Atul Ramachandran of Filed Inc. argues that effective vertical AI products require moving beyond chat and citations to enable autonomous agentic delegation, where users supervise AI agents performing complex, long-running tasks tailored to their preferences. He emphasizes building trust through control and monitoring, integrating traditional features, and adopting new success metrics focused on agent-driven task completion rather than user activity.
Atul Ramachandran, CTO and co-founder of Filed Inc., discusses the challenges and evolution of building AI products in vertical industries such as healthcare, legal, and taxes. He emphasizes that relying solely on chat interfaces and citations for AI agents is insufficient to fulfill the promise of saving customers time and money. While chat allows flexible user input and citations provide verifiable outputs that reduce hallucinations, these tools are synchronous and place the burden of verification on users, which contradicts the goal of agents working autonomously on behalf of users.
Ramachandran outlines the historical evolution of product interaction through three levels of abstraction. Initially, users interacted with human employees directly, limiting value creation to the number of employees. The digital transformation shifted this bottleneck to the number of users interacting with online platforms. Now, with agentic delegation, users delegate long-running tasks to AI agents, removing the bottleneck of user presence and enabling continuous value creation as agents work independently, akin to workers on a conveyor belt supervised by users.
To build effective agentic products, Ramachandran identifies four essential components: identifying delegatable tasks that save significant user time, enabling users to teach agents their unique work preferences through customizable skills, providing robust monitoring tools to track agent progress and build trust, and ensuring users retain control to intervene when necessary. This framework transforms users from active participants to supervisors who delegate, monitor, and control AI agents performing complex tasks.
He also highlights the importance of maintaining trust by integrating traditional product features alongside agentic capabilities. Users must feel confident they can take back control if something goes wrong, and irreversible or sensitive actions should require explicit user approval through clear plans. This balance ensures users do not feel abandoned by the system but rather empowered to oversee and guide AI agents effectively.
Finally, Ramachandran calls for a shift in how product success is measured. Traditional metrics like weekly active users are less relevant in an agentic delegation model. Instead, weekly active sessions—counting tasks completed by agents even when users are not actively engaged—better capture value creation. By designing for delegation, viewing products as conveyor belts with users as supervisors, and adopting new success metrics, companies can build more successful vertical AI products that truly deliver on their promises.