Y Combinator CEO: Corporate friction helps AI startups win

This article summarizes an episode of a16z’s video series featuring Garry Tan, CEO of Y Combinator.

Garry Tan, president and CEO of Y Combinator / Photo credit: Garry Tan

A simple text file can now do the work of a human employee. Y Combinator CEO Garry Tan says founders should stop chasing tech trends. Instead, they should use AI to turn their industry knowledge into large businesses.

AI redefines startup headcount

Companies with customer insight are using automated systems to replace traditional software interfaces. Tan says identifying the right bottleneck to automate is what matters most.

“You can go from zero to US$15 million ARR in about four months with two or three people, hundreds of agents, and a few hundred skill files,” he explains.

This shift alters how early-stage companies scale operations:

• Headcount no longer correlates with company maturity or revenue generation.
• Small teams use hundreds of AI programs to take over repetitive tasks and improve company operations.
• Customers prefer paying small groups for guaranteed results over buying software platforms to manage themselves.

Bureaucracy protects disruption

While small teams scale faster with automation, corporate friction creates advantages for emerging companies:

• Long approval chains prevent large organizations from adopting new tools quickly.
• Software investments fail when trapped in management layers.
• Small businesses maintain a competitive edge because they adapt at a faster pace.

Tan views this institutional slowness as an advantage for founders. He argues that the layers of middle management slowing down large companies serve as a protective barrier, giving startups the time they need to develop competing products.

Turning expertise into a permanent employee

Exploiting this corporate sluggishness requires founders to ignore hype and focus on domains they understand.

Leaders must follow a sequence to convert experiments into autonomous routines:

• Train AI models: Invest resources to explain leadership decisions to the system when learning justifies the computing cost.
• Identify errors: Allow the agent to run tasks repeatedly until patterns and mistakes become obvious.
• Automate templates: Trigger the workflow on a schedule without needing human reminders or planning meetings.

Once a workflow runs independently, Tan argues that companies must go beyond executing tasks by choosing instead to “skillify it” and turn the process into a markdown file that functions as a permanent employee.