What happened

In an interview published by Business Insider on 29 August 2026, Valon CEO Andrew Wang described a policy requiring most new hires to learn their roles without AI before managers approved access.

The restriction had already been introduced; the interview was a new account of it, not the policy’s first day. Wang’s public explanation connected the decision to preserving the experience through which employees develop judgement.

He also projected annualised token spending could fall from roughly $15-20 million to $4-5 million. That was his forecast, not an independently verified saving. Engineers were treated differently because their code went through peer review.

Why it matters

A person needs enough knowledge to recognise when a plausible answer is wrong. That knowledge may include facts, but it also includes understanding how work moves through the organisation and where exceptions occur.

An onboarding programme should therefore identify what new employees must learn to verify. For a finance administrator, that might include tracing a discrepancy back to its source. For a service adviser, it might involve distinguishing a standard policy from an authorised exception.

I would avoid making access depend on a manager’s unexplained impression. Define a small set of observable tasks: find the authoritative record, identify an error, explain the correction and know when to ask for help. Give employees practice and feedback before assessing them.

The purpose is to develop capability, rather than turn AI access into a status symbol.

The bigger shift

AI changes the design of apprenticeship. If experienced staff can complete routine work quickly themselves, managers should deliberately preserve opportunities for newcomers to understand that work.

One option is a protected practice environment. Give a new hire a realistic case with safe sample data. Ask for an initial diagnosis, then permit an approved AI tool and compare the two approaches. Include a case in which the tool’s suggestion is incomplete.

Another option is a structured review with an experienced colleague. The newcomer explains the evidence and the reviewer asks what could invalidate the conclusion. This develops a habit of verification while keeping the employee connected to the team’s practical knowledge.

Different roles may need different paths. Some work has strong peer review already; other work leaves errors less visible. The training design should reflect that difference, with clear and equitable criteria.

Include the cost of mentoring in the plan. The guide to corporate AI training budgets helps frame learning as more than tool instruction.

My take

I would borrow the underlying question from Valon before copying its restriction: how will new employees learn to challenge an answer confidently?

A temporary limit may be appropriate in a particular setting. In another, supervised use and deliberate practice may work better. The organisation should test the approach against learning, quality and the burden placed on colleagues.

Keep cost control visible but separate. Lower usage can reduce a bill without proving that employees understand their jobs better. Assess both outcomes.

The strongest onboarding process gives a new colleague a route to increasing independence. AI should fit that route, with clear expectations and support. A manager’s responsibility is to help people acquire the judgement that makes faster work dependable.