Our commitment to compliance with the EU AI Act is reflected in how we integrate AI into our product offerings. This document outlines the usage of AI in our products, focusing on transparency, data handling, and responsible AI practices.

The EU AI Act categorizes AI systems into four risk levels: unacceptable risk, high risk, limited risk, and minimal risk.

At a high level, Constructor’s service helps online retail customers optimize how they present their products to their customers, for example, by helping retail customers search for the right product based on their preferences or previous purchases on that website. Constructor’s service uses AI to enhance the quality of these functions. With regards to the EU AI Act, the use of AI in this way presents minimal risk due to its sole focus of better presenting the customer’s existing products to its audience. As such, there should be no environmental, bias or adverse impacts caused to particular persons through its use.

Products

Additional detail on how we integrate AI into our product offerings can be found below:

Search

Product description: Search allows users to find specific items or content on a website by typing queries into a search bar. It quickly scans the website's catalog and returns the most attractive results, making it easier for users to find what they are looking for.

AI integration: Search utilizes AI for item retrieval and ranking, enhancing the user experience by presenting the most relevant items based on the user's prior actions (for example, clicks, add-to-carts or purchases).

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data.
• Outputs: The AI generates a ranked list of items that are most relevant to a search query.
• Training and fine-tuning: The AI model is trained on aggregated and anonymized behavioral and catalog data, and some fine-tuning is conducted using open source pre-trained models to improve relevance.

Risk level: Minimal

Browse

Product description: Browse lets users explore a website by navigating through categories. Instead of searching for a specific item, users can look through different options, much like walking through the aisles of a store.

AI Integration: Browse utilizes AI for item ranking and enhancing the user experience by presenting the most relevant items based on the user's prior actions (for example, clicks, add-to-carts or purchases).

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data.
• Outputs: The AI generates a ranked list of items that are most relevant to the user.
• Training and fine-tuning: The AI model is trained on aggregated and anonymized user behavioral and catalog data. No fine-tuning is conducted.

Risk Level: Minimal

Autosuggest

Product description: Autosuggest generates suggestions as users type in the search bar. It predicts what users might be looking for based on the initial letters or words entered, helping them find what they need more quickly and efficiently.

AI integration: Autosuggest utilizes AI to suggest search terms (based on queries entered by other users) and by presenting the most relevant items based on the user's prior actions (for example, clicks, add-to-carts or purchases).

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data.
• Outputs: The AI generates a ranked list of items that are most relevant to the user.
• Training and fine-tuning: The AI model is trained on aggregated and anonymized user behavioral and catalog data. No fine-tuning is conducted.

Risk level: Minimal

Recommendations

Product description: Recommendations are suggestions of items provided to users based on a user’s preferences, prior search history and interaction, or context on the page (for example, presenting recommendations of alternatives to a viewed item, bestsellers, etc.).

AI integration: Recommendations utilizes AI to predict the most relevant items for users based on prior actions and items they are interacting with.

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data.
• Outputs: A ranked list of recommended items tailored to user preferences and situations.
• Training and fine-tuning: The AI model is trained on aggregated and anonymized user behavioral and catalog data. No fine-tuning is conducted.

Risk level: Minimal

Quizzes

Product description: Quizzes are interactive tools that engage users by asking questions and providing tailored results based on their answers. They can help users find items that match their needs or preferences, making the shopping experience more personalized and fun. There is no AI in quizzes. All experience is manually created and curated by customers.

AI integration: None

Attribute Enrichment

Product description: Attribute Enrichment enhances product information by adding more detailed attributes, such as size, color, material, and other relevant features. This makes it easier for users to understand and compare items, leading to more informed purchasing decisions.

AI integration: Attribute Enrichment utilizes AI to predict and add attributes based on item image and description.

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data.
• Outputs: Enriched product attributes (for example, color) that improve search and filter functionalities.
• Training and fine-tuning: For specific domains and attribute types, the AI models are trained to enhance accuracy. No fine-tuning is conducted.

Third-party AI models: OpenAI.

Risk level: Minimal

AI Shopping Assistant

Product description: the AI Shopping Assistant is a virtual helper that uses AI to assist users in finding items, answering shopping-related questions, and providing personalized recommendations. It aims to make the shopping experience smoother and more efficient by offering intelligent support. Instead of just a query, users can write a phrase or intent related to their shopping experience, and AI will focus on returning everything which relates to a user's perceived intent, making the whole experience more intuitive.

AI Integration: AI is utilized to parse a free-form user-provided textual query and generate a dynamic response to fit the requested intent specifically by finding relevant items, answering shopping-related questions, and providing personalized recommendations. An example would be providing a list of ingredient, instructions, and related items for a query "I would like to prepare a carrot cake for my son's birthday."

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data. Retailer materials such as blogs, articles, FAQs, sales transcripts and other materials are optional inputs to train the LLMs to speak in the customer’s voice.
• Outputs: The assistant generates product recommendations, answers, and search results in free-form text. The assistant’s responses are passed to third party and homegrown guardrail and safeguarding models to flag misuse of the product, and to limit exploration and answer generation to the customer’s catalog context (for example, it will attempt not to answer any questions unrelated to the shopping experience).
• Training and fine-tuning: Responses are powered by customer-specific domains which provide instructions to prompt LLM responses. Constructor also provides some learning examples to train the model how to respond to typical customer-specific queries. Domains are fine-tuned to respond and textually speak according to the tone of hundreds of training samples in our fine tuning process.

Third-party AI models: OpenAI (primary), AWS Titan, Claude, Mistral, LLama 2 (AWS Bedrock).

Risk level: Limited. To the extent a customer’s use of the Constructor service incorporates AI-powered customer service bots (such as our AI Shopping Assistant), both parties should ensure that they comply with applicable transparency requirements so that users are aware they are interacting with AI.

Collections

Product description: Collections are curated groups of items organized around a theme, occasion, or category. They help shoppers easily find related items that go well together, making it simpler to shop for specific needs or inspirations.

AI integration: Collections utilizes AI to provide customers with an option to automatically group related items in collections, which customers have the opportunity to manually review and edit.

AI functionality:

• Inputs: The AI model processes behavioral data such as user interactions, viewed items, purchase history, and product catalog data.
• Outputs: Curated collections of items organized by theme or topic.
• Training and fine-tuning: No training or fine-tuning.

Third-party AI models: OpenAI (primary), AWS Titan, Claude, Mistral, LLama 2 (AWS Bedrock).

Risk level: Minimal

Merchant Controls & Intelligence

Product description: Merchant Intelligence is a group of product analytics reports and tools provided to businesses that allow them to gain insights into shoppers’ behavior using slotting rules performance dashboard and A/B testing of merchandising rules.

AI functionality: None

Conclusion

Our products incorporate AI in ways that are consistent with the guidelines set out by the EU AI Act. We are committed to transparency in our AI processes, ensuring that all AI usage is aligned with ethical and regulatory standards. While it is ultimately up to each customer to decide how to categorize our service based on how they choose to implement it, we can confidently say that Constructor's use of AI would not be categorized as a High Risk AI system.