Valon Limits AI for Its Hires and Aims for 75% Cost Reduction

Valon is now regulating access to models for its new recruits, with a stated goal of significantly reducing token expenses. Management describes errors and automation reflexes that could hinder the learning of the job. Exceptions exist, and the initial expected effects are quantified against a broader debate surrounding the use of AI at work.

Targeted token expenses between $4 and $5 million instead of $15 to $20 million

Andrew Wang anticipates that Valon's annual token expenses will decrease this year to around $4 to $5 million, down from the previous range of $15 to $20 million. According to him, the new policy could reduce these costs by approximately 75%. So far, he claims to have encountered no internal opposition and indicates that experienced employees have welcomed the measure after having to correct errors generated by the AI of new recruits. Since the implementation of the policy, new recruits have been seeking more help from experienced colleagues, as they are not allowed to rely on AI to solve problems.

Access to AI subject to managerial approval and exemption for engineers

The directive applies to new recruits in nearly all departments, including senior-level positions. They are only allowed to use AI if their manager believes they can identify errors in the tools. Andrew Wang specifies that engineers are exempt, as all code is reviewed by peers before publication. He emphasizes that in finance or human resources, there are no comparable safeguards. The measure was instituted last month after observing broad access to AI systems.

Stated motivations: models too costly and weakened judgment

Andrew Wang has observed that employees were using the most expensive AI models for simple tasks, a practice he considers costly and not conducive to new recruits learning the job. He asserts that excessive use of AI prevents some new employees from fully understanding their roles and generates errors. When questioned about these choices, employees responded that AI was almost always correct, which he believes reduces the tendency to question results and exercise their own judgment. He argues that doing the basic work without resorting to automation helps develop this understanding.

Profile of Valon and the acknowledged paradox around AI

Based in New York, Valon has about 320 employees and develops mortgage management software based on AI agents. In 2024, the company was valued at $1.75 billion and raised $275 million from investors such as Andreessen Horowitz, WestCap, and 166 2nd. Andrew Wang highlighted the paradox of introducing such a rule in a company that makes advanced use of AI. He previously worked as an analyst at Goldman Sachs.

Broader corporate context: generated content and "meat proxies"

In September 2025, a survey conducted by BetterUp and Stanford's Social Media Lab revealed that 40% of the 1,150 full-time American office employees had received AI-generated work from a colleague in the previous month, with each case requiring an average of nearly two hours of processing. This phenomenon gave rise to the term "meat proxies" to describe those who share unverified AI results, a term popularized by German developer Niklas Gruhn in a blog post on August 3. Outside of Valon, some AI enthusiasts have criticized the company's policy, to which Andrew Wang responds by inviting suggestions for better methods to ensure learning.