Copilot Code Review Comes to Azure Repos With Per-Review Billing and a Two-Day-Late Report

Microsoft has rolled out GitHub Copilot code review for Azure Repos to all Azure DevOps customers, removing the signup that had limited access since June, according to InfoQ. The feature posts review comments on pull requests, bills per review through the linked Azure subscription, and reports the spend two days after it happens.

For teams building software on DevOps teams in Brazil, this ends a long wait: the AI-powered automatic review that already existed on GitHub now runs without requiring a repository migration. But the detail that decides whether it's worth turning on lies in the billing and the limits, not in the announcement.

Why Microsoft brought the feature to Azure Repos

The June announcement, signed by Dan Hellem and Andrew Brenner, was unusually candid about the strategy. The company had spent years pushing customers to migrate repositories from Azure Repos to GitHub in pursuit of AI experiences and agentic development, and then admitted the outcome:

While many customers are actively planning and executing migrations to GitHub, many others are still not ready to move and continue to rely on Azure Repos for day-to-day development.

>

-- Dan Hellem and Andrew Brenner, Microsoft

The complexity of migrating varies with organization size, customizations, compliance requirements, tooling, and industry restrictions. Instead of waiting for these teams to move everything, Microsoft brought GitHub's capability to where they already are. A reader of the source summed up the sentiment well:

Features like this have already been available on GitHub for a long time. For many companies it's too much work and risk to simply migrate to GitHub. I'm glad they're finally coming to DevOps.

>

-- Marc Selman, commenter

The move isn't isolated. The Azure DevOps Remote MCP Server reached general availability this month, giving AI assistants a hosted endpoint for work items, pull requests, repositories, and pipelines, even though Claude, ChatGPT, and Cursor can't connect because Entra still lacks the client registration mechanisms these clients require. Both releases serve the same population: teams that want agentic tooling in Azure DevOps without moving repositories.

What the August release adds

The update brings onboarding controls at the organization, project, and repository level, support for Managed DevOps Pools, custom instructions, and automatic reviews triggered by branch policies. Custom instructions can apply to the whole organization or project, be defined per repository, or targeted at specific paths.

The billing deserves a close read

This is where the Brazilian team needs to do the math before enabling it at scale. Each completed review consumes input, output, and cache tokens, converted into GitHub AI credits at a rate of one credit per cent. Higher review effort analyzes more context and consumes more; PR size, custom instructions, and model changes affect the total.

Charges appear in Azure Cost Management under the GitHub Copilot for AzDO product, 48 hours after the review finishes. Cost attribution has improved: charges now carry Azure DevOps project tags, so you can filter or group cost by project instead of seeing just an organization-wide total. Since these are standard Azure tags, they work in analysis views, exports, and budgets.

But the documentation is explicit about the limits of budgets:

Budgets only notify. They don't stop reviews or change any resource.

>

-- Microsoft documentation

In practice, this creates a risky scenario: a team can set a limit and receive an email within an hour of the next evaluation, but combined with the 48-hour delay in reporting, an automatic review policy applied to a busy project would run for two days before the cost shows up anywhere.

Concurrency and infrastructure limits

Before turning on automatic review broadly, it's worth checking the concurrency limits:

| Scope | Concurrent review limit | |---|---| | Per organization | 5 | | Per user | 2 | | Per pull request | 1 |

An organization with several teams reviewing every new PR will queue up. The feature runs on top of Azure Pipelines infrastructure, using the organization's default agent pool by default, which some organizations disable. Managed DevOps Pools are supported and must run the latest Ubuntu Server image; Windows images are not supported, and Microsoft advises against mixing Ubuntu and Windows agents in the same pool. Self-hosted agents are not supported.

It's advisory review, not a blocking gate

By design, Copilot always leaves a Comment-type review, never approving or requesting changes. In other words: its feedback doesn't satisfy mandatory-reviewer policies and doesn't block the merge. It also doesn't re-review after new commits unless asked to, and it doesn't read replies or follow up. For anyone expecting to use AI as an automatic quality gate, expectations need recalibrating: it's a second pair of eyes, not a gate.

Reliability and restrictions still open

Reliability during the preview was uneven. One customer reported that reviews would stall after the plan-tool execution phase, sit idle for about 60 minutes, and get auto-cancelled; Hellem confirmed the behavior as a known issue, with a fix taking about a week to reach all organizations. Another couldn't find the automatic review setting where the documentation pointed, and Hellem clarified that it lives at the branch policy level.

Preview terminology also remains inconsistent: the June post referred to a limited public preview running a signup called technical preview; the August one calls it public preview with no signup; and the documentation still carries a limited preview banner, with no SLA and limited support.

There are restrictions limiting where the feature applies:

• Repositories of 10 GB or less
• Pull requests with 100 changed files or fewer and no merge conflicts
• TFVC is not supported
• Gradual rollout by region, with an estimate of two to three weeks or more to reach all organizations

What to do before enabling it

Microsoft's own guidance is conservative: enable the feature for one or two repositories, compare effort levels, and monitor daily usage before expanding. The path that makes sense for a Brazilian DevOps team is the same: start with a low-volume repository, track cost tags by project in Cost Management, and only then evaluate automatic review via branch policy, knowing that the actual cost will only show up 48 hours later. Since the cost of each review scales with PR volume, turning on automatic review in a busy monorepo without this monitoring is a recipe for a billing surprise.