Why chat plus citations is a dead end for vertical AI, and how to shift toward delegation-style products that actually do the work for your users.
Chat and citations won't save your vertical AI - Atul Ramachandran, Filed Inc · Atul Ramachandran
15 min total·Actually worth watching closely: ~4 min·2 must-watch clips
- 0:01 – 2:30Listen
Why chat and citations won't save your vertical AI
Opens straight at the pain: chat is a synchronous medium that keeps users waiting on the platform, and citations throw the burden of verifying results, one by one, back to the user. In healthcare, legal, and taxes, that adds work for professionals rather than removing it.
Chat and citation interfaces can't keep the promise of saving your customers time and money. They make users do more work, not less.
Pure spoken argument with nothing on screen; fine to listen to this problem statement while commuting or doing something else.▶ Jump to 0:01Speaker · Atul Ramachandran - 2:30 – 5:00Listen
Three levels of abstraction in product value
Walks through three modes of delivering value: an employee doing the task, capped by how many employees a company has; digital self-serve, capped by how many users visit; and agentic delegation, where users hand off a task and leave, so value is no longer tied to time spent on the platform.
Only agentic delegation breaks the ceiling of value being capped by how many times a user visits.
A conceptual framework delivered verbally, with nothing worth watching; just follow the three-step progression.▶ Jump to 2:30Speaker · Atul Ramachandran - 5:00 – 7:03Listen
The conveyor belt metaphor: design for delegation
Introduces the core mental model of the talk: think of your product as a conveyor belt, with AI agents as the workers and the user as the supervisor who comes in to delegate tasks and monitor progress. Design for delegation, not participation.
The design question shifts from how users use it to how users hand work off with confidence.
The metaphor is developed in speech rather than on screen, but it's the key to the second half, so worth full attention.▶ Jump to 5:00Speaker · Atul Ramachandran - 7:04 – 9:20Listen
Where to start: long-running agents and skill capture
The first step in practice: find tasks that take your users more than an hour or two and are repeatable, and build long-running background agents for them. An end-to-end agent only gets eighty to ninety percent of the way; the last twenty percent of personalization has to come from skills captured automatically from product usage rather than created by hand.
The real moat is the last twenty percent, doing the work the user's way, and skills should be learned from behavior.
Verbal walkthrough using Filed and Wispr Flow as examples, no live demo; the value is in the method.▶ Jump to 7:04Speaker · Atul Ramachandran - 9:20 – 11:43Listen
Visibility: task lists and traces
Agents run long and many tasks run in parallel, so you need a task list to track where each one stands, plus traces that let every value the AI produced be traced back to its source.
Most user complaints happen where visibility is missing. Traces are the infrastructure of trust.
This covers product mechanics rather than live operation; grasping the intent behind traceability is enough.▶ Jump to 9:20Speaker · Atul Ramachandran - 11:46 – 13:20Listen
Control: takeover, the self-serve layer, and plan approval
Users only delegate when they're confident they can take back control. Takeover should feel like pausing the belt, fixing the problem, and starting it back up (Filed's agents pause wherever they'd make an assumption, and users reply by tagging the agent as they would in Slack). You also need to keep the self-serve layer in your product, and irreversible, dangerous actions must present a plan for approval before running.
Trust isn't won with accuracy; it's won with a control mechanism that lets users take the wheel at any time.
Mostly a verbal description of Filed's interaction design, with no live demo; catching the three control mechanisms is enough.▶ Jump to 11:46Speaker · Atul Ramachandran - 13:20 – 15:10Listen
The new north star: weekly active sessions instead of WAU
Closes with a shift in measurement: WAU is fundamentally at odds with a delegation product, so switch to weekly active sessions, tasks completed by a human or an agent whether or not the user is on the platform. The healthy signal is sessions going up while WAU goes down, though not to zero.
The less users come in and the more the platform completes, the better the product is working.
Pure argument and the most disruptive conclusion in the talk; worth listening through and writing down the metric definition.▶ Jump to 13:20Speaker · Atul Ramachandran