Every spring, American employers enter a lottery for the right to hire foreign professionals. For fiscal year 2024, they submitted 780,884 registrations for 85,000 places. For fiscal year 2027, the count fell to 211,600. A tempting story writes itself from those two numbers: AI coding tools arrived, demand for imported programmers collapsed, and technology quietly settled a fight Congress never could. That story is too neat, and the evidence says so. The truer version is more interesting and more consequential.
A technology program in all but name
Legally, H-1B covers “specialty occupations” requiring at least a bachelor’s degree. Practically, it is a software labor program. Computer-related occupations accounted for 65 percent of the 386,318 petitions USCIS approved in FY2023. One category, systems analysis and programming, accounted for 54 percent by itself. Custom computer programming services firms received a quarter of all approvals, and Indian nationals received 72 percent. None of this is new. Computer occupations made up 46 percent of initial approvals across 2000 to 2012 and nearly 70 percent by FY2017.
Anything that changes how software gets built will eventually show up in this program. The live question is whether it already has, and how we would know.
What the productivity evidence says
The headline claims about AI coding tools come mostly from laboratory tasks. One controlled GitHub Copilot study had developers build an HTTP server in JavaScript; the assisted group finished in 71 minutes, the control group in 161. That is a real result on a narrow problem. Production software is not a narrow problem.
Field experiments tell a more sober story. Randomized trials covering 4,867 developers at Microsoft, Accenture, and a Fortune 100 company found a 26 percent increase in completed tasks. A study of more than 5,000 customer support agents found AI assistance raised issues resolved per hour by 14 percent, with the least experienced agents improving by roughly 35 percent. Meanwhile, METR’s randomized study of experienced open-source developers working on mature codebases found early-2025 tools made them 19 percent slower, even as the developers believed they were faster.
Read together, these studies suggest an AI-assisted developer produces the output of roughly 1.1 to 1.3 unassisted developers in real workflows. That is nowhere near the “one engineer replaces three” rhetoric. It is still enough to matter. A team that once needed six people may need five, and nobody gets fired. The sixth job simply never gets posted.
The canaries are young
That quiet mechanism is exactly what payroll data now show. Stanford’s Digital Economy Lab, using ADP records, found that employment among workers aged 22 to 25 in the most AI-exposed occupations sits about 19 percent below where it would be had it tracked less-exposed fields. The adjustment came through lower hiring rather than layoffs. Software development and customer support were at the center of it.
Junior work is where coding assistants are strongest: boilerplate, first-pass tests, documentation, translating tight specifications into code. Federal projections reflect the split. BLS expects traditional computer programmer employment to shrink 6 percent through 2034 while software developer employment grows 15.8 percent, data scientists 33.5 percent, and information security analysts 28.5 percent. The market is forming a barbell, thin at the entry level and heavy with experienced specialists. Firms that stop hiring juniors save money today and hollow out the pipeline that produces tomorrow’s architects.
Why the lottery numbers prove less than they seem
Deming taught a generation of managers not to treat every movement in a chart as a signal. A process whose rules change midstream produces data that cannot be compared across the change, and the H-1B registration series changed its rules repeatedly. USCIS moved to beneficiary-centric selection for FY2025, ending the practice of registering the same worker many times; FY2024’s peak was inflated by duplicates. New fees, a $100,000 charge on certain new petitions, and wage-weighted selection further raised the cost of speculative filings. All of that happened alongside the unwinding of a pandemic hiring boom that had pushed Indeed’s software-development postings to more than 120 percent above their February 2020 level.
The counterevidence deserves equal weight. Initial-employment H-1B approvals actually rose in FY2024, to about 141,000 from 119,000 the year before. Software postings bottomed in May 2025 and had recovered meaningfully by August 2026, precisely while coding tools were growing more capable. FY2027 registrations, even after the fall, still exceeded the cap by a factor of two and a half. A clean substitution story would not look like this.
Where the pressure will land
The most exposed model is the IT services firm, whose economics historically ran on billable headcount multiplied by utilization and rate. AI breaks the link between revenue and bodies. Cognizant, Infosys, TCS, and HCL remain among the largest H-1B sponsors, yet the seven largest India-based firms together received only about 5 percent of initial approvals in FY2024. Product companies, manufacturers, universities, and hospitals use the program too, and several of those employers sponsor outside the cap entirely.
The likely near-term shift is qualitative rather than quantitative. Fewer routine IT placements, more advanced-degree specialists in machine learning, security, semiconductors, and research, and a growing share for cap-exempt universities and hospitals.
The elasticity question
Everything turns on one ratio. Technical employment roughly equals demand for technical output divided by output per worker. Should AI raise productivity 20 percent while cheaper software lifts demand 30 percent, employment grows. Should demand rise only 10 percent, it shrinks. History offers both outcomes, and software is unusual because it feeds every other industry. Cheaper code could spawn enough new products, firms, and integration work to absorb the gains. Nobody yet knows which force dominates across a full business cycle.
What we can honestly say
AI will not solve the H-1B debate. It is, however, changing the economic ground beneath it. The case for importing large volumes of routine programming labor is weakening at the margin, and the entry-level squeeze is real. The fight over scarce talent in AI research, chip design, medicine, and science will not go away, and may intensify. Technology may end the lottery for commodity IT roles long before Congress acts, while leaving the harder argument over specialists exactly where it has always been.