Hiring managers across the country have quietly rewritten job reqs as the fastest-spreading software tool in memory changed what one programmer can do alone, and the result shows up not first in paychecks or postings but in filled seats. A new Federal Reserve Board study traced a decisive break in trend that lined up with the mainstream debut of large language models, finding that programmer-heavy roles that once grew just under 5 percent annually slowed sharply and even stalled in programmer-dense corners of the economy. The headline: companies kept building software, but staffed it differently. By holding industry size constant and letting only programmer shares vary, the researchers inferred that firms began shrinking the proportion of dedicated coders on payroll. That inference did not rest on a single series; it linked headcount data with postings, wages, and sector detail, offering a narrative of reallocation rather than wholesale retrenchment.
After Late 2022: The Break in Trend
The study’s core method started simple and then got disciplined: if industries had kept the same programmer share they held before generative AI took off, what would employment look like once sector growth alone was applied? That counterfactual became a yardstick, and the actual path slid below it by roughly three percentage points per year since the second half of 2024. Translated into people, the gap implied on the order of half a million positions that likely would have existed by now under the prior hiring intensity. Crucially, the authors did not frame that figure as a pink-slip count. Some would-be entrants probably landed in adjacent analytics or product roles, while existing staff used tools like ChatGPT, GitHub Copilot, and Replit’s Ghostwriter to stretch farther, making a smaller bench cover more ground.
Before November 2022, programmer employment outpaced the broader labor market, buoyed by cloud migrations, digital sales channels, and the scramble to modernize back-office systems. After generative models proved reliable for code scaffolding, refactoring, and test generation—with IDE integrations from JetBrains and Microsoft streamlining use—employers revised the mix. The change did not arrive overnight. The divergence opened around mid-2024, roughly 18 months after the initial shock, matching a period in which teams hardened workflows, addressed security and license questions, and saw model quality jump with better context windows and retrieval patterns. This timeline fit a diffusion curve: pilot, learn, then adjust headcount plans during annual budget cycles when staffing targets are set and vendor contracts get renewed.
Where the Impact Lands: IT Services and Reallocation
The largest effects appeared not on Big Tech campuses but in the sprawling IT services and contract development market that builds and maintains systems for everyone else. This segment, which employs a sizable share of U.S. programmers, showed growth flattening as clients leaned into AI-enabled throughput. A regional hospital network that once hired a vendor team for an EHR customization now used internal analysts, Copilot-assisted engineers, and low-code extensions to ship smaller increments without adding headcount. A midsize manufacturer that relied on a contracting shop for QA offloaded basic unit-test creation to AI and cut external hours. These are unglamorous, repeatable tasks, the exact territory generative tools accelerated, and they disproportionately shaped demand for contractors who used to capture that work at scale.
Signals outside payroll reinforced the story. Programmer wages held steady, with no broad, systematic decline, suggesting the market valued experience that could orchestrate AI-assisted delivery. Meanwhile, job postings painted a two-step: a plunge through 2023 gave way to stabilization and a modest uptick since 2024 on Indeed and major ATS platforms. Titles morphed. Ads for “software engineer” ceded space to “platform engineer,” “ML ops engineer,” and “solutions architect,” roles that blend coding with infrastructure, data governance, or vendor management. In internal teams, product managers and data analysts absorbed more coding tasks—querying APIs, writing glue scripts, or building dashboards—while senior engineers handled code reviews, security, and integration complexity. The net effect looked less like collapse and more like a reshaped production line.
Timing, Causes, and What Comes Next
Causality claims always attract scrutiny, and the paper anticipated it. By comparing programmer trajectories with occupations plausibly less touched by generative AI over the same period, only the coding group exhibited a pronounced break from expected employment paths. As a sanity check, the method recovered known historical patterns: ATMs depressed the occupation-specific headcount of bank tellers even as banking expanded, while offshoring hit seamstresses through industry channels. The programmer result mirrored the occupation-specific profile. Still, the authors acknowledged confounders—higher interest rates, post-pandemic normalization, and a crypto unwind—that weighed on tech broadly. Their adjustment for industry size helped strip those out, leaving a residual that matched when GenAI became a production tool rather than a novelty.
Tax policy likely compounded the shift at the margins. The change that forced firms starting in 2022 to amortize research expenses, including software development, muted some appetite for headcount in R&D-adjacent units. Yet the effect varied. Where projects tied directly to revenue or compliance, the work proceeded; where initiatives were discretionary or speculative, teams stretched with fewer hires, relying on AI to prototype, generate tests, or migrate code between frameworks. Measurement noise also persisted. Discrepancies between household and establishment surveys since the pandemic muddied short-run signals, so the authors triangulated with postings, wages, and sector detail. Across different exposure frameworks, programmers stood out as a rare point of agreement: over 98 percent landed in the top tier of GenAI exposure, aligning with usage metrics showing coding prompts dominating interactions on leading models like Claude.
Practical Implications: Hiring Playbooks and Next Moves
This research suggested several concrete responses that have already been paying off for employers and workers. Companies that redesigned delivery around AI-assisted workflows—pairing Copilot-style tools with trunk-based development, robust code review gates, and security scanning in CI/CD—recovered cycle time without wholesale expansion. Vendor managers who rewrote statements of work to price by outcome rather than hours captured AI-driven efficiency while preserving quality. HR teams that rewired ladders to reward integration, prompt design, and systems thinking alongside language mastery retained senior engineers who coordinated more output per person. For individuals, the best moves involved stacking skills: instrumenting services, automating tests, and managing data contracts, not just writing features.
Looking ahead, the most durable strategies had emphasized measurement and optionality. Leaders who tracked programmer intensity as a ratio—engineers per dollar of software throughput—could set targets that floated with tool gains rather than legacy staffing norms. Buyers of contract development who ran pilots to benchmark vendors’ AI toolchains under their own security and compliance regimes secured leverage in renewals. Universities and bootcamps that folded code reasoning, data governance, and API composition into curricula placed graduates into roles resilient to task reshuffling. Taken together, these steps provided a roadmap: treat generative AI as an accelerator that shifted the production boundary, staff to orchestrate that shift, and keep hiring flexible until the next wave clarified whether cheaper code ultimately expanded demand or pushed the same work into fewer, broader jobs.
