Focus on Meaningful Problem-Solving

While navigating the complexities of generative AI, our guiding principle was to learn directly from our customers. We made sure not to prioritize technology over the objectives it serves by asking ourselves several key questions:

#### Determine Relevant Challenges

• What unique challenges do our customers and organization face?
• What current solutions exist for these challenges?
• What added value can generative AI technologies bring to solving these issues?
• How critical are these challenges, and what are the best and worst-case scenarios?

We realized that most customers shop with a specific event or occasion in mind, making it impractical to assign personal shoppers to 50 million customers.

An overload of choices can lead to customer frustration, as one remarked, “I love shopping online but it often takes too much time to really find what I need.” Conversely, some customers know precisely what they want, and we aim to facilitate that process.

#### Define Learning Objectives

Once we had a working prototype, we conducted research to identify potential challenges and customer reactions. Since this was new territory for our users, we encouraged them to interact with the Assistant. The research revealed their expectations, which influenced our roadmap and long-term vision.

Key questions for our learning objectives included:

• What is the significance of these learning goals?
• What opportunities will they unlock?

Transform the Interaction Paradigm

One crucial insight from our initial research was that customers struggled to visualize how to engage with the Assistant. As one user expressed, “I would have liked to better understand how to ask questions to the Assistant.”

For years, customers have relied on traditional e-commerce techniques like searching and filtering. Therefore, we needed to actively guide them in this new paradigm.

#### Show, Don’t Just Tell

To address various use cases (e.g., seeking fashion advice or finding a specific item), we prepared conversation starters to visually demonstrate how to engage with the Assistant.

As the project developed, this guidance helped shape the interaction in numerous ways, including the creation of a user-friendly search function showcasing how our generative AI algorithm works.

#### Introduce Meaningful Moments

Identifying key moments in the customer journey where the product adds value is essential. We aimed to avoid interrupting the customer experience. Our team brainstormed ways to establish meaningful interactions, such as offering advice on product detail pages or helping users combine recent purchases for desired looks.

Manage User Expectations

While generative AI holds promise across various scenarios, it is not a one-size-fits-all solution. Therefore, managing expectations and acknowledging both its strengths and limitations is vital.

#### Educate on the Model’s Boundaries

This process begins with the large language model itself. When developing the Zalando Assistant, we had to provide clear instructions to shape the experience from the Assistant's perspective, akin to training a personal shopping assistant.

We defined specific roles for the Assistant, including:

• Providing fashion advice
• Assisting with styling items
• Suggesting item recommendations
• Helping customers choose outfits for specific events

We also instructed it to avoid discussing customers' private lives, financial situations, or mental health.

#### Emphasize Transparency

To set proper expectations, it is crucial to communicate transparently about the product's current capabilities and limitations. This openness helps customers grasp the system's functioning, fostering trust and encouraging constructive feedback.

Embrace Failure Gracefully

Engaging with large language models necessitates anticipating and managing potential failures, such as misinterpretations or irrelevant suggestions. As generative AI is still evolving, learning from failures is essential.

#### Scale Learning

Each customer interaction with a generative AI product is unique, making it challenging to identify and learn from issues. To address this, we developed an internal AI tool to monitor performance at scale, quantify problems, and comprehend customer intentions. This tool is vital for maintaining quality and supports a human-in-the-loop approach to continuously enhance the Zalando Assistant and customer experience.

#### Ensure a Safe Experience

Proactively promoting safe and ethical usage of AI-powered products is essential. We must make intentional decisions, collaborate across functions, and take preventive measures.

Risk levels can vary based on the product, industry, and use cases. Instances like Netflix's generated images for a true-crime documentary or an airline facing backlash for poor chatbot advice highlight the importance of recognizing potential customer harm.

Our team assessed scenarios that could lead to various types of harm:

• Customer Harm: Physical, Psychological
• Organizational Harm: Financial, Social, Legal

We categorized these scenarios and prioritized them according to their likelihood and associated risks, helping us determine where to focus our mitigation efforts.

Key Takeaways

The Zalando Assistant project has expanded to four countries, reaching 500,000 users in just one year. The Assistant is primarily utilized for product discovery and receiving outfit recommendations or fashion advice for specific occasions, confirming our hypothesis that customers desired personalized guidance.

We aim to continue enhancing our Assistant in alignment with our vision of creating a fashion and lifestyle ecosystem. Significant potential exists to weave this solution into meaningful moments throughout the customer journey while bolstering security and reliability.

Our project has yielded valuable insights into leveraging generative AI technologies responsibly and ethically.

Are you embarking on a similar generative AI project? Here’s a summary of our key takeaways to help you navigate the complexities of developing generative AI-powered products.

Solving Meaningful Problems:

• Identify Relevant Challenges: Focus on real challenges and the unique needs of customers and the organization.
• Set Learning Objectives: Research and prototype testing are vital for understanding customer expectations and shaping product direction.

Transforming Interactions:

• Show, Don’t Just Tell: Customers require guidance to interact with generative AI; provide examples and proactive support.
• Introduce Meaningful Moments: Integrate AI into customer journeys where it adds value without disrupting the user experience.

Managing Expectations:

• Educate on Model Limits: Clearly define the roles and limitations of the large language model to ensure focused assistance.
• Practice Transparency: Communicate AI capabilities and limitations to both internal teams and customers.

Learning from Failures:

• Scale Learning: Utilize internal tools to monitor performance, quantify issues, and ensure a human-in-the-loop approach for continuous enhancement.
• Safeguard Experiences: Address potential risks proactively and ensure safe and ethical AI usage through intentional decisions and cross-functional collaboration.

If you have tips or principles for designing generative AI-powered products, we invite you to share them in the comments!

Next, ?eyda will share her insights on navigating ambiguity in product design.