At the AI startup Valon, a new policy requires most new hires to learn their jobs the old-fashioned way — without AI.
Andrew Wang, CEO and cofounder, said he instituted it last month after Valon gave employees broad access to AI. When he examined their usage, he said he found employees turning to the priciest models for even simple tasks, a habit he worried was both costly and preventing newer workers from developing a real grasp of their jobs.
“By doing the basic work, rather than relying on AI, you start to form an understanding,” Wang, a former Golden Sachs analyst, told Business Insider.
New York-based Valon, which has roughly 320 employees, makes mortgage-servicing software powered by AI agents. In a recent blog post describing the new mandate, Wang acknowledged the irony of the move.
“It’s a weird decision for a company that is at the frontier of AI usage,” he wrote.
Wang told Business Insider it was necessary, though, because of the common response he got when he asked employees why they were using the most powerful AI models for simple assignments. They told him, he said, the AI was almost always right. That answer suggested to him that employees were becoming less likely to question AI’s output or develop the judgment to spot mistakes, he said.
The new policy applies to new hires in almost every area of the business, including senior recruits. They can only use AI once their manager is confident they can identify when AI is wrong, said Wang. He added that engineers are exempt because all code is peer-reviewed before release. In functions such as finance or human resources, he said, there are no comparable guardrails.
Valon was last valued at $1.75 billion in 2024 and has raised $275 million in venture funding from investors including Andreessen Horowitz, WestCap, and 166 2nd. So far, Wang said he hasn’t come across any internal resistance to the new policy. If anything, tenured employees have welcomed the change, he said, because they had been cleaning up AI-generated slop from new hires.
The kind of frustration Wang described extends beyond Valon. In a September 2025 survey by BetterUp in partnership with Stanford’s Social Media Lab, 40% of 1,150 full-time US desk workers said they had received AI-generated work from a colleague in the past month. Respondents also said that dealing with each instance took nearly two hours on average.
The problem has become so ubiquitous that techies have coined a nickname for perpetrators: “meat proxies.” Popularized by German software Niklas Gruhn in an August 3 blog post, the term refers to those who share unchecked AI outputs.
Valon’s new policy is already having an impact. Wang now projects the company’s annualized token spending to shrink this year to roughly $4 million to $5 million, from around $15 million to $20 million. Meanwhile, new recruits are seeking help from their more tenured colleagues because they’re not allowed to rely on AI to solve problems. As a result, he said, they are developing a deeper understanding of their work.
Some AI enthusiasts outside Valon have said that the policy is the wrong approach, said Wang, pointing to some comments on a LinkedIn post he made about it. His answer to them, he said, is simple: “If you have a much better idea here of how to make sure people learn, please tell me.”
