Juji, Inc.

Top User Intents in Human-AI Interactions In a mixed-initiative engagement—the kind that human-centered AI agents are designed to support—user input can take many forms. Sometimes, users provide a direct response to the AI agent's prompt (e.g., answering a question or confirming a step). But often, users take initiative to ask questions, make requests, express emotion, or simply go off-topic. These user-initiated actions can signal a wide range of user intents, each requiring a thoughtful and appropriate AI response. Below are 10 intent categories we've observed across a wide range of task-based AI applications. 1. Information Seeking A user asks a question to obtain more information about a task, the ongoing engagement, or the AI agent. A user may also ask off-topic questions irrelevant to current engagement. Examples: • "What are the financial aid options?" • "What do you mean?" • "What do you know?" 2. Information Disclosing A user provides certain information in response to an AI agent's question or voluntarily. Examples: • "My major is Marketing." • "My favorite movie is The Sting." 3. Service Request A user requests a particular service. Examples: • "Could you help me practice job interview?" • "I'd like to update my insurance info." 4. Engagement Management A user requests a particular way of handling the current engagement, e.g., pause, continue, go back, or jump ahead. Examples: • "Could you pause for a second?" • "Can we restart?" 5. Social Engagement A user socializes with an AI agent by expressing gratitude, satisfaction, frustration, etc. Examples: • "Thank you very much for your help!" • "I really enjoyed our chat!" 6. Response Avoidance or Excuses A user is unwilling to respond to an AI agent's question or request. Examples: • "I don't want to answer this." • "I cannot tell you, it's confidential." • "I don't care, you decide." 7. Task Struggle A user has difficulty performing a task requested by an AI agent, such as answering a question or performing an action. Examples: • "This is a hard question." • "It's really difficult to do." 8. Other Struggles A user expresses struggles physically or emotionally. Examples: • "I feel sick." • "I'm so depressed." 9. Error Handling A user reports a system/agent error or exception to an AI agent. Examples: • "You have a bug." • "You are stuck." 10. Abusive/Improper Behavior A user responds with improper content, such as violent or hostile comments. Examples: • "You are so stupid!" • "I will kill you." Human-centered AI agents must be able to detect, interpret, and respond appropriately to all of these intents. Doing so is critical not just for task completion, but for building trust, resilience, and real-world usability. Have you seen these behaviors in your own conversations with AI tools? Are your agents prepared to handle them? Share your observations! #HumanCenteredAI #ConversationalAI #AgenticAI #UserIntent #Chatbots #MixedInitiative #IntentRecognition #HumanAIInteraction 

Human-AI Interaction Patterns: A New Design Language for Smarter Agents Generative AI and large language models (LLMs) have made it easier than ever to build AI agents that understand and generate natural language. But there's still a major challenge: enabling these agents to handle the diverse, nuanced, and nonlinear ways humans actually interact, especially in complex domains like advising, coaching, or counseling. That's where Human-AI Interaction Patterns come in. Just as software design patterns help developers build robust systems, interaction patterns provide a structured way to design AI agents that can interpret, respond to, and adapt to real-world user behavior—even when users deviate from scripted workflows. Based on our research and a review of work in HCI, NLP, and conversational UX, we’ve identified five foundational interaction patterns (see the figure below) that frequently occur when an AI agent initiates a task and the user responds in various ways: 1. Progression Pattern The user cooperates with the AI agent to successfully complete the initiated task (T1 in the figure) . 2. Expansion Pattern The user cooperates partially, and the current task (T1) is expanded into sub-steps (T11 and T12), which further engage the user to ensure its completion. 3. Digression Pattern The user deviates from the task (T1) by initiating a new task (e.g., asking for clarification) or making a side comment (e.g., social commentary). The AI agent must handle the digression properly and return to the original task. 4. Regression Pattern The user amends or revisits a previously completed task (T1). The AI agent must determine how to update affected task(s), including redoing one or more previous steps if necessary. 5. Transgression Pattern The user's current response impacts future tasks (e.g. T2)—by jumping ahead or providing information prematurely. The AI agent must update the workflow to avoid asking redundant questions and ensure completeness. These patterns aren't edge cases—they're commonplace user behaviors. Recognizing and responding to them correctly is essential to: • Maintain task coherence • Avoid user frustration • Ensure data quality • Build trust and satisfaction This taxonomy offers practical building blocks for anyone designing AI systems that must engage in rich, multi-turn conversations with real users. Have you seen these patterns in your own use of AI tools? If you're building AI agents, do your users exhibit these behaviors? Let's compare notes. #HumanCenteredAI #ConversationalAI #InteractionPatterns #AIInteractionPatterns #AIWorkflow #AIChatbots #HumanAIInteraction #AgentDesign #TaskCentricAIAgent

Thank you Rodney Brooks for sharing this article - one of the best on AI since AI became hot again. It highlights AI’s strengths and weaknesses and how/where AI can offer us humans the realistic or best ROI. It’s a must-read for those who want to build, adopt, or invest in AI. My favorite one is point #5. #AI #AIforReal FYI Wenxi Chen Jie Lu, Ph.D. Huahai Yang Henry Truong