Chat and citations won't save your vertical AI - Atul Ramachandran, Filed Inc

The broken promise of chat and citations

Atul opens with a provocative claim: the two interfaces most AI products rely on, chat and citations, cannot deliver on the core promise vertical AI companies make to customers. The pitch is that AI agents will save time and money by doing work while users sleep. But chat is synchronous, meaning users must wait for responses, and citations put the verification burden back on the customer, requiring them to review outputs one by one. In high-stakes verticals like healthcare, legal, and taxes, this extra verification work breaks the value proposition entirely.

Filed's rapid growth validates the thesis

Atul introduces himself as CTO and co-founder of Filed, a company building products for US tax professionals that has raised over $17 million. The company has seen explosive growth: in just the last month, they closed more revenue than they did in their entire first year. After two years building AI agents for the tax industry, Atul argues the learnings transfer to any vertical AI product.

Three abstraction layers in product history

Atul traces three abstraction layers in product history. First, physical presence: users delegated tasks to employees, and company value was bottlenecked by headcount. Second, digital transformation: users performed tasks themselves via apps, and value scaled with user count. Third, the current shift to agentic delegation: users delegate work to AI agents, removing the user bottleneck entirely. Agents can work while users sleep, unlocking value generation that was previously impossible.

The conveyor belt metaphor for agentic products

Atul offers a mental model: think of an agentic product as a conveyor belt. The AI agents are the workers on the belt, and the user is the supervisor who delegates tasks and monitors progress. This framing clarifies what tools you need to build. The product is the infrastructure that keeps the belt running, and the user's role shifts from operator to overseer.

Delegation: target multi-hour tasks

The first component is delegation. Atul advises finding tasks that take users more than a couple of hours, tasks that are repeatable across use cases. These become candidates for long-running background agents. In taxes, Filed identified three such tasks in the workflow. This is where the core value lives: removing hours of work from users' plates.

Skills: capturing the last 20%

The second component is skills. Background agents might get you 80-90% of the way there, but users want the work done their way. Skills capture the last 20%, the quirks and preferences that define professional work. Filed captures skills automatically from product usage rather than forcing users through a separate interface, similar to how Gmail learns from behavior over time.

Monitoring and control: building trust

The third and fourth components are monitoring and control. Users need visibility into what agents are doing through task lists and traces. They also need confidence they can intervene when something goes wrong. Atul uses the metaphor of taking the wheel, not abandoning the car. Filed pauses agents when assumptions are needed and lets users respond via chat-like interactions to resolve conflicts.

Flip your metrics: sessions up, users down

Atul closes with a metrics shift. Weekly active users made sense when users did the work themselves. In an agentic world, the right metric is weekly active sessions, tasks completed by humans or agents even when the user is offline. The goal: weekly active sessions go up while weekly active users go down. That proves users trust the platform enough to delegate and leave.