GitHub's 2026 Octoverse report has a number every engineering lead should sit with: in September, more than 40% of pull requests on GitHub were started by an AI agent, up from 17% in March (Octoverse 2026). Writing code got cheap this year. Reviewing it didn't. That's why Copilot code review moved from a nice extra to a core part of the platform, and why GitHub gave it a new top tier at Universe. Here's how it works today, what it costs, and the rollout we use with clients.
What Copilot code review does
Copilot reviews a pull request the way a reviewer would: it reads the diff in context, leaves comments on specific lines, and proposes suggested changes that the author can apply with a click. You can request it like any other reviewer, set it to review automatically, and, through the REST and GraphQL APIs, request it from your own tooling. It's available on Copilot Pro, Pro+, Max, Business and Enterprise (GitHub docs).
Automatic review is configured with rulesets, at the level that suits you:
- Repository: a branch ruleset with Automatically request Copilot code review turned on.
- Organization: the same branch rule in an organization ruleset, targeting repositories by name pattern.
- Enterprise: an enterprise-level branch ruleset, so every organization inherits it.
Optional settings let Copilot review new pushes to an open pull request and review drafts. Developers also have a personal code review settings page on every plan to turn on automatic review for their own pull requests (GitHub changelog).
Lite, Balanced and Max
Every review runs at an effort level, and the level is now the main lever on both quality and cost.
- Lite is a standard review: quick feedback for everyday changes.
- Balanced is the default. It goes deeper on complex logic, security-sensitive code and changes that cross service boundaries.
- Max, announced at Universe in public preview, is the most intensive level. It runs several independent review passes using high-reasoning models and reconciles the results. GitHub says Max posted the highest F1 score on ReviewBench for high-severity issues (GitHub blog). It hasn't published the full benchmark results, so treat that as a vendor claim to test on your own code.
Admins can set a default effort for the enterprise, organization or repository, and cap the maximum effort at any of those levels so nobody runs Max on every typo fix.
What it costs
Code review is billed in AI credits, like the rest of Copilot's premium features. Balanced and Max use more credits than Lite. On private repositories, reviews also consume GitHub Actions minutes, and the deeper levels can use more of those too. Organization owners decide who pays: by default each review draws on the requesting member's allowance, and the review fails if that allowance is exhausted. Alternatively the organization can pay, with its own budget.
Our advice: bill reviews to the organization, put a budget on it, and look at a month of data before deciding which repositories deserve Balanced by default. If you're also rethinking how you pay for Copilot overall, our Copilot billing comparison covers seats versus the new included model.
Should Copilot approvals count?
Copilot can now submit approving reviews, not only comments. It's off by default, and it's two separate decisions: allow Copilot to approve pull requests, and allow those approvals to count toward merge requirements. Repositories can restrict counted approvals to up to 15 file path patterns, and the enterprise default is disabled everywhere.
Our recommendation is to turn on approvals as a signal but keep them out of merge requirements, except for narrow, low-risk paths you'd already trust automation with: documentation, generated files, dependency bumps that pass CI. Everything else should still need a human approval, ideally from CODEOWNERS. The goal is to spend human attention where it matters, not to remove it.
The rollout we use
- Start with one organization ruleset that requests automatic review on your default branches at Lite or Balanced. Leave drafts off at first so developers aren't flooded.
- Give Copilot your house rules. Copilot reads your repository's custom instructions file (
.github/copilot-instructions.md). Put the review standards you actually enforce in it: error handling, logging, security patterns, what not to touch. If you already keep conventions in AGENTS.md, point to them so you aren't maintaining two copies. - Cap effort, then raise it where risk is. Set a maximum at the organization level and request Max deliberately on the changes that deserve it: authentication, payments, data migrations, infrastructure.
- Keep the deterministic gates. Linting, tests and security scanning belong in CI, where they pass or fail every time. Copilot review is for what those tools can't see: intent, logic and design.
- Measure it. Track how many Copilot comments developers actually act on, time to first review, and defects that escape to production. If a repository's comments are mostly ignored, fix the instructions or lower the effort level before you pay for more.
Copilot or a dedicated review tool?
There's a healthy market of AI code review tools, and some teams run one alongside Copilot. For GitHub Enterprise customers, Copilot's advantage is that review lives inside the same rulesets, policies, audit logs and billing as everything else, and it can review agent-authored pull requests in the same flow as human ones. Whatever you choose, the principle is the same one behind every agentic SDLC that works: when agents write more of the code, the review system has to be engineered, not improvised.
Deop helps engineering teams roll out GitHub Copilot, including code review, approvals and governance, as the 2026 GitHub Growth Services and Channel Partner of the Year and a Microsoft partner specialized in Agentic DevOps. We'll be at GitHub Universe On the Road Toronto on November 18. Book a discovery.
