Atul Ramachandran: Why Chat and Citations Alone Won't Save Your Vertical AI

In the rapidly evolving world of artificial intelligence, many startups are racing to build vertical AI solutions for industries like healthcare, legal, and taxes. While the promise is often to save time and money by having AI agents perform tasks while users sleep, the current interfaces of chat and citations are falling short. This is the core thesis put forth by Atul Ramachandran, CTO and Co-founder at Filed, in a recent YouTube presentation. Ramachandran argues that these standard interaction methods, while useful, ultimately break the promise of true AI-driven efficiency by placing too much burden back on the user.

Companies working on this

StartupHub profiles of the companies this article names, with funding and a one-liner from our database.

Descript

$50M

Descript is an AI-powered video and podcast editor that simplifies content creation by allowing users to edit media as easily as text documents.

Naver

$676M

South Korean internet conglomerate offering search, AI, cloud, e-commerce, and content services.

Outbrain

$394M

A leading technology platform that drives business results by engaging people across the open internet.

Glance

$390M

AI-powered lock screen content platform with 450M+ users globally

Who Is Atul Ramachandran

Atul Ramachandran is the CTO and Co-founder of Filed, an AI company focused on tax professionals in the US. With over a decade of experience in product development, Ramachandran has overseen significant growth at Filed, which has raised more than $17 million. His insights are drawn from two years of building and scaling AI agents in the complex tax industry, experiences he believes are transferable to any vertical AI product.

The Shortcomings of Chat and Citations in Vertical AI

Ramachandran acknowledges the initial benefits of chat and citations. Chat interfaces offer flexibility, allowing users to interact with AI agents in a conversational manner for various tasks. Citations are crucial for grounding AI responses in truth, reducing hallucinations, and enabling users to verify outputs. However, these benefits come with significant drawbacks in a vertical AI context.

The primary issue, according to Ramachandran, is that "Chat is synchronous." This means users must actively wait for the AI's response, preventing them from leaving the platform to perform other work. Similarly, "Citations also puts the verification burden back into the customer." In highly regulated or detail-oriented fields like healthcare, legal, and taxes, this verification process adds substantial manual work, negating the promise of time and cost savings. Users often feel that the AI isn't truly working for them while they sleep if they still need to meticulously review every output.

The Evolution of Product Abstractions: From Physical to Agentic Delegation

To illustrate his point, Ramachandran outlines three historical levels of product abstraction:

• Physical: This era involved direct human interaction, like visiting a bank branch and speaking with an employee. The bottleneck was the number of employees.
• Digital Transformation: The rise of online banking and mobile apps shifted the bottleneck to the number of users. More users meant more value.
• Agentic Delegation: The current and future state, where users delegate long-running tasks to AI agents. Here, the bottleneck of user interaction is removed, as agents can perform work autonomously. "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." This paradigm shift allows for unprecedented value generation.

Ramachandran conceptualizes an agentic product as a "conveyor belt," with users acting as supervisors. Users delegate tasks to AI agents (the workers) and monitor their progress, intervening only when necessary. This model necessitates a different approach to product design.

Building the Agentic Conveyor Belt: Four Key Components

Ramachandran identifies four essential components for building a successful agentic product:

1. Delegation: The product must allow users to hand off tasks to AI agents. These tasks should be repeatable and take a significant amount of human time (e.g., more than a couple of hours), making them ideal for long-running background agents. For example, in tax preparation, tasks like tax prep, review, and planning often exceed two to four hours.
2. Teach (Skills): Predefined agents can get users 80-90% of the way, but the last 20% lies in capturing user-specific preferences and quirks. This requires the ability for users to "teach the agents how your users do the work." This can be achieved through automatic skill learning based on product usage, similar to how tools like Descript learn from user behavior.
3. Monitor: Since agents perform long-running tasks, users need clear visibility into their progress and outcomes. This involves building task lists and detailed traces within the product, showing how the agent executed each step. "This is where the trust is built," Ramachandran emphasizes.
4. Control: Users must feel confident that they can intervene if something goes wrong or requires human judgment. The system should allow users to "pause the belt, fix the problem, and start it back up again." This could involve pausing an agent when it makes an assumption and allowing the user to provide direction, much like tagging an agent in a chat platform.

Bonus Points for Trust and Measurement

Ramachandran offers two additional crucial considerations. First, Level 2 (digital transformation) features will not disappear with agentic delegation. Users will only delegate work if they trust they can regain control. Therefore, Level 2 features must be integrated to build this trust. For instance, in sensitive operations, presenting a plan for user approval before execution is vital, such as a tax software generating a data entry plan that users can review before data is modified.

Second, the way product success is measured needs to evolve. Traditional metrics like weekly active users (WAU) are insufficient. Instead, Ramachandran advocates for "weekly active sessions (WAS)." This metric accounts for tasks completed by a human or an agent, even when the user is not actively on the platform. The ideal scenario is for weekly active users to decrease while weekly active sessions increase, indicating successful delegation and agent autonomy.

Key Takeaways for Vertical AI Builders

In summary, Ramachandran's presentation offers three critical takeaways for developers and product engineers in the vertical AI space:

• Design for delegation, not participation, enabling users to hand off tasks rather than perform them directly.
• View your product as a conveyor belt, where users are supervisors overseeing autonomous agents.
• Measure what work is done, not time on the platform, by tracking weekly active sessions instead of weekly active users.

By adhering to these principles, AI and startup companies can build truly impactful vertical AI products that deliver on the promise of saving time and money for their customers.