Atul Ramachandran, CTO and Co-founder of Filed, argues that the chat-and-citation interfaces dominating vertical AI products today are undermining the very efficiency they promise. His critique, drawn from two years of building AI agents for tax professionals, carries weight for any product team working on industry-specific AI tools.
Ramachandran acknowledges that chat and citations solved real problems. Chat interfaces gave users flexibility to interact conversationally with AI agents across different tasks. Citations grounded responses in verifiable sources, cutting down hallucinations and letting users check the AI's work. But in vertical AI - healthcare, legal, taxes - these same features create new bottlenecks.
"Chat is synchronous," Ramachandran said. Users must wait for the AI to respond before they can move on to other work. And citations, while necessary, "puts the verification burden back into the customer." In regulated fields where accuracy is non-negotiable, reviewing every cited output eats up the time the AI was supposed to save. The result: users feel the AI isn't working for them while they sleep if they still need to scrutinize every result.
The three eras of product abstraction
Ramachandran frames the problem through three historical shifts in how products deliver value. The physical era meant showing up in person - visiting a bank branch, talking to an employee. The bottleneck was how many staff you could hire. Digital transformation moved interactions online through apps and websites, but shifted the bottleneck to the user. More value required more user visits.
The third era, agentic delegation, removes the user bottleneck entirely. "It's no longer the amount of value that you generate is the amount of the number of times the user have visited your platform because agents can do the work while the users have gone to sleep," Ramachandran said. In this model, the product becomes a conveyor belt. Users act as supervisors, delegating tasks to AI agents and stepping in only when something needs human judgment. This shift demands a fundamentally different approach to product design and a rethinking of how teams measure success in AI for Product Development.
Four components of the agentic conveyor belt
Building an agentic product requires four pieces to work together, according to Ramachandran.
Delegation comes first. The product must let users hand off tasks that take significant human time - more than a couple of hours - to AI agents that run in the background. In tax preparation, tasks like prep, review, and planning routinely exceed two to four hours, making them ideal candidates for delegation.
Teach addresses the last mile. Predefined agents get users 80-90% of the way, but the remaining 20% lives in user-specific preferences and quirks. The product needs mechanisms for users to teach agents how they do the work, whether through explicit instruction or automatic skill learning based on usage patterns.
Monitor builds trust. Since agents perform long-running tasks without direct oversight, users need clear visibility into progress and outcomes. This means building task lists and detailed execution traces that show exactly how the agent handled each step. "This is where the trust is built," Ramachandran said.
Control gives users confidence they can intervene. The system should let supervisors "pause the belt, fix the problem, and start it back up again" - whether that means stopping an agent when it makes a wrong assumption or providing direction mid-task. Without this safety valve, delegation won't happen.
Measuring what actually gets done
Traditional metrics fall apart under agentic delegation. Weekly active users (WAU) measures human visits, but the whole point is that humans visit less. Ramachandran advocates for tracking weekly active sessions (WAS) instead - counting tasks completed by either a human or an agent, regardless of whether the user was on the platform at the time.
The ideal signal of a healthy agentic product: weekly active users decrease while weekly active sessions increase. That gap represents work getting done autonomously, which is the core promise of AI Agents & Automation in vertical markets.
Ramachandran also warns that digital transformation features won't disappear. Users will only delegate work if they trust they can regain control. That means building familiar interfaces - like presenting an agent's plan for human approval before it executes sensitive data changes - alongside the new agentic layer.
Why this matters for product development professionals
Ramachandran's framework gives product teams a concrete alternative to bolting chat interfaces onto industry workflows and calling it vertical AI. The shift from designing for participation to designing for delegation changes everything: what you build, how you measure it, and how users experience the product's value. For teams evaluating whether to invest in agentic features, the three-part test is straightforward - can users delegate tasks that take hours, teach agents their preferences, and monitor progress with the option to intervene? If those pieces aren't in place, the product is still asking users to do the work, just through a chat window.
Your membership also unlocks:
