I’m a big fan of the work that Shyvee Shi, now Product Lead at Microsoft, has been doing through the LinkedIn PM series where she speaks with leading product and tech people. Together with Caitlin Cai and Dr. Yiwen Rong, Shi has now written “Reimagined: Building Products with Generative AI” , providing readers with guidance on how to integrate generative AI into product strategy and careers.
“Reimagined” is pretty comprehensive in its scope, covering the generative AI landscape, the process of building generative AI products and navigating product careers in an AI era. Reading the book, it became apparent to me again that foundational product management skills aren’t going to disappear overnight in the era of (generative) AI; in the book there are many useful sections about what I’d deem to be critical skills applicable to both AI and non-AI products. The book contains many real-life use cases of generative AI products such as copy.ai, Galileo AI and Redfin.
The book’s section about things to watch out for when working with generative AI was one that grabbed my attention:
- Creating a coherent sentence — Generative AI can struggle in mastering language intricacies. Despite advancements, AI models still struggle with aspects like lexicon, grammar, context and tonality.
- Infusing a desired degree of control and determinism — Unlike humans, these AI models lack an innate grasp of human values, making it difficult to tailor their outputs to our subjective preferences, such as humour or ethics.
- Maintaining consistent output quality — For instance, GANs might produce visually striking images marred by imperfections. Language models can generate text that lacks clarity or consistency upon detailed examination, revealing a gap in quality assurance.
- Requiring vast amounts of high-quality data for training — Despite the advent of few-shot and transfer learning, the procurement and curation of expansive datasets are still resource-draining. Furthermore, there is the added challenge of ensuring that this data is representative and diverse, to prevent unintentional biases from infiltrating AI models given the nuances of human experiences.
- Lack of memory and context limit — The lack of memory in generative AI technology, or the context limit, restricts the AI’s grasp of preceding interactions, often leading to a disjointed and incoherent dialogue flow. This is a stark contrast to human communicators who effortlessly reference past exchanges to enrich conversations.
“Reimagined” offers a number of useful product principles for generative AI products:
- Innovate to serve the needs of people — Generative AI products should be built to solve real user needs, bring true value, and enhance social good.
- Design for transparency and explainability — Users need to trust and accept generative AI products, and transparency is key in helping users comprehend, and make sense of these AI-generated experiences and outcomes.
- Implement continuous feedback loop — Generative AI models are heavily dependent on the quality and quantity of input data.
- Balance automation and control — Striking the right balance between automation and user input is essential for fostering creativity and productivity in generative AI products. By giving users the ability to set parameters and constraints, or to adjust the AI’s operation according to their specific needs, product teams can create a sense of empowerment and avoid a situation where the AI system dictates the outcome.
- Prioritise safety and ethics — Ensuring thorough safety measures, security protocols, obtaining user consent, and adhering to privacy regulations and ethical guidelines are non-negotiable aspects of AI product development.
- Design with accessibility and inclusivity in mind — Consider accessibility and inclusivity in the design process is vital for creating user-centric generative AI products that cater to diverse user groups. By providing options for different language settings, supporting screen readers, and offering alternative input
- Aim at augmenting human capabilities — This principle emphasises that generative AI should aid users in achieving their goals more efficiently and effortlessly, not necessarily as a replacement for human skills, but rather, as a supportive tool.
Building on these general product principles, “Reimagined” then offers practical tips on how to design and build generative AI products. There are several design and build considerations specific to generative AI products that we should take into account when creating generative AI products:
Conversational AI
Conversational AI aims to simulate human-like dialogue, making interactions more natural and intuitive. Some of the related design challenges include crafting a user experiences that accounts for various conversational context and maintaining a coherent and engaging dialogue flow.
- Dynamic responses
- Context preservation
- Sentiment analysis
- Real-time adaptation
- Variable dialogue flow
Pi, a companion chatbot is mentioned as a good example of compelling conversational AI.
Content Generation
In the realm of content generation, generative AI produces text, images, or even music autonomously. Key UX patterns include:
- Prompt guidance
- User-guided constraints
- On-the-fly content creation
- Preview and edit
- Auto-complete and suggestions
- Ethical and responsible generation
The book mentions Canva Magic Studio as a good example of how to apply these UX patterns.
Search AI
LLM-powered search engines are able to provide a crisp, articulate answer in response to a search query, accompanied by options for further exploration:
- Concise answer curation
- Source attribution
- Interactive, chat-like search interfaces
- Generative content suggestions
- Visual & multimodal search
The book mentions Consensus AI as a good example of applying search AI , providing concise answers from a vast database of scientific papers.
Personalised AI
Personalisation takes user engagement to the next level by tailoring experiences to individual preferences. Designing for personalisation involves creating interfaces that can intelligently adapt without becoming intrusive.
- Adaptive UI
- User behaviour tracking
- Dynamic content loading
- Context-aware notifications
- Ethical personalisation
Not surprisingly perhaps, the book mentions Netflix as a good example of personalised AI, providing streaming journeys tailored to the individual user (e.g. adapting images of recommended shows based on user taste).
Predictive AI
Predictive AI anticipates user needs and actions, often before the user explicitly states them. This requires a design that seamlessly integrates these predictions into the natural flow of interaction, avoiding any abrupt or jarring user experiences.
- Next best actions
- Predictive search
- Pre-loaded information
- Contextual predictions
- Data generation for simulation and prediction
The book mentions Amazon’s predictive shopping experience as as a good case in point of predictive AI. Amazon is well known for its product recommendation and order prediction capabilities, and earlier this year it released “Rufus” — a virtual shopping assistant.
Assistive AI & Generative Productivity
Assistive AI aims to make tasks easier and more efficient for the user. In contrast to traditional software that might offer static help menus or FAQs, Assistive AI dynamically guides the user based on their actions.
- Contextual tooltips
- Task automation
- Real-time error detection
- Adaptive workflow assistance
- Data-driven decision making
- Interoperability
The book mentions the Generative Productivity Workflow by copy.ai as a good example of assistive AI and generative productivity.
Main learning point: As a product person, I really like how “Reimagined” has been written through a customer and product centric lens. The book provides a useful perspective on the opportunities and limitations of generative AI products, illustrated through many real-life use cases and product examples.
