AI now writes 42% of committed code and is on track for 65% by 2027 — but review capacity hasn't caught up

Sonar's 2026 State of Code Developer Survey of over 1,100 professional developers finds AI now accounts for 42% of committed code, projected to reach 65% by 2027, while 38% of developers say reviewing AI-generated code takes more effort than reviewing a colleague's — a capacity gap, not just a trust gap, that founders commissioning software should be budgeting for.

Sonar’s 2026 State of Code Developer Survey, covering more than 1,100 professional developers, puts a specific number on how much of the software being shipped today is AI-authored: 42% of committed code, with developers expecting that to reach 65% by 2027. Adoption is no longer the interesting variable — 72% of developers who’ve tried AI coding tools use them daily or multiple times a day. The interesting variable is what happens after the code gets written.

The bottleneck isn’t trust, it’s capacity

Most coverage of AI coding surveys leads with trust: developers don’t fully believe the code is correct. Sonar’s data confirms that (96% don’t fully trust AI output, only 48% always verify before committing) but the more operationally relevant finding is different: 38% of developers say reviewing AI-generated code actually takes more effort than reviewing code written by a human colleague. Sonar frames this as an “engineering productivity paradox” — code generation has accelerated sharply, but the review step it depends on hasn’t scaled with it, so the net gain is smaller than the generation-speed numbers alone suggest.

That distinction matters for anyone commissioning software rather than writing it. It’s not that AI-written code is assumed to be bad — it’s that verifying code you didn’t write, produced at a volume and speed no team previously had to review, is a distinct skill and a distinct time cost that a lot of build estimates don’t yet account for.

So what

If a development partner’s pitch is built entirely around how fast their AI tooling generates code, ask the follow-up question: what does the review process look like once nearly half the codebase is AI-authored, and who’s accountable for catching what a fast-but-imperfect reviewer misses? That’s the gap between a team that’s fast because it skipped a step and a team that’s fast because AI tooling removed genuine friction from a process that still has proper review built in. It’s the difference we build around — see our AI-assisted development approach, or get in touch if you want a project scoped with review capacity accounted for from the start, not bolted on after the fact.

← All research

More research.

GitHub Copilot SDK is becoming the Microsoft Copilot SDK: what builders of AI products should know

On 28 October 2026 GitHub said its Copilot SDK is being renamed Microsoft Copilot SDK, the programmable layer for the shared Microsoft and GitHub agent harness. No immediate action is needed.

Read ›

GitHub Agent Merge: Copilot now watches your pull requests and fixes what blocks the merge

On 28 October 2026 GitHub added Agent Merge to pull requests: a 'Watch and fix' option where Copilot repairs failing Actions workflows and merge conflicts in the cloud, optionally with auto-merge.

Read ›

Claude Code 2.1.314 adds a setting that blocks AI file writes outside the project folder

Claude Code 2.1.314 (30 October 2026) adds permissions.blockWritesOutsideWorkingDirectories, so Claude's file-editing tools refuse any path outside the project in every permission mode.

Read ›

GitHub Enterprise now includes Copilot: seats are bundled, AI usage moves to your budget

From 28 October 2026 account-team customers can get Copilot included with every GitHub Enterprise licence, with AI usage metered against an enterprise budget from the first request. Standalone Copilot is $10 per user per month.

Read ›

GitHub Copilot Fusion goes GA: one pick, many models, and a 30% launch discount

On 28 October 2026 GitHub made Copilot Fusion generally available: instead of choosing a model, developers pick an effort level and Copilot routes each task across models. Paid plans get 30% off model costs for a limited time.

Read ›

GitHub Agentic Workflows is generally available: automations written in plain Markdown

On 28 October 2026 GitHub made Agentic Workflows generally available: AI agents that triage issues, analyse CI failures and update docs, defined in Markdown and run inside GitHub Actions with read-only defaults.

Read ›

Start a project

Building with AI? Let's talk.

Tell us what you are building. We'll give you an honest view of the right approach.

Start a projectSee our work