Responsible AI FAQs for Personalized Shopping Agent (Preview)

Personalized Shopping Agent is a Microsoft-managed solution that provides a personalized consumer experience with contextual, informative interactions and recommendations. Customers can chat with this agent to get product suggestions and guidance in natural language, much like talking to a store associate, helping them make informed purchase decisions.

In this article, you find answers to some frequently asked questions (FAQs) about Personalized Shopping Agent, how it uses AI, the safeguards in place, its limitations, and how it aligns with key Responsible AI principles (fairness, transparency, privacy, safety, and accountability) for external customers and partners.

What is Personalized Shopping Agent?

Personalized Shopping Agent is a chat-based solution that you can embed into your own shopping websites, customer-facing apps, or your in-store apps like POS or custom apps. It helps you enrich customer interaction and increase sales. Your customers can chat with the AI-enabled agent, ask for suggestions, and get guidance to make informed shopping decisions. The agent converts these insights into sales in a seamless way.

What are the system’s capabilities?

Personalized Shopping Agent uses a machine learning model called GPT-4 Turbo. GPT-4 Turbo is trained on a large amount of text from the internet and can generate new text that looks and sounds like human-written text. The agent takes a natural language statement from a customer through a chat interface and provides a response that is conversational and relevant to your data. By combining a powerful generative AI model (GPT-4) with search and your custom data, the Personalized Shopping Agent can both "think" (understand and generate text) and "know" (retrieve factual information) in order to answer customer queries. The agent’s responses are thus grounded in a mix of the customer’s query, the retailer’s proprietary data, and broader knowledge when appropriate, providing reliable and contextually relevant assistance. Microsoft discloses these technical components so that you know what technologies are at work behind the scenes. This transparency helps you trust the system’s operations and ensures you can make informed decisions about its use.

What can Personalized Shopping Agent do?

Personalized Shopping Agent provides a chat-based interface to the end customers. Customers can have a conversational shopping experience on this interface in a natural language format.

Customers can have natural language-based queries related to any event. For example, "I'm going backpacking in Yosemite National Park in March. This adventure is my first such experience. Can you help in buying the essentials." Or they can have specific queries related to a product or product line. For example, "Show me some beginner-friendly ski boots with good insulation."

The inherent idea about this conversational experience is to give an experience to customers where they can get informed, aware and can then make the right product decision, as they would do while talking to a shopping personnel in a regular retail store.

What languages does Personalized Shopping Agent support?

Personalized Shopping Agent is currently supported in English language only. The AI model and the solution’s prompts are tuned for English language interactions. Microsoft plans to localize Personalized Shopping Agent to other languages over time, prioritizing languages in markets based on customer demand.

If multi-language support is critical for your scenario, you can discuss this requirement with the Microsoft for Retail product account team or email the Microsoft for Retail community (email: msfrcommunity@microsoft.com). Feedback from preview users helps guide which languages are added next.

Where does Personalized Shopping Agent process data?

Personalized Shopping Agent processes data in your Dataverse tenant by using respective Power Platform and Dataverse services. You define the Azure resources and choose the country/region where the environment needs to be set up as per the availability of the Copilot Studio resources.

What data does Microsoft store for Personalized Shopping Agent?

Microsoft doesn’t collect or store personal data or user data. Microsoft gathers aggregated telemetry (stored for 30 days) about usage that helps us improve the quality of the product.

All shopper data and usage information are stored in Dataverse storage, which is present on your tenants. Microsoft doesn’t have access to this information. For more information on data privacy and security for Azure OpenAI service, refer Data privacy.

Microsoft’s approach is to give you full control over customer data. The agent operates on your data, in your cloud, and Microsoft’s role is to provide the AI technology without taking possession of your users’ personal information. Always inform your customers how their data is used when interacting with the agent (transparency builds trust), and ensure you configure the system in line with your organization’s privacy and security standards. With Personalized Shopping Agent architecture, you have the tools to do so while benefiting from advanced AI capabilities.

How did Microsoft evaluate Personalized Shopping Agent?

Microsoft implemented multiple safety mechanisms to minimize the risk of harmful or inappropriate outputs from the Personalized Shopping Agent. These safeguards span the AI model configuration, content filtering, and testing processes.

Microsoft uses a set of evaluation methods that include a collection of sample questions, answers, and relevant documents to test the quality of the AI answers vs the human answers baseline.

• Focused domain and grounding. Personalized Shopping Agent is designed to stick to the domain of shopping assistance. By grounding answers in the retailer’s product data and related info, the system naturally remains focused on relevant content (for example, it talks about products, recommendations, etc., rather than veering off into unrelated or problematic topics). This grounding reduces the chances of the AI producing random inappropriate text because it constantly refers back to factual data.
• Microsoft conducted internal evaluations to catch potentially harmful behaviors. They specifically assessed categories of harm, including harassing or harmful content and biases related to protected characteristics.

We evaluate the solution by following Microsoft quality guidelines. It’s advisable for you to conduct an evaluation check for each new implementation, as the product is currently in preview, and we warmly welcome feedback. It’s important to note that the evaluation is conducted thus far for English.

As part of the mandatory prerelease check, the product goes through a Responsible AI assessment following Microsoft RAI guidelines. The features in the agent also go through multiple quality reviews such as privacy review, security review, and threat reviews before launching to customers.

Microsoft’s Responsible AI assessment and testing process for Personalized Shopping Agent aligns with the principle of Reliability and Safety, aiming to prevent harm while enabling helpful interactions. Users should still use their judgment and report any questionable output, but the system significantly reduces risks out-of-the-box.

What are the current limitations of Personalized Shopping Agent? How can customers minimize the impact of Personalized Shopping Agent?

Transparency means not only touting what the AI can do, but also being clear about what it can’t. Microsoft openly notes that answers from Personalized Shopping Agent might not be perfect and could occasionally be incorrect or incomplete due to the nature of AI models.

• Personalized Shopping Agent is built on the GPT-4o Mini model in Copilot Studio. The underlying technologies are trained on a wide range of sources and some answers might not be perfect. It’s important for you to do a thorough test with internal users and provide feedback. This testing can help in setting up the right prompt based on retailer's needs and brand requirements.

• Personalized Shopping Agent undergoes a three-step process of launch- private preview, public preview, and General Availability. Since Personalized Shopping Agent is in preview, we recommend you to use the solution in test environments only. You can minimize the impact in two ways:

  • Define specific prompts based on your domain and nature of business.
  • Optimize appropriate prompts to align with your brand guidelines.