AI-Assisted Product Attribute Extraction for E-commerce Content Creation

zalando · E-commerce · 2024 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: E-commerce product content enrichment

Zalando needed faster, more accurate product attribute enrichment across markets while keeping copywriters in control.

Constraints:

25 markets
attribute code translation
human final decisions
cost/accuracy tradeoff

Architecture

Zalando’s Content Creation Copilot integrates OpenAI GPT into the product-onboarding QA flow rather than replacing human content experts. The Content Creation Tool sends image URLs and article metadata to a Prompt Generator, which uses Article Masterdata and an attribute-translation layer to turn Zalando-specific codes into human-readable prompt context. GPT returns suggestions, the translation layer maps them back into API-compatible codes, and the UI preselects/marks AI suggestions while leaving final decisions to human QA.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  content-creation-tool["Content Creation Tool"]
  prompt-generator["Prompt Generator"]
  article-masterdata["Article Masterdata"]
  openai-gpt["OpenAI GPT"]
  attribute-translation-layer["attribute translation layer"]
  suggestion-source-indicators["suggestion source indicators"]
  human-qa["human QA"]
  content-creation-tool --> prompt-generator
  prompt-generator --> article-masterdata
  article-masterdata --> attribute-translation-layer
  prompt-generator --> openai-gpt
  openai-gpt --> attribute-translation-layer
  attribute-translation-layer --> content-creation-tool
  content-creation-tool --> suggestion-source-indicators
  suggestion-source-indicators --> human-qa

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Content Creation Tool`	Internal UI that receives generated attribute suggestions, preselects them, and leaves final QA to content experts.	`Prompt Generator`, `human QA`, `suggestion source indicators`
`Prompt Generator`	Builds GPT prompts from product images, article metadata, and translated attribute values.	`Content Creation Tool`, `Article Masterdata`, `OpenAI GPT`, `attribute translation layer`
`Article Masterdata`	Metadata store for article attributes, attribute sets, optional/mandatory fields, and market-specific codes.	`Prompt Generator`, `attribute translation layer`
`OpenAI GPT`	Model endpoint used to generate product-attribute suggestions from prompt context and images.	`Prompt Generator`, `attribute translation layer`
`attribute translation layer`	Converts Zalando-specific attribute codes into readable values for GPT and maps GPT output back to API-compatible codes.	`Article Masterdata`, `Prompt Generator`, `OpenAI GPT`, `Content Creation Tool`
`suggestion source indicators`	UI markings that show which suggestions came from the prompt generator or other sources.	`Content Creation Tool`, `human QA`
`human QA`	Content experts make final decisions and contribute prompt improvements.	`Content Creation Tool`, `suggestion source indicators`

Data flow

Photographers/content users work in the Content Creation Tool. The tool sends product-image URLs to the Prompt Generator; the generator combines those images with Article Masterdata, translates attribute codes into readable labels for GPT, calls OpenAI GPT, converts suggestions back into Zalando-specific codes, and returns preselected suggestions with UI source indicators for human review.

4 supporting quotes

“This diagram illustrates the current workflow involving the interaction between four components: Content Creation Tool, Prompt Generator, Article Masterdata and OpenAI - GPT.”article, implied

“Content Creation Tool: Internal content creation tool used by photographers to upload images, which URLs are sent to the Prompt Generator.”article, stated

“One of the biggest challenges was translating the Zalando-specific attribute codes provided by Master Data”article, stated

“The solution was to get the English translation of the possible attribute values (in this case it’s Petite and Tall), wait for the GPT response, and then translate it back into the attribute_code.”article, stated

Why this architecture

The design addresses a high-cost manual QA step while protecting customer experience. Zalando needed model suggestions compatible with its internal APIs, visible source attribution in the UI, controlled rollout switches for products/attributes, and human final approval because generated attributes directly affect search, discovery, and customer-facing product information.

4 supporting quotes

“As the manual process contributed to approximately 25% of the overall content production timeline, we've prioritized the development of assistive functions to support the QA process.”article, stated

“Our system is designed to combine AI input with other sources, while the user interface clearly shows what suggestions come from which source, leaving the final decision in the hands of the human.”article, stated

“As the suggestions directly impact customer experience, it was imperative for us to ensure the output of OpenAI was compatible with our APIs.”article, stated

“The way we built our architecture helped us do a controlled rollout where we could easily include/exclude products or attributes with minimal effort.”article, stated

Models used: OpenAI GPT-4 Turbo · OpenAI GPT-4o

Scale

The manual content process was about 25% of the overall content production timeline.
“the manual process contributed to approximately 25% of the overall content production timeline”article, stated
Zalando operates in 25 markets, which is why article attributes are stored as codes.
“As Zalando operates in 25 markets with different languages, we are storing the attributes of the article as attribute codes.”article, stated
The copilot reached approximately 75% accuracy and enriches about 50,000 attributes weekly.
“We've achieved an accuracy rate of approximately 75%, and we're enriching around 50,000 attributes on average per week.”article, stated
GPT-4o reduced response time and cost compared with the earlier GPT-4 Turbo setup.
“The new model not only provided better results but also delivered faster response times and proved to be more cost-effective.”article, stated

Alternatives considered (roads not taken)

Manual-only copywriter QA

Rejected because: Manual enrichment consumed a large share of production time and left opportunities to reduce errors earlier.

“Copywriters enrich attributes using a Content Creation Tool and perform Quality Assurance (QA) themselves to guarantee the four-eyes principle.”article, stated

Use raw master-data codes directly with the model

Rejected because: The model needed human-readable values, then API-compatible codes on the way back.

“One of the biggest challenges was translating the Zalando-specific attribute codes provided by Master Data”article, stated

Stay on GPT-4 Turbo for production generation

Rejected because: GPT-4o produced better results, faster response times, and lower costs.

“As GPT-4o was announced relatively early in the copilot's development, we initially performed a human inspection, comparing the accuracy of different sources for sample articles.”article, implied

Patterns

`human-in-the-loop-gate` (content-operations)

AI suggestions are shown with a visual cue while copywriters retain final selection authority.

`structured-output-contract` (content-operations)

The system converts visual product analysis into controlled product-attribute codes.

“assortment_type_7841) into human-readable language understandable by the GPT model and then translating the suggestions back into the Master Data-specific code.”article, stated

Tradeoffs

Model choice for production attribute extraction.

Chose GPT-4o after comparison over staying on GPT-4 Turbo.

Rationale: The newer model improved speed and results, but selection was based on domain expert inspection and accuracy tradeoffs.

“The new model not only provided better results but also delivered faster response times and proved to be more cost-effective.”article, stated

Gotchas (what bit them)

Zalando-specific attribute codes were not understandable to the model.

Root cause: Master data stored codes such as assortment_type values rather than human-readable labels.

Resolution: Translate possible attribute values to English for the model, then map suggestions back to codes.

“One of the biggest challenges was translating the Zalando-specific attribute codes provided by Master Data (e.g.”article, stated

Evaluation

Zalando evaluated model predictions on catalog samples reviewed by domain experts.

Metrics: 75% accuracy · 50,000 attributes weekly · 25% timeline contribution

“We evaluated the results of various models on a large sample of articles from our catalog assortment measuring accuracy by having the predictions reviewed by domain experts.”article, stated

Outcomes

The copilot enriches about 50,000 attributes weekly with 75% accuracy.

AI-assisted copilot, not a fully autonomous agent; marked non-agentic.

“translation layer for Zalando-specific attribute codes, the system now enriches approximately 50,000 attributes weekly with 75% accuracy.”summary, stated