Recommendations for designing conversational user experiences
Applies to this Power Platform Well-Architected Experience Optimization checklist recommendation:
| XO:10 | Design conversations that align with users' needs and help the AI accomplish its objectives. Make it clear what the AI can do. Create natural interactions that feel intuitive and human-like. Provide fallback mechanisms to ensure the AI gracefully recovers from problems. |
|---|
This guide describes the recommendations for designing the user experience for conversational AI in a workload. Conversation design is the process of creating structured, intuitive, and natural dialogues between users and AI. Conversation design significantly enhances the user experience, by making interactions with conversational AI more intuitive and efficient, reducing user frustration, and by helping users complete their tasks more quickly.
Key design strategies
Conversation design involves crafting the flow of interactions to ensure that the AI understands user intent, responds effectively, and guides users toward their goals. Good conversation design enhances the user experience by making interactions feel more human-like and seamless. Therefore, user engagement and satisfaction increase. When you create conversational AI for your workload, aim to develop human-like, user-friendly interactions that facilitate natural communication and help users achieve their goals efficiently.
Explain the capabilities of the AI
Be upfront about the AI's capabilities and help ensure that users understand what the AI can do. For example, if the AI can perform only specific tasks, let users know about this limitation right from the start. Make sure that users understand why the AI suggests specific actions or produces specific results. Provide explanations for the AI's recommendations to help users be successful and increase their trust in its capabilities and accuracy.
Clear communication about the AI's abilities helps manage user expectations by preventing unrealistic assumptions and minimizing disappointment and frustration. The result is a more positive user experience. An AI system that is transparent about its strengths and limitations boosts user confidence in the technology and builds trust. This trust is essential for user engagement and satisfaction.
Awareness of the AI's specific capabilities improves usability because users can navigate interactions more effectively and use the AI more efficiently. Clarifying the AI's capabilities reduces miscommunication because users are less likely to make requests that are beyond the AI's scope. This approach reduces potential frustration and dead ends.
Understand the user's intent
Understanding user intent is a cornerstone of conversation design because it directly influences how well the AI can fulfill user requests and navigate complex dialogues. Accurately capturing intent ensures that the AI not only understands what the user is asking for but also responds in ways that feel intuitive and relevant to the user's needs.
When the AI recognizes and responds to intent appropriately, it creates a more natural and seamless conversation, reduces frustration, and helps users accomplish their goals more efficiently. Proper intent recognition helps the AI avoid dead ends by ensuring that it understands the purpose behind a user's input, even when the phrasing or terminology varies. The interaction remains fluid, with the AI proactively guiding users toward solutions, recommendations, or further actions, ultimately making the conversation more effective and satisfying.
Start by identifying what users want to accomplish. Map out key intents for your workload, and design targeted interactions that match those intents. Study user behavior, preferences, and contextual cues to anticipate the various ways that users might communicate their intentions.
Optimize how the AI interprets input
At the core of any conversational AI is its ability to accurately interpret and understand human language. This ability involves natural language understanding (NLU), which enables the AI to discern the user's intent, even when they phrase their requests in unconventional or varied ways. Users often communicate in a natural, unstructured manner. AI systems must be designed not only to process structured commands but also to handle open-ended questions and informal language. To ensure that the AI can navigate these complexities and deliver accurate and relevant responses, it's critical that you optimize how the AI interprets user input.
Anticipate the various ways that users might phrase their requests and structure the AI's responses to accommodate different intents. For example, enable the AI to handle both structured commands and open-ended questions to ensure that users feel understood, regardless of how they phrase their requests. Design AI responses to mirror natural human communication. The AI's replies should sound conversational and context-aware, to reflect how people typically interact.
For more complex user inputs, break down the conversation into manageable steps. Guide users through a series of questions or actions that help resolve their issue without overwhelming them. If a user provides a request that involves multiple pieces of information, the AI should guide the user through the process step by step. This approach avoids overwhelming the user with too many questions at once and helps clarify intent. A structured flow of follow-up questions ensures that all necessary details are captured without confusion. At the same time, it helps maintain a natural conversation.
Guide the user through interactions
Conversation flows determine how the conversation progresses based on what the user says or chooses. A good conversation flow should always be goal-oriented. Each step in the interaction should bring the user closer to completing their objective, whether it's scheduling a meeting, retrieving information, or resolving an issue. By designing with the user's goal in mind, you ensure that the conversation remains focused and purposeful. In this way, you help reduce frustration and improve the overall experience.
To create an effective conversation flow, first map out all possible interactions that users might have with the AI. This process includes designing clear decision trees that branch out based on different user responses or choices. By anticipating the various paths that a conversation might take, you ensure that the AI is prepared to handle both expected and unexpected inputs. The conversation flow should account for not only responses to questions but also possible follow-up questions if details are unclear or incomplete.
In a well-designed conversation, the natural back-and-forth between the user and the AI mirrors the dynamics of human interaction. The AI must not only respond to user inputs but also recognize when it should take the initiative; for example, by prompting the user for more information, asking clarifying questions, or by providing suggestions. Design the conversation to include moments where the AI can offer options, confirm user choices, or propose next steps so that the interaction feels cooperative and engaging.
An effective conversation handles interruptions and deviations gracefully. Users might change their mind mid-interaction or ask unrelated questions. A flexible conversation flow enables the AI to manage these situations effectively. For example, it might pause the current task to address a new query and then return to the original task when appropriate. Ensure that the AI can maintain context but still accommodate dynamic user behavior.
To enhance the fluidity of the conversation, transitions between different stages of the flow must be managed with care. When the AI shifts from gathering information to carrying out an action, it should clearly communicate this transition to the user.
Design fallback mechanisms
Conversations don't always follow a predictable path. Users might enter unexpected, unclear, or off-topic responses that the AI struggles to understand. Good conversation design prepares for these moments by ensuring that the AI can recover gracefully and continue to keep the interaction productive and user-friendly. Don't design the AI to end the conversation abruptly if it encounters something that it doesn't understand. Instead, design the flow to prompt the user with clarifying questions, offer alternative suggestions, or redirect the conversation in a way that keeps the user engaged and moving toward their goal.
Plan for edge cases. Sometimes, users say something unexpected, irrelevant, or out of context. Design the AI to include fallback responses that it can use to help put the conversation back on track. For example, if a user's input is too vague, the AI shouldn't just say, "I didn't understand that." Instead, design it to ask a clarifying question such as, "Could you provide more details about what you need?"
If users provide incomplete information, the AI should prompt them with contextual follow-ups. For example, if a user says, "Schedule a meeting," but doesn't specify a time or participants, the AI might ask, "What time would you like the meeting to be?" or "Who should be invited?" In this way, the AI can fill in gaps in the user's input without causing frustration.
For more complex scenarios, where a user might be asking for something that the AI can't handle or something that is off-topic, design the fallback responses to offer alternative suggestions. For example, imagine in a system for scheduling meetings that a user asks, "Can you book a flight?" Because booking a flight is outside the AI's capabilities, the AI might respond, "I'm currently able to schedule meetings, but I can also help with other tasks, like drafting emails or managing your calendar." These types of alternative paths help keep the interaction flowing smoothly and avoid alienating the user. They also help the user understand what the AI is capable of doing.
Anticipate repeated misunderstandings. If the AI fails to understand a user multiple times in a row, the fallback strategy should offer escalation through alternative solutions. For example, it might redirect the user to a human for assistance or provide links to relevant help documentation. This approach ensures that the user feels supported, even when the AI's limitations are reached.
Power Platform facilitation
Copilot Studio includes prebuilt entities that are designed to understand and categorize common user intents across a variety of domains. Use the prebuilt entities to map user intents to the relevant actions or responses without having to build everything on your own. Prebuilt entities are optimized to handle common variations in user input.
You can use agent templates as a starting point for building copilots. The templates are preconfigured with core conversational patterns, capabilities, and workflows that are designed to handle common tasks and scenarios. They incorporate built-in intents, entities, and dialogue flows tailored for specific use cases, such as customer support, productivity tasks, or FAQs. You can customize the templates to meet the specific needs of your workload and users.
Provide custom instructions in your copilot through prompt modification, to gain better control over how the AI engages with users and answers specific questions. By altering the underlying instructions that shape the copilot's comprehension and language generation, prompt modification enables developers to tailor the AI's behavior, responses, and emphasis. Because instructions can be customized to fit various use cases, sectors, or company requirements, prompt modification ensures that the copilot offers contextually appropriate and useful replies.
Configure the fallback topic that a copilot can use when it doesn't understand a user's input or can't handle the request. The fallback topic acts as a type of "safety net." It provides a preset response that keeps the conversation on track and prevents user frustration when the AI reaches the limits of its capabilities. By configuring the fallback topic, you control how the copilot responds when it can't interpret a user's intent. For example, the copilot can politely inform the user that it didn't understand the request and then offer helpful suggestions, such as asking the user to rephrase the query or providing a set of related options or commands that the AI can process. In more advanced configurations, the fallback topic can guide users to other resources or escalate the conversation to a human representative.
Additional information
- Intelligent application workloads
- Introduction to conversational experiences
- Conversational AI best practices