Inside Big Tech: Developer AI Usage & Engineering Output
Verified disclosures from 53 companies, from Google, Microsoft, Amazon and Meta to HubSpot, Airbnb, Uber, IBM, Bank of America and Baidu, on how internal AI adoption translated into code share, pull request throughput, modernization velocity, and workforce capacity. Every figure was checked word-for-word against the company's own words; claims found only in news reports were left out.
AI is writing a fast-growing share of code. Google's AI-generated share of new code rose from >25% (Oct 2024) to 75% (Apr 2026), and Zeta Global reports 89.6% (Q2 2026). But code share is not shipping output: almost no company ties these numbers to cycle time or delivery throughput. Our reading of the evidence is that the human constraint has shifted to code review, automated testing, and architectural validation.
Disclosed AI Share of Code, Over Time
Every point is a figure the company itself stated, plotted on the date it applied. Output effects use different units (PRs, developer-years, minutes), so they are in the Productivity Evidence chart below, not on this axis.
Plotted: the four companies with three or more points, or two points on a consistent definition. DocuSign (60% → ~75%), Snap (~40% → 65%+) and Pinterest (~15% of the codebase → ~50% of new code) also disclosed twice; their latest figures are in the Disclosure Register. Definitions differ: Zeta and Compass report share of new code; Google mixes measures (new code; code "involving" accepted AI suggestions in Apr 2025; code "written by agents" per the CFO in Feb 2026); Cognizant switched from AI-generated (2025) to AI-assisted (2026). Google's 50% "last fall" point is dated approximately. The latest figure from all 39 companies that disclosed one is ranked in the Disclosure Register.
Where Productivity Is Actually Measured
The chart above, and the Disclosure Register below, measure adoption: how much code AI writes. This one collects every verified figure that measures what AI does to output, grouped from the strongest evidence (what ships, and how fast) to the weakest (a "productivity" number the company never defines). Each quote was re-checked against its source on 13 Oct 2026.
Bar length = size of the stated improvement, on one shared 0–200% scale. Measures differ between rows (features, PRs, commits, effort), so compare within a group, not across groups. "3x" is plotted as +200% and "2x" as +100%; cuts in time or effort are drawn at the stated percentage cut.
Not plotted: Stripe's >1,000 merged agent-written PRs per week (a count with no before-and-after baseline), and figures that measure inputs rather than output (Infosys's 28M AI-generated lines across client work, FactSet's coding-token growth).
When Companies Replaced People with AI
Companies rarely put AI setbacks in writing. These are the cases where a company moved work done by people to AI, or where an AI agent doing a person's work failed, and the company's own documents (or a tribunal ruling) record what happened next.
What independent research found
Checked but not included, because no company document or ruling could be verified: Commonwealth Bank reinstating 45 roles after an AI voice bot (media and union only), McDonald's ending its AI drive-thru test (statement to press only), Taco Bell keeping humans "in the loop" (press interview only), Cursor's support bot inventing a policy (the company's reply was a since-deleted Reddit post), IBM tripling entry-level hiring (press only), the Klarna CEO's "lower quality" remark (press interview), and AWS outages attributed to an AI agent (Amazon's own statement says AI was not the cause).
What Companies Have Actually Said
The latest AI share of code (or of pull requests) that each company has stated itself, in an earnings call, SEC filing, investor document, company blog or CEO post. Each quote was re-checked word-for-word against its source on 13 Oct 2026. Claims found only in news reports, podcasts or vendor pages were removed.
Definitions differ (new vs committed vs all code; AI-written vs AI-assisted), so the ranking is indicative, not like-for-like. Ranges are plotted at their midpoint; "over X%", "over half" and "majority" at the stated lower bound.