AI startup Valon has introduced a policy requiring most new employees to learn their jobs without artificial intelligence before they are allowed to use the technology, as concerns grow that widespread reliance on AI could weaken workers’ judgment and leave companies paying heavily for tools that employees may be using unnecessarily.
Andrew Wang, Valon’s CEO and cofounder, introduced the policy last month after giving employees broad access to AI tools and reviewing how they were being used. He said employees were frequently turning to the most expensive models for relatively simple tasks, raising concerns about both soaring computing costs and the effect on employees’ ability to understand their work.
“By doing the basic work, rather than relying on AI, you start to form an understanding,” Wang, a former Goldman Sachs analyst, told Business Insider.
The policy is an unusual position for Valon, a New York-based company that builds mortgage-servicing software powered by AI agents. The company employs about 320 people and has made artificial intelligence central to its products and operations.
“It’s a weird decision for a company that is at the frontier of AI usage,” Wang wrote in a blog post explaining the policy.
But Wang said the decision followed a pattern he observed after employees were given unrestricted access to AI. When he asked why they were using the most powerful models for basic assignments, employees told him the systems were almost always correct.
For Wang, that answer was itself a warning sign.
He said employees could become less inclined to question AI-generated answers or develop the expertise needed to recognize when a system is wrong. That creates a potential problem for companies deploying increasingly capable AI systems: workers may become more productive in the short term while losing the underlying knowledge required to supervise those systems effectively.
Under Valon’s policy, new hires in almost every part of the business, including senior employees, must initially work without AI. They can begin using AI only after their managers determine that they understand their responsibilities well enough to identify incorrect AI output.
Engineers are exempt because Valon requires all code to undergo peer review before it is released. Wang said functions such as finance and human resources do not have equivalent safeguards, making unrestricted AI use more difficult to justify.
The policy also highlights a less obvious cost of the AI boom. While companies have focused heavily on how much AI can save by automating work, the technology can create additional costs when employees use powerful models for tasks that do not require them.
Wang said Valon’s annualized spending on AI tokens is now expected to fall to roughly $4 million to $5 million this year, from an estimated $15 million to $20 million previously.
The savings come alongside what Wang considers a more important benefit: new employees are now asking experienced colleagues for help rather than asking an AI system to solve problems for them. That interaction recreates an older form of workplace training in which junior employees acquire institutional knowledge by observing experienced colleagues, performing basic tasks and gradually taking on more complicated responsibilities.
The issue is becoming bolder as companies deploy AI across entry-level functions that traditionally served as training grounds for young workers. If AI takes over those tasks immediately, employees may reach more senior positions without having developed the practical judgment that those assignments were intended to teach.
The concern extends beyond Valon.
A September 2025 survey by BetterUp and Stanford’s Social Media Lab found that 40% of 1,150 full-time U.S. desk workers had received AI-generated work from a colleague during the previous month. Respondents said dealing with each instance took nearly two hours on average. That suggests companies can incur a hidden productivity cost when employees submit AI-generated material that colleagues must check, correct, or rewrite.
The emergence of the term “meat proxies” reflects the growing frustration with this behavior. The phrase, popularized by German software developer Niklas Gruhn in an August 3 blog post, describes people who effectively act as intermediaries for unchecked AI output, passing machine-generated work to others without adequately reviewing it.
For companies, the central issue is not necessarily whether employees should use AI, but whether they have sufficient expertise to supervise it.
That could get into play more as AI agents move from generating text and code to carrying out multi-step tasks with limited human intervention. The more responsibility companies delegate to AI, the greater the need for employees who understand the underlying processes well enough to detect errors and intervene when systems go off course.
Valon’s approach points to a reversal of the assumption that faster AI adoption is always better. Instead of measuring success simply by how much work AI can perform, Wang is placing greater emphasis on whether employees understand the work being automated.
The policy has not faced significant internal opposition, according to Wang. He said experienced employees have generally welcomed it because they had been spending time correcting poor-quality AI-generated work produced by newer hires.
Still, the approach has drawn criticism from some AI enthusiasts, including people who responded to a LinkedIn post in which Wang described the policy.
Wang’s response is straightforward: “If you have a much better idea here of how to make sure people learn, please tell me.”
Valon’s experiment could offer an important lesson for companies racing to integrate AI into their operations. The technology may reduce the need for humans to perform routine tasks, but eliminating those tasks entirely could also eliminate some of the mechanisms through which workers acquire expertise.
