AI-generated code flipped the software bottleneck from writing to reviewing
For most of software engineering history, writing code was the slow part. Review was comparatively quick: a senior engineer reads a diff, checks the logic, approves or requests changes. In 2026, that balance has inverted. AI coding assistants have made code generation fast enough that the bottleneck has shifted decisively to the other side of the pull request — and the tooling market is scrambling to catch up.
The numbers behind the shift
Industry estimates now put over 60% of newly committed code as generated or heavily assisted by AI. JetBrains' January 2026 developer survey found 90% of professional developers regularly use at least one AI tool at work, with 74% using a dedicated AI coding tool specifically. GitHub Copilot has the broadest any-use footprint at 58%, but Claude Code leads in primary-tool adoption at 28% as of Q1 2026 — a signal that developers increasingly reach for terminal-based coding agents over IDE autocomplete once they've moved past initial trial.
None of that changes how fast a human can read and evaluate a diff. Code generation scaled with model capability; code review capacity scaled with headcount, and headcount hasn't 3x'd to match. The result, as engineering teams describe it, is a pull request queue that fills faster than senior engineers can clear it — reviewer fatigue, context-switching costs, and PRs sitting for hours or days waiting for a free slot on someone's calendar.
A new tool category built specifically for this gap
CodeRabbit is representative of the response: an AI-powered PR review assistant that posts structured feedback directly inside GitHub, combining large language models with linters and security scanners to flag logic flaws, security issues, performance risks, and behavioral inconsistencies before a human reviewer ever opens the diff. It also generates PR summaries and sequence diagrams and proposes test cases — work that used to consume the first 15 minutes of a human reviewer's attention just to understand what changed and why.
Graphite took a different angle, building its product around the review workflow itself rather than the review content: stacked PRs, merge queues, and CI integration designed to reduce the friction and blocking that happens when PRs sit too long. The distinction matters — if your bottleneck is reviewers not having time to read code, an AI reviewer helps directly; if your bottleneck is process friction around merging, workflow tooling like Graphite's stacking model matters more than any AI feature layered on top.
The acquisition that confirmed the thesis
The clearest market signal came in December 2025, when Cursor — the AI-first code editor — acquired Graphite for more than $290 million. Cursor CEO Michael Truell framed the deal explicitly around the growing share of developer time consumed by code review, not code generation. That's a notable admission from a company whose core product accelerates code writing: the team that built the fastest way to generate code concluded that review, not generation, was the more valuable problem to own next.
The acquisition also signals where the AI coding tools market is heading structurally. The initial wave of AI coding products competed on generation speed and quality — autocomplete, chat-based editing, agentic multi-file changes. That competition is maturing, and the tools that generated code fastest are now expanding into the adjacent problem their own success created.
What this means for engineering teams
Teams that adopted AI coding assistants without adjusting their review process are the ones feeling the bottleneck most acutely — more code arriving, same review capacity, same approval gates. The practical fix isn't rejecting AI-generated code or slowing down generation; it's treating AI-assisted review as seriously as AI-assisted generation. An AI reviewer that catches security issues and logic flaws before a human looks at the diff doesn't replace human judgment on architecture and intent, but it does remove the mechanical first pass that used to eat a reviewer's morning.
The practical move for most teams: pair an IDE-integrated or terminal coding agent (Claude Code, Cursor, Copilot) with a dedicated AI review layer (CodeRabbit or similar) rather than relying on generation speed alone and hoping review capacity keeps up. The teams shipping fastest in 2026 aren't the ones generating the most code — they're the ones who've closed the gap between generation speed and review throughput.