At Valon, a New York-based mortgage-technology company, most new hires are no longer allowed to touch generative AI tools until their manager decides they've earned the privilege. The policy, put in place last month by CEO and cofounder Andrew Wang, came after tenured employees found themselves cleaning up AI-generated work from junior staff who were leaning on chatbots for tasks they hadn't yet learned to do themselves.
Wang told Business Insider that he concluded employees were becoming less likely to question AI output or spot mistakes in it, even as the company built its business on AI-powered mortgage-servicing software. He also said he found employees firing up the priciest AI models for simple tasks, and when he asked them why, the answers convinced him something had gone wrong in how junior staff were learning to work. Under the new rule, new recruits in almost every part of the business, including senior recruits, can't use AI until a manager is confident they can identify when it's wrong.
A Founder Who Calls His Own Policy Weird
Wang, a former Goldman Sachs analyst who cofounded Valon in 2019 with former Twilio engineer Jon Hsu, described the policy as a strange move for a company sitting at the frontier of AI usage. Valon builds mortgage-servicing software powered by AI agents, and the company has appeared on Forbes' Fintech 50 list for five straight years, with revenue doubling from $41 million in 2024 to $86 million in 2025, according to Forbes. Valon employs roughly 320 people, was valued at $1.75 billion in 2024, and has raised $275 million in venture funding from backers including Andreessen Horowitz, WestCap, and 166 2nd, per Business Insider's reporting.
Before the crackdown, Valon employees had broad access to AI and, according to Wang, routinely told him the AI was almost always right. That confidence is part of what worried him. By doing the basic work themselves rather than outsourcing it to a chatbot, Wang said, new hires start to form an actual understanding of the job, and without that foundation, they can't tell when the machine has led them astray.
The 'Meat Proxy' Problem
The concern echoes a slice of tech-industry slang that's been spreading this month. German software developer Niklas Gruhn popularized the term “meat proxies” in an August 3 blog post to describe people who forward unchecked AI outputs without reviewing or understanding them, as reported by Futurism. Techies have since adopted the nickname as shorthand for exactly the behavior Wang says he was seeing among Valon's newest employees.
Research backs up the cost of that dynamic. A September 2025 survey by BetterUp and Stanford's Social Media Lab, covering 1,150 full-time U.S. desk workers, found that 40% had received AI-generated work from a colleague in the past month, and that dealing with each instance took nearly two hours on average, according to BetterUp. Separately, a January 2026 study published in Societies by SBS Swiss Business School researcher Michael Gerlich, surveying 666 participants, identified a negative correlation between frequent generative AI use and critical thinking, driven by what researchers call cognitive offloading.
Guardrails Don't Apply Everywhere
The restriction isn't universal inside Valon. Engineers are exempt from the new AI-use policy, though all code still has to go through peer review before release. Valon's finance and human-resources functions, meanwhile, have no comparable AI guardrails at all. Wang said Valon previously spent around $15 million to $20 million a year on annualized AI token costs, a figure he now projects will shrink to roughly $4 million to $5 million this year as usage narrows.
Not everyone inside the company pushed back, according to Wang: tenured employees who'd been stuck fixing junior staffers' AI-generated mistakes welcomed the change, and he said he hasn't encountered internal resistance to it. New recruits, unable to lean on AI as a crutch, are now seeking help from more experienced colleagues instead, and Wang said they're developing a deeper understanding of their work as a result. Still, he acknowledged some AI enthusiasts outside the company have told him the policy is the wrong approach, and he's open to being challenged on it. “If you have a much better idea here of how to make sure people learn,” Wang told Business Insider, “please tell me.”
A Company Reshaped Around Software
The policy also arrives just as Valon has remade its own business. The company completed its sale of roughly 810,000 loans in its mortgage-servicing portfolio to Carrington Mortgage Services on August 4, freeing Valon Technologies to focus entirely on developing and selling its ValonOS software platform, according to Business Wire. Valon originally became a licensed mortgage servicer in 2019 specifically to test its software in-house, and in 2021 it became the first fintech to win mortgage-servicing approval from every major government-sponsored enterprise and federal agency, including Fannie Mae, Freddie Mac, FHA, VA, USDA, and Ginnie Mae, per FinTech Futures.
That regulatory track record underscores why Wang's team has so little tolerance for AI hallucinations: mortgage servicing platforms handle trillions of dollars in consumer loans under close federal oversight, leaving little room for the kind of unverified errors that plague other software fields. A February 2026 study from University of Texas at San Antonio researchers, previously covered by Hoodline's report on AI code slop, found that nearly 20% of package dependencies suggested by AI code generators simply don't exist, forcing developers to review flawed or unsafe suggestions.
Part of a Wider Workforce Debate
Valon's approach lines up with a broader shift in how companies think about AI and junior staff. An August 2026 CNBC and SurveyMonkey poll of 1,686 U.S. workers found that 35% believe entry-level and junior employees should be barred from using AI tools entirely, while 42% support access only with explicit guidelines, according to CNBC. Some researchers have also pointed to a decline in entry-level tech hiring, with U.S. postings requiring three years of experience or less dropping from 43% in 2018 to 28% in 2024, according to a University of Warwick and Oxford study cited by the London School of Economics, though analysts continue to debate how much of that shift stems from AI adoption versus remote-work trends.
-4.webp?w=160&h=400&fit=crop&crop=edges)
