AI is writing almost all startup code. That's creating a new problem.

At Alma, a Menlo Ventures-backed AI nutrition coaching app, nearly every line of code is now written by AI.

"I'm not exaggerating," said Rami Alhamad, Alma's cofounder and CEO. "Nearly everything we ship now is AI-generated."

In a survey of more than two dozen startup founders and venture capitalists, Business Insider found that AI has very quickly become the primary author of startup code, with Anthropic's Claude Code the overwhelming tool of choice. But that speed is a double-edged sword, with founders worrying about slop and low quality.

Coding has become one of generative AI's clearest business use cases, with VCs pouring billions into AI coding startups such as Lovable, Replit, and Cursor. Last week, SpaceX announced it would acquire Cursor for $60 billion. Anthropic has filed paperwork to go public, likely later this year.

Writing AI-generated code is like going from a hand saw to power tools in woodworking, according to Dan Lorenc, cofounder and CEO of Chainguard, an open source cybersecurity company. It is powerful yet dangerous.

"AI showed up and gave everyone a circular saw," he said. "It's way faster, but also a lot easier to lose a finger. Today, everyone is figuring out what guardrails to put in place to do this safely."

Lorenc said 100% of his code is now created via Claude Code. Last year, he put that figure at 60%.

"A year ago, you would write code yourself, and the LLMs might save you a bit of time typing," he said. "In the past four to six months, the models, the tool calls, and the harness got really good. You still have to prompt and steer it, but you can crank out in hours or days what would have taken weeks or months before."

At Wordsmith AI, an AI platform for legal teams, CTO and cofounder Volodymyr Giginiak said the company's code is "nearly 100%" AI-generated.

"Humans write very little code directly," Giginiak said. "The distinction is no longer who writes the code, but how much autonomy the AI has."

Giginiak said fully autonomous tasks account for about 10% of work today, but he expects that to rise quickly. A year from now, he predicted "80—90% of tasks" will be fully autonomous.

"Engineering is not disappearing, but being fundamentally restructured," he said. "The highest-leverage engineers will be those who can design the right environments and context for AI to operate in."

The problem: AI can generate code faster than startups can trust it

In our survey, there was a clear downside to all the AI-generated code: Lots of slop and bugs.

"The trend I'd flag for 2026: the 'vibe coding' bubble will produce a wave of fragile, unmaintainable products built by people who can't support them beyond launch," said Jason Alan Snyder, a futurist and cofounder of SuperTruth and Artists & Robots.

A December report from Menlo Ventures, which was an early backer of Anthropic, called this the "Cleanup Tax."

"The speed gains in writing code can be offset by the time spent on cleanup and quality assurance, an 'ROI Paradox' that complicates the simple productivity narrative," the report said.

At Blueprint, which is building the AI operating system for therapists, nearly all the company's code is now written by AI, up from 40% in August, according to CEO and founder Danny Freed.

He appreciates that the cost of "trying things" has dropped dramatically, but says his human employees have only become more valuable.

"Taste and judgment matter more than ever," he said. "Just because something can be built doesn't necessarily mean it should be built."