does bonprix say about pricing and size advice?
Blue Yonder, the software provider, lists bonprix as a customer for AI price optimisation software and says it helped the retailer improve its margins. The public summary contains no dates, figures or method details, so the effect cannot be assessed. For size advice, bonprix uses Fit Finder from Fit Analytics, which combines information the customer enters with purchase and product data. bonprix's stated goals are higher loyalty, fewer returns and fewer multi-size orders, and its digital product manager reported significantly higher use of Fit Finder from April 2020 compared with before the pandemic.
What data and controls make this work?
Three features of bonprix's approach are transferable. First, the models rest on behavioural data that an online retailer collects daily, such as clicks and baskets. Second, availability is part of the logic: a ranking model that promotes products which cannot be delivered would harm customers. Third, the migration of all processes to Google Cloud, completed in 2021, shows that the data platform came before the more advanced model.
- Humans keep the final decision on the range.
- Models were tested through A/B testing before rollout.
- bonprix states that customer data analysis complies with the GDPR.
How does this compare with typical retail AI claims?
Much public discussion of AI in fashion retail rests on vendor claims of large uplift. bonprix's own statements are more measured, and in that respect they are more credible: the company describes what each model does, which data it uses, how it was tested and how it was rolled out, and it declines to quote figures it has not published. This is the right way to read a case, by separating the mechanism from the outcome.
The mechanism is plausible. Culling predicted flops before production reduces the risk of unsold stock, which is among the most expensive problems in fashion. Ranking by forecast demand puts available, popular products in front of shoppers. Size advice addresses the leading cause of returns. Each of these links to a cost that retailers recognise. What is missing is the size of the benefit.
- Ask what data feeds the model and how often it refreshes.
- Ask how the model treats stock availability, so it does not promote what cannot ship.
- Ask whether a human can override the output, as bonprix says its range planners can.
- Ask for measured effects against a control group, not only a before and after comparison.
The Learning Collection also illustrates a common limit of range prediction. A model trained on historical data predicts which new styles resemble past successes, and may be weaker for genuinely new ideas. That is one reason bonprix keeps human judgement in the loop.
The sources also show a pattern in how bonprix communicates. In both press releases it names its partners, states what is in-house and what is not, and notes the legal basis for data use. That openness helps readers judge scope. A reader can see, for example, that the ranking model is a self-developed system and that the fit tool is supplied by a specialist, which affects who controls the data and how easily the tool can be replaced.
It also shows the time dimension. The Learning Collection was announced in 2020 and the ranking model in 2021. Retailers evaluating AI should expect multi-year programmes with several model generations, not a single launch, and should plan for ongoing maintenance, monitoring for drift and periodic retraining.
What are the limits of what is known?
The published material dates from 2020 and 2021, so it describes the state of bonprix's systems at that time. The company said it planned further use in inventory management, personalised communication and behaviour-based ranking, but the sources reviewed do not confirm what was delivered. Readers should therefore treat bonprix as evidence that such models can be run in production at scale, not as evidence of a particular return on investment.
Frequently asked questions
Is bonprix using artificial intelligence?
Yes. bonprix has described a Learning Collection system for range planning, a convolutional neural network for daily product ranking and an AI-based Fit Finder for size advice, as well as fraud prevention. It says AI is used alongside human decisions.
How does bonprix rank products online?
bonprix uses an in-house model that analyses the previous seven days of clicks, wish-list additions, basket additions and purchases to forecast next-day sales. Products with low stock are ranked lower, and registered users get rankings refined by five age groups.
What results has bonprix reported from AI?
bonprix has not published quantified results in the sources reviewed. It says its ranking model is more reliable than the previous one and expects a positive effect on conversion, but gives no percentages.
Who supplies bonprix's AI pricing software?
Blue Yonder states that bonprix is a customer for its AI-based price optimisation software and that it improved margins. No figures or timelines are published in the summary reviewed.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
