AI coding adoption is nearly universal. AI code impact is not.

88.3% of developers now use AI coding tools regularly, up from just under 72% in early 2024. In the same dataset, AI-assisted pull requests merge at 32.7%, against 84.5% for unassisted code. Adoption has saturated; the work AI produces still gets through the delivery pipeline at less than half the rate of human-authored work.

Those two numbers, read together, are the state of AI coding adoption in 2026. This page, part of LinearB's research and data library, lays out the figures behind them and the method they came from. They are also the baseline for a question every engineering leader is being asked this year, whether the AI coding investment is paying off.

The figures come from LinearB's 2026 Engineering Benchmarks Report, which analyzed 8.1 million pull requests from roughly 4,800 engineering teams in 42 countries, paired for the first time with a qualitative survey of engineering leaders. Ben Lloyd Pearson, LinearB's Director of AI Innovation, led the analysis.

The report classifies every pull request into one of three groups:

• Unassisted: authored by a human with no AI involvement.
• AI-assisted: authored by a human but significantly shaped by AI, whether in code generation, planning or research.
• Agentic: created entirely by an AI agent, with no human leading the effort.

Size and timing figures are reported at the 75th percentile (P75) unless stated otherwise. The adoption percentage comes from the survey component; the PR-behavior figures come from the git data. Full definitions and distributions are in the report itself.

"AI adoption is effectively maximized at this point. It's almost universal."

— Ben Lloyd Pearson, Director of AI Innovation at LinearB, on Dev Interrupted, Why AI-assisted PRs merge at half the rate of human code

The report introduces a benchmark it calls acceptance rate, defined as the share of opened pull requests merged within 30 days. This is deliberately different from vendor acceptance rates, which count suggested lines a developer kept in the editor. A line accepted in the editor can still be deleted before commit; a merged PR reached the codebase.

"Of all the things that AI is generating for your organization, which of them are actually useful enough to be deployed to production? Like that's a far more meaningful metric to track."

— Ben Lloyd Pearson, Director of AI Innovation at LinearB, on Dev Interrupted, Why AI-assisted PRs merge at half the rate of human code

AI-assisted PRs run roughly two and a half times the size of unassisted ones at P75. LinearB's long-standing guidance is to keep PRs under about 300 lines, the size a reviewer can hold in their head; AI-assisted work lands above that line on average.

The two rows point in opposite directions, and that's the finding. AI-assisted PRs wait 4.6 times longer to be picked up (agentic PRs, 5.3 times), then get reviewed faster once someone starts. Survey respondents described the first half as hesitancy (uncertainty about the mental load and whether the code can be trusted) and the second half as reviewers scanning for high-level issues rather than reading deeply. One respondent reported more code reaching production without proper review.

AI is almost exclusively generating new code paths rather than modifying existing ones. The report's read is that this isn't the model being told to write new code; it's the model not being told to reuse what exists.

From the survey component, modeled on the DORA research's AI capability categories:

The adoption question is settled, and the impact question is open. Nearly nine in ten developers use AI; a third of what AI helps produce reaches the codebase within a month. The gap sits in review: larger PRs, unclear ownership, and reviewers who either wait or skim.

That reframes what an engineering leader should measure. Usage statistics confirm what the industry already knows. The numbers that separate organizations in this dataset are merge rate by PR type, pickup time, and refactor rate, because those are where AI's output either becomes delivery or doesn't. The adoption measurement guide in this library covers how to track them, and LinearB's AI developer productivity insights report them by PR type against these benchmarks.

"Don't confuse adoption with impact. Just because everyone is using AI today, it doesn't mean that you're getting real value from it yet."

— Ben Lloyd Pearson, Director of AI Innovation at LinearB, on Dev Interrupted, Why AI-assisted PRs merge at half the rate of human code

The dataset is drawn from organizations using LinearB, which skews toward teams that already measure delivery. PR classification depends on detecting AI involvement from git signals and tool integrations, so some AI-assisted work is likely classified as unassisted, which would make the real gap larger rather than smaller. Survey figures reflect engineering leaders' self-assessment. P75 figures describe the upper-middle of each distribution, not the median.

What percentage of developers use AI coding tools?

88.3% use them regularly, meaning multiple times a week, per LinearB's 2026 benchmarks. Other 2026 studies place the figure between 90 and 95%.

What is the merge rate for AI-generated code?

32.7% of AI-assisted pull requests merge within 30 days, against 84.5% for unassisted PRs, in a dataset of 8.1 million pull requests.

Why do AI-assisted PRs merge less often?

They're larger (400+ lines at P75 versus 157), they wait 4.6 times longer to be picked up for review, and reviewers report uncertainty about trusting and owning code they didn't write.