ASOS
catalog-scale deep learning product recommendation, plus a conversational style assistant
ASOS serves roughly 5 billion recommendation requests per day across nearly 50,000 products; its move to a transformer model in 2024 improved its evaluation metrics by more than 20%, complemented by a conversational AI Stylist on Azure OpenAI.
Key points
- Personalized deep learning product recommendation across the whole catalog.
- In-house transformer model plus a conversational AI Stylist on Azure OpenAI.
- Evaluation metrics improved by more than 20%, about 5 billion requests per day.
- Evidence level B, confirmed active status.
Objective
Make a catalog of tens of thousands of items, refreshed every week, browsable and relevant for a largely mobile audience.
The deployment
ASOS sells fashion online to a young audience, with nearly 50,000 items available at any time and hundreds of new products each week. To keep this catalog navigable, the company ranks products by personalized recommendation across its various touchpoints, based on customer interactions (purchases, saves, cart additions) processed by deep learning. In 2024, the team replaced its longstanding matrix factorization model with a transformer model, which improved its evaluation metrics by more than 20% over that baseline. The recommendation system serves about 5 billion requests per day, in nine languages and more than 200 markets. ASOS has also launched an AI Stylist, a conversational assistant built with Microsoft on Azure OpenAI, that helps the customer discover products through dialogue based on their style and the occasion.
Results Proof B
Figures published on the official ASOS technical blog (recommendation team) and corroborated by a Microsoft customer story for the conversational assistant. The 20%+ improvement is an offline evaluation metric, not an audited business result; the scale, on the other hand, is well documented.
How it works
Documented architectureThe stack in detail
- outil Modele transformer de recommandation (in-house ASOS) In-house model that replaced matrix factorization in 2024 (+20% on evaluation metrics) and serves about 5 billion requests per day.
- plateforme Microsoft Azure AI Studio Azure environment used with Microsoft to build the conversational AI Stylist assistant.
- llm Azure OpenAI OpenAI models served via Azure, the LLM component of the AI Stylist; the exact model version is not named in the sources.
- outil AI Stylist Product discovery assistant through dialogue (style, occasion), connected to the ASOS catalog.
How it runs, concretely
For ops teams-
1Aggregate interactions AI / data team
Purchases, saved products, and cart additions continuously feed the customer's taste profile.
-
2Rank the catalog per customer AI / data team
The transformer model orders products by relevance for each person, on each surface.
-
3Serve at scale AI / platform
About 5 billion recommendation requests per day are served in nine languages and more than 200 markets.
-
4Open discovery to dialogue AI / conversational assistant
The AI Stylist lets the customer describe what they are looking for in natural language and suggests products in response.
Customer interactions (purchases, saves, cart additions). On a catalog refreshed every week, a new product starts with no history and must be attached to signals quickly, or it stays invisible in the feeds.
How your customers perceive this type of use
Sourced studiesLe paradoxe est documente des deux cotes : 71% des consommateurs attendent des interactions personnalisees et 76% sont frustres quand elles manquent (McKinsey, 2021), mais 75% declarent ne pas acheter aupres d'organisations auxquelles ils ne confient pas leurs donnees (Cisco, 2024). La « creepy line » est localisee : messages recus quelques secondes apres une recherche et suivi de localisation sont les pratiques qui mettent le plus mal a l'aise (Periscope by McKinsey, 2019).
Acceptance conditions
- La confiance dans le traitement des donnees precede l'achat : 75% ne achetent pas sans elle (Cisco 2024)
- Un cadre legal protecteur rassure : 59% des consommateurs disent que des lois fortes sur la vie privee les rendent plus a l'aise pour partager des informations dans des applications IA (Cisco 2024)
- La personnalisation elle-meme est attendue quand elle est consentie : environ la moitie des consommateurs (US 55%, UK 52%) disent s'inscrire souvent ou parfois a des services personnalises (Periscope by McKinsey 2019)
Red lines
- Le message declenche quelques secondes apres une recherche ou un achat : deuxieme ou troisieme cause de malaise selon les pays (Periscope by McKinsey 2019)
- Le suivi de localisation percu comme de la surveillance : 40% de malaise en Allemagne et au Royaume-Uni (Periscope by McKinsey 2019)
- Le mesusage des donnees personnelles par l'IA, devenu la premiere inquietude des consommateurs, a 53% et en hausse (Qualtrics 2025)
Sources: McKinsey & Company 2021 · Periscope by McKinsey 2019 · Cisco 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- History of customer interactions (purchases, saves, cart)
- Catalog with product attributes and a stream of new arrivals
- Real-time serving infrastructure
Org prerequisites
- ML team to maintain and retrain the ranking models
- For the conversational assistant, a governed LLM component (safeguards, transparency)
Possible stack
- Managed recommendation engine (AWS Personalize, Google Recommendations AI) for a first version
- Azure OpenAI or equivalent for a style assistant, connected to the catalog
The plan, step by step
- Step 1Audit the collection of customer interactions (purchases, saves, cart additions) and attach it cleanly to a per-person identifier.Deliverable: Usable interaction dataset, per customer and per product.
- Step 2Put in place a first personalized ranking on a high-traffic surface (app home), via a managed engine or a simple model, in A/B test against the current sort.Deliverable: Personalized feed in test with a locked measurement plan.
- Step 3Read the A/B test (engagement, conversion) and handle the cold start of new arrivals via product attributes.Deliverable: A/B report and new-product attachment rules.
- Step 4Extend the ranking to the other surfaces and industrialize retraining as the catalog refreshes.Deliverable: Retraining pipeline and production monitoring.
- Step 5Prototype a conversational style assistant connected to the catalog, in closed beta, with safeguards and transparency.Deliverable: Assistant tested on a limited scope before opening.
First step: Replace the default sort on a high-traffic surface (app home) with a personalized ranking based on interactions, and measure engagement and conversion against the current sort.
The same pattern, proven in other industries
Function of Beauty
personalized product formulation by algorithm from a quiz
Henkel
genAI content supply chain for one-to-one personalization at scale
Hims & Hers
AI-driven personalized treatment routing on intake data (MedMatch)
Sources
- S1 Transforming Recommendations at ASOS - ASOS Tech Blog (Ed Harris) Interested party archive pending
- S2 ASOS uses Azure AI Studio to surprise and delight young fashion lovers - Microsoft Customer Stories Interested party archive pending
- S3 ASOS Plc Annual Report and Accounts 2025 Primary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.