BurntAI

How to forecast which developers will go over their AI budget

By Jiri Sifalda

To forecast which developers will go over their AI budget, project each person's month-end spend from their recent daily rate, then compare that projection to their budget. The provider tools do not do this for you. They alert or block when spend reaches a limit, which is after the money is already spent.

This guide shows what the built-in controls do, why they are not a forecast, and how to build a per-developer forecast from the data GitHub and Anthropic already give you.

What the built-in controls do

GitHub Copilot. Organization and enterprise budgets send alerts when usage reaches 75%, 90% or 100% of a budget, to the recipients you choose (setting up budgets). User-level budgets cap the AI credits one user can consume in a billing cycle, and they always enforce a hard stop: credit features are blocked until the next cycle, until an administrator raises the budget, or until an administrator approves a request for more (budgets for usage-based billing).

Claude Team and seat-based Enterprise. Owners and Primary Owners can set a monthly spend limit for each member and for the organization. When a member reaches their limit, their usage credits pause until the end of the month. Anthropic notes that it is possible to slightly exceed a limit. The admin view shows each member's month-to-date spend (manage usage credits).

Both are useful guardrails. Neither one is a forecast:

  • An alert at 75% fires on the amount spent, not on the pace. A developer at 60% on day 10 is on track to finish far over budget, and no alert fires yet.
  • A hard stop protects the budget but blocks the work. A developer who hits the cap on day 20 loses credit features for the rest of the month, often in the middle of a task.
  • Each control sees one provider. A developer can be under their Copilot budget and under their Claude limit, and still be over the total you planned for them.

A forecast gives you the days between "on pace to go over" and "went over". That is the time to have a conversation, move budget, or accept the spend on purpose.

The data you need per developer

You need three things for each person: month-to-date spend, a recent daily rate, and the date of the last complete day of data.

  • Copilot credits. The Copilot usage metrics API reports ai_credits_used per user (GitHub changelog). One AI credit is $0.01 USD (usage-based billing). Data for a day is available within two full UTC days after that day closes (About Copilot usage metrics).
  • Claude API and Claude Code cost. The cost endpoint of the Usage and Cost Analytics API, GET /v1/organizations/analytics/cost, returns cost in USD and can group by actor, with daily buckets. Data typically refreshes within 4 hours and can take up to 24 hours. Every response has a data_refreshed_at timestamp, and values can still change until about 7 days after the month ends (Usage and Cost Analytics API).
  • Seats. A seat fee is fixed for the month. It belongs in the total, but not in the daily rate.

The per-developer AI spend tracker guide shows how to join these sources into one row per developer.

A simple forecast that works

Use a run rate over recent complete days. It is easy to explain, and it reacts when someone's use changes.

  1. Find the last complete day for each source. For Copilot, that is usually two days ago. For Anthropic, use the last full UTC day before data_refreshed_at.
  2. Take the daily rate. Average each developer's usage spend over the last 7 complete days. Leave seat fees out.
  3. Count the days left. Count the days from the day after the last complete day to the end of the month. Days without data count as days left, not as days with zero use.
  4. Project the month end. Projected spend = seat fees + month-to-date usage + daily rate × days left.
  5. Compare with the budget. Flag each developer whose projection is above their budget, and show the projected overspend in dollars.
  6. Name the date. Divide the budget left by the daily rate to get the date the developer will cross the budget. That date is what makes the flag actionable.

A worked example with made-up numbers: a developer has a $150 budget, a $19 Copilot seat, and $70 of usage after 10 complete days in a 30-day month. Their last 7 days averaged $8 per day. The projection is $19 + $70 + $8 × 20 = $249, about $99 over. With $61 of budget left at $8 per day, they cross the budget in about 8 days. The forecast flags them today. A 75% alert on the same budget would fire in about 3 days, and leave about 5 days to act.

Where a run rate goes wrong

  • The first days of the month. With 2 or 3 days of data, one heavy day swings the projection. Show the forecast, but mark it as low confidence until a full week is in.
  • Weekends and leave. A 7-day window covers one weekend, which keeps the rate fair for most teams. A developer back from leave can look like a sudden spike. That is a reason to check, not to act.
  • Data delays. If you divide by calendar days instead of complete data days, every developer looks cheaper than they are. Step 1 exists to prevent this.
  • Late revisions. Anthropic values can change until about 7 days after the month ends. Treat the forecast as an early warning, and reconcile the final month against each invoice.
  • Caps change the outcome. A developer with a user-level budget on Copilot or a spend limit on Claude will stop at that cap. The forecast then tells you the date their work gets blocked, which is the useful number.

What to do with a flag

A flag is the start of a conversation, not a verdict. High spend can mean a developer runs an expensive model for a task a cheaper one would do. It can also mean they ship more than anyone else. Look at what the spend produced before you change a budget.

Then pick one action and do it before the crossing date: raise the budget on purpose, agree a change in how the developer works, or let a cap stop the spend.

Skip the spreadsheet

BurntAI is built to do this for you: every developer's Copilot and Claude spend in one row, checked against their budget, with a month-end forecast that names the overspender while you can still act. See pricing to get on the list.

See every dollar your team spends on AI

See every developer's GitHub Copilot and Claude spend in one dashboard. Track per-person AI costs, set budgets, and catch overspend before the invoice lands.