Valon’s AI onboarding rule: AI banned until new starters learn the job

---

By Esther Smith

Andrew Wang runs an AI company and has just taken AI away from most of his new hires. Wang is CEO and cofounder of Valon, a 320-person mortgage servicing software firm in New York that markets itself as AI-driven. New starters in finance, HR and operations no longer get the tools on day one. Neither do senior recruits.

Wang has inverted the standard AI onboarding sequence. Tool training usually comes in week one and job knowledge arrives later. At Valon it now runs the other way.

Any L&D leader has read a flawless graduate deck and then watched its author fail to defend a single number in it. That is the problem Wang is trying to solve.

Engineers are the exception, because peer code review already catches bad output before it ships. Everyone else waits until their manager judges that they understand the work well enough to spot AI getting it wrong.

“It’s a weird decision for a company that is at the frontier of AI usage,” Wang wrote.

Token bills AND WEAK WORK LED TO THE REVIEW

Valon’s AI spending was running at a reported $15m to $20m a year, according to Business Insider. Across 320 people that is roughly $50,000 a head. The restriction is expected to bring it down to $4m to $5m, closer to $14,000 a head. So this is an AI onboarding decision with a number attached, which is rare.

Cost alone would have been a thin argument. The quality problem was the stronger one. Wang describes new hires reaching for expensive models to handle simple tasks, then producing work that experienced colleagues had to correct.

That pattern has a name now. BetterUp Labs and Stanford Social Media Lab surveyed 1,150 US desk workers in September 2025. They found 40% had received “workslop” in the previous month, meaning AI output that looks finished and is not. Each instance took a colleague around two hours to sort out, which the researchers costed at $186 per employee per month.

The developer Niklas Gruhn put the same complaint more bluntly. “I can talk to Claude myself,” he wrote. “It’s going to be faster and I get to control the context. I don’t need a meat proxy in between.”

Novices gain most from AI, which is exactly the problem

Here Wang’s decision gets uncomfortable.

The best-known study on AI and productivity found the largest gains among the least experienced workers. Danielle Li at MIT Sloan, with Lindsey Raymond and Stanford’s Erik Brynjolfsson, tracked contact centre agents using a generative AI assistant. Output rose 14% on average. Among the least experienced agents it rose 35%. Among the most experienced it barely moved.

“Without access to an AI tool, less-experienced workers would slowly get better at their jobs,” Li said. “Now they can get better faster.”

So Valon is withdrawing the tool from the exact group the evidence says gains most. The defensible reading of Wang’s position is that faster output and durable competence are different purchases. An employer buys the second one in year one.

There is a second risk. An AI onboarding restriction does not reliably stop use, as 42% of employees already using AI in secret suggests. A rule people work around teaches them to hide the workaround.

Manager judgement is now the access control

AI onboarding at Valon now has a gatekeeper, and it is the line manager. They decide when someone knows the job well enough to supervise the tool.

That places serious weight on a population that is often unsupported. 58% of new managers receive no training for the role. A manager who has never been taught to assess competence is now rationing the company’s most expensive tool.

Wang is open about not having solved it. “If you have a much better idea here of how to make sure people learn, please tell me,” he wrote.

AI onboarding has to teach the job before the tool

The AI onboarding question is no longer whether new starters can operate the tools. It is whether they can tell when a tool is wrong. Those are separate assessments and most induction programmes only run the first one. Name, for each role, the handful of judgements a person must make unaided, and treat those as the release condition. KPMG rebuilt its early careers training on a similar logic. It sits close to the argument in our piece on AI literacy and judgement.

Wang’s evidence is simple and timeless. You cannot supervise work you have never done.