Why AI Won't Fix Small Businesses' H-1B Problem

Picture a 40-person fintech startup, Series B, three H-1B engineers on staff. They open their email and learn the government just proposed a $103,000 fee per H-1B visa. Not per year. Per visa. The founder's first call is to their head of product: "We need to figure out what we can automate."

That call is happening at hundreds of companies right now. The thing nobody in those meetings is confronting directly is that they don't actually have the people in place to make it work.

The policy shock is twofold: the proposed H-1B fee that DHS itself estimates would paralyze 76% of small businesses that rely on the program. And the State Department is preparing to revoke business and tourism visas for up to 200,000 asylum seekers . It's what NBC News is calling the largest mass visa revocation in US history. The US has broadly halted immigrant visa applications while the machinery gets reworked. And in the meantime, the foreign visa pipeline stops completely.

For large tech companies, this is painful but survivable. Google and Microsoft will absorb the fee, pass some of it along through pricing, and keep hiring. The 200,000-visa revocation will be tied up in federal courts for years. The administration almost certainly knows that. But you don't need to win in court if the announcement alone freezes hiring decisions for 18 months.

Small businesses have no buffer. They'll have to make decisions now, under duress, with incomplete information. And the decision most of them are making is to buy AI tools to try to fill the gap.

The problem with that is that talent infrastructure was already hollowed out from the inside. It's been trending that way for nearly a decade. College recruitment programs got cut because they're "inefficient" compared to poaching experienced hires. Internships became resume theater. The mentorship structure that used to turn a smart 22-year-old into a capable engineer in 18 months disappeared because senior people are too stretched to run it, and nobody's measuring it. Then one day a company needs someone mid-level and there's nobody waiting in the wings. The $103,000 visa fee is almost beside the point. Sadly, this current situation is a pressure test applied to a system that was already rotten.

But it goes back even further, and the mistakes of the past are illustrative. GM spent over $600 million automating its Hamtramck assembly plant in the 1980s, then watched productivity crater as robots collided with each other, painted cars the wrong colors, and smashed windows while workers stood by, unable to intervene. All because the humans who understood the old process had been let go before the new one was stable. The plant ran at a fraction of capacity for years. Toyota's Georgetown, Kentucky facility survived that era by investing in process redesign and people at the same time they bought the equipment.

We are doing this again. Except the equipment is a Cursor license, and the workers being skipped aren't line workers, they're the senior engineers who would have known what to catch. And, honestly, at this point I don't have a solution. I don't know if the companies making this bet are wrong to make it. The fee is real. The timeline is real. Maybe you buy the tools now and figure out the infrastructure later, and it works out fine. Maybe the half-trained generalist is good enough for what most of these companies actually need. I'm not sure. What I've seen anecdotally is that the ones who don't build any institutional memory around AI failure end up with engineers who've absorbed just enough AI output to stop questioning it, and that's a bad place to be.

The quickest and most efficient response for a 40-person startup isn't necessarily a university partnership or a six-month bootcamp. It's a designated senior engineer whose job includes a weekly hour with every junior on the team, and a failure library (a Notion database, a structured post-mortem log, whatever your team will actually maintain) that records every time your AI tools produced something confidently wrong. The hallucination that made it into a client deliverable. The code suggestion that broke production three weeks after it passed review, on a Friday, when the person who approved it was already on a plane.

This new fee may not have created this rot, but it absolutely exposed it with chilling swiftness. Let's hope companies put in the work to fix their empty talent pipelines for the long term, even as they scramble for those quick short-term solutions.

Editor's note: The sourced reporting on this policy dates to August 25, 2026. All linked articles were published and available at time of writing.