Copilot Code Review Hits Azure Repos: Per-Review Billing, 48-Hour Lag

Microsoft removed the sign-up gate on GitHub Copilot Code Review for Azure Repos on August 4, 2026, making it available to all Azure DevOps customers as a public preview. The move is a direct concession to a migration reality Microsoft stated plainly in its June announcement: many enterprise customers are not ready to move their repositories from Azure Repos to GitHub, and waiting for them to migrate before delivering agentic developer tooling is no longer tenable. Engineers already running GitHub Copilot in the desktop app will find the feature familiar in function but meaningfully different in billing mechanics and infrastructure constraints.

What the August Release Adds

The August update layered several capabilities onto the June limited preview. Onboarding controls now operate at organization, project, and repository granularity. Custom instructions follow the same three-tier hierarchy — organization- or project-wide, per repository, or scoped to specific file paths — letting teams encode style guides, security conventions, or language-specific rules without modifying review configuration per pull request.

Automatic reviews are the highest-leverage addition: a branch policy can trigger Copilot the moment a pull request opens, removing the manual invocation step. The feature runs on Azure Pipelines infrastructure using the organization's default agent pool. Managed DevOps Pools are supported provided they run the latest Ubuntu Server image; Windows images are explicitly unsupported, and Microsoft advises against mixing Ubuntu and Windows agents within a single pool. Self-hosted agents are also unsupported.

Concurrency is capped at five concurrent reviews per organization, two per user, and one per pull request. A project with ten simultaneous pull requests will saturate the organization limit immediately.

Per-Review Billing and the 48-Hour Delay

Each completed review consumes input, output, and cached tokens, converted into GitHub AI credits at one credit per cent. Pull request size, custom instructions, and model configuration changes all affect the total. Charges surface in Azure Cost Management under the product label GitHub Copilot for AzDO, but they appear 48 hours after the review completes — not in real time.

This reporting lag has a direct operational consequence: an automatic review policy applied to a high-volume project will accumulate two days of spend before any signal appears in cost analysis. Microsoft's own guidance is to enable the feature for one or two repositories first and monitor daily usage before widening scope.

The August release improved attribution by attaching Azure DevOps project tags to charges, enabling filtering and grouping across standard Azure cost tools. The documentation is explicit that budgets issue notifications only — they do not halt reviews or modify any resources. A budget threshold alert fires within an hour of the next evaluation cycle, meaning cost overrun notification and actual spend diverge by up to 48 hours plus one evaluation interval.

Operational Constraints

Copilot always posts a Comment review type — it never approves or requests changes. Its output cannot satisfy a required-reviewer branch policy and will not block a merge regardless of what it flags. It does not re-review after new commits unless explicitly invoked again, and it does not read reply threads or follow up on its own comments.

Preview reliability has been unsteady. One customer reported that reviews stalled after the plan-tool execution phase and sat idle for approximately 60 minutes before being automatically canceled. Product manager Dan Hellem confirmed the behavior as a known issue with a fix in rollout expected to take about a week to reach all organizations. A separate customer could not locate the automatic review setting where the documentation described it; Hellem clarified that the setting lives at branch policy level, not in the Copilot configuration panel.

The Broader Pattern

This release, alongside the Azure DevOps Remote MCP Server reaching general availability this month, forms a coherent pattern: Microsoft is shipping agentic capabilities into Azure DevOps rather than using feature gaps to push repository migration. The MCP server carries its own unresolved constraint — Claude, ChatGPT, and Cursor cannot connect because Entra lacks the required client registration mechanisms — illustrating how governance infrastructure often lags capability delivery, a dynamic explored in infrastructure governance as the real unlock for safe agent deployment. For teams operating under enterprise compliance requirements, the billing attribution improvements and policy-level controls matter more than review quality; a tool that cannot be cost-attributed or policy-governed does not clear procurement regardless of feedback accuracy.