A few years ago I treated AI coding assistants like a slightly smarter autocomplete. Helpful, sometimes magical, but still a tool. That framing already feels dated. In 2026, AI is not assisting; it is authoring. The question is no longer whether your team uses it, but whether the code it produces survives the rest of your pipeline.
The latest data makes that shift concrete. Faros analyzed two years of telemetry from 22,000 developers and 4,000 teams and found that 60% of AI-generated code is now accepted into codebases, up from 20% in their prior dataset. 1 Eighty percent of teams exceed a 50% weekly active user threshold for AI tools. 1 The assistant became the author while most organizations were still calling it a copilot.
That jump in authorship shows up in throughput first. Task completion is up 34%, epics completed per developer are up 66%, and code-specific tasks have increased 210%. 1 LinearB’s 2026 benchmarks, drawn from 2.7 million pull requests and 83,000 developers, show the heaviest AI users merging code at 2.3 times their June 2025 rate by May 2026. 2
But throughput is a vanity metric if you only measure what ships.
The downstream bill
Faros found that bugs per developer are up 54%, monthly incidents are up 57.9%, and the incidents-to-PR ratio has more than tripled. 1 Median review time increased 5x, and 31% more PRs are merging without any review. 1 Code churn, the ratio of lines deleted to lines added for merged code, has risen 861% for teams with high AI adoption. 1
That is the acceleration whiplash. AI generates more code than the surrounding system can absorb. The bottleneck moves downstream, into review, testing, incident response, and maintenance. You do not feel it while you are writing the code; you feel it three weeks later when the ticket comes back, or at 2 a.m. when the alert fires.
DORA’s 2025 report argues that AI amplifies existing strengths and weaknesses, and that strong engineering foundations offer protection. 3 Faros takes the opposite view: mature organizations are experiencing the same deterioration as everyone else. 1 I think both can be true if you read them carefully. Good foundations help, but they help at the margin. If you flood any system with 2x or 3x the raw output, the handoff points break unless you redesign them for the new volume.
What the gap actually looks like
LinearB splits organizations into performance tiers. At the top 10%, 54% of pull requests involve AI coding assistance, 57% get an AI code review, 45% of merged code lines are written by AI, and fewer than 5% of PRs come from autonomous agents. 2
Those numbers matter because they show adoption without dependency. Elite teams use AI heavily, but they still treat humans as the owner of the work. Autonomous agents tell the same story from the other side. At the top 10% of organizations, 79% of agent-opened PRs merge within 30 days. That sounds decent until you compare it to the 92% merge rate for human-only PRs. 2 At the top 60% of organizations, agentic yield drops to 37%. 2
The problem is not agent capability. It is ownership. An agent can open a PR, but if no engineer feels responsible for landing it, the PR rots. That is why AI code review is currently the cleaner win. LinearB found that PRs with an AI review attached merge at the highest rate of any category, lifting PR yield by up to five percentage points over the baseline. 2
Cost is not the constraint
AI spend gets attention because it is new and easy to invoice. LinearB puts the 90th percentile token cost at $481 per developer per month, or $35.20 per coding day. 2 That is under 4% of a fully loaded developer’s cost. The top cohort’s daily cost rose nearly 75% in three months, but the median stayed roughly flat. 2
So cost is real, but it is not the bottleneck. The bottleneck is abandoned work. PR yield, the share of opened PRs that merge within 30 days, is the connection between spend and delivery. The top 10% of organizations convert 90% of opened PRs. The top 60% convert 81%. Agentic PRs at the top 60% convert 37%. 2 Every abandoned PR is tokens, salary, and attention with no return.
The teams closing the gap treat AI as a production contributor, not a magic typing machine. They measure leverage, predictability, efficiency, and developer experience together. They automate review routing and pull request sizing. They turn on AI code review before they scale agents. They coach with the data instead of ranking by it.
What I am changing
At Symbol and at Wawandco, we are past the phase of handing out licenses and hoping for the best. We are now asking three questions before any AI-assisted work starts:
- Who owns the outcome? Not the prompt, not the branch, the outcome.
- How does this get reviewed? Bigger PRs need smaller units, not faster approvals.
- What signal tells us this was worth it? Merge rate, yield, incidents, rework, not lines generated.
The 2026 data is a wake-up call. AI can double your output, or it can double your mess. The difference is whether you redesigned the handoffs for a system that produces code at machine speed.
References
[1] Faros AI, “AI Engineering Report 2026: The Acceleration Whiplash,” March 2026. https://pages.faros.ai/hubfs/AI_Engineering_Report_2026_The_Acceleration_Whiplash_Faros.pdf
[2] LinearB, “The AI engineering productivity gap: 2026 benchmarks,” 2026. https://linearb.io/resources/ai-engineering-productivity-gap
[3] DORA, “State of AI-assisted Software Development 2025,” 2025. https://dora.dev/dora-report-2025/