Copilot budget requests GA: your dev stalls until you approve

Usage-based billing for AI coding tools has a failure mode nobody puts on the pricing page: a developer hits the credit ceiling mid-task and simply stops. GitHub made its fix generally available on September 16 — Copilot budget increase requests — and the mechanism is worth reading carefully, because it quietly makes someone on your team the person who unblocks engineers by approving spend.

What actually happened

Per the GitHub changelog, members who exhaust their Copilot AI credits can now request more budget at the moment they hit the limit. Before this, hitting the limit just blocked them from every Copilot feature that consumes credits — no in-product path forward.

The flow: organization owners, enterprise owners, or billing managers see pending requests under Requests from members, set a new amount per request, and click Approve and increase. Approval restores the member's access immediately. Requests route automatically to whichever account pays — an org-owned budget surfaces in org settings, an enterprise-owned budget in enterprise settings.

The scope line matters: this is Copilot Business and Copilot Enterprise on usage-based billing. If you are on an individual plan or a seat-only arrangement, this does not apply to you yet.

Why it matters for your business

This is a small feature that tells you exactly where the AI tooling market has landed. Per-seat pricing was predictable and you could ignore it. Usage-based pricing is cheaper at the median and unpredictable at the tail, and the tail is where the work happens — the long agent run, the big refactor, the incident at 11pm. GitHub is not fixing the unpredictability. It is giving you a queue.

So decide now who owns that queue, because the default is that nobody does and a developer sits idle waiting for an approval that is buried in a settings page they cannot see. Three things to put in place:

Name an approver and a backup. One person with billing rights, one alternate, both with notifications on. An approval workflow with a single un-paged owner is an outage with extra steps.

Set a standing rule, not a case-by-case judgment. Something like: approve any request under $X without discussion, anything above gets a two-minute conversation. The whole value of the feature is speed. If each request becomes a debate, you have converted a billing control into a delay.

Watch the pattern, not the individual request. Repeated increases from the same person are not a spending problem. They usually mean the work got routed to an expensive model, or an agent loop is burning credits on retries. That is a configuration finding, and GitHub's auto model selection cost/quality controls — shipped two days earlier — are the lever for it.

The broader point applies to every metered AI vendor you use, not just Copilot: if a limit can stop work, the escalation path for that limit is part of your operations, and it needs an owner before it needs a policy.

Sources: GitHub Changelog, GitHub Changelog — Copilot auto model selection.