Agentic AI systems are emerging as a key frontier in advancing intelligence, with early adoption seen in areas like deep research, software development, and customer service. Despite their promise, current systems struggle with reliability and can be unpredictable for simpler tasks as well. This limits their use to tasks where lower reliability can be managed. To unlock broader applications, we need to rethink how these systems are built. By designing workflows that incorporate human-in-the-loop interfaces, we can balance AI-driven execution with human-guided planning and ideation. This talk will showcase how such an approach can enable more complex, high-stakes tasks-demonstrated through a real-world deep research example.

The practical implementations of Human-In-The-Loop Agentic Systems span a variety of complex tasks. In deep research, these systems can assist in navigating vast amounts of information and synthesizing insights. For consumers, they can facilitate complex planning, such as organizing intricate travel itineraries or guiding high-value purchases by providing structured information and suggestions. Businesses can leverage these agents for critical planning activities like optimizing supplier selection, streamlining inventory management, and enhancing overall business process management. These applications highlight how human-in-the-loop design can elevate the reliability and effectiveness of AI for demanding and high-stakes scenarios.