Skip to content

tools.md

Latest commit

 

History

History
149 lines (102 loc) · 5.94 KB

tools.md

File metadata and controls

149 lines (102 loc) · 5.94 KB

Tools

Copilot Billing Forecast groups its tools into three categories. Upload a GitHub usage report CSV once, then switch between the tools below without your data ever leaving your browser.

All figures are estimates. AI Credit (AIC) costs use 1 AIC = $0.01 USD, and monthly projections run to the end of the calendar month containing the report's latest day, because GitHub AI Credit entitlements reset monthly. Always refer to your GitHub billing statements as the source of truth.

Usage Forecast

Forecast total AI Credit consumption and track it against your entitlement.

What it shows

  • Cumulative actual usage to date, with a projected trend over the remaining horizon.
  • A shaded 95% prediction band around the projection that widens the further out it goes.
  • An optional entitlement cap line, the projected exhaustion date, and any projected overage in USD.
  • A trend indicator (rising / falling / flat) and the fit's R² as a confidence cue.

How to read it

  • Enter your entitlement to see when, and by how much, you are projected to exceed it.
  • Use the projected daily run rate control to model a different rate from today onward. Set it as a percentage change relative to the fitted rate (the slider) or as an exact figure (the input) - the two stay in sync, so you only ever set one. When a scenario rate is active the projection becomes a flat line whose band shows expected day-to-day variability (rather than the trend's prediction interval), and a grey baseline line tracks the unadjusted trend for comparison.

Key assumptions

  • The projection is an ordinary least squares linear regression fit to daily usage. The most recent day is excluded from the fit when enough earlier days remain, since reports are usually pulled mid-day and the partial total would drag the trend down.

How the forecast range is calculated

The Usage Forecast fits an ordinary least squares linear regression to your daily AI Credit usage and projects that trend forward over the horizon. The shaded band around the projection is an approximate 95% prediction interval, calculated as follows:

  • The fitted line gives the central forecast value for each day.
  • The band half-width at a given day is t · SE · √(1 + 1/n + (x − x̄)² / Sₓₓ), where SE is the residual standard error of the fit, n is the number of days used, and the square-root term is the standard prediction-interval leverage factor (the band widens the further you project from the observed data).
  • t is an approximate 95% critical value that widens for small samples (≈4.3 for ≤2 days down to 1.96 for >30 days) to account for added uncertainty.
  • Both bounds are floored at zero since daily usage can't be negative, and the cumulative projected total is reported as a range (projectedLower to projectedUpper) by summing each day's bounds.

A couple of adjustments keep the run rate realistic:

  • The most recent day is excluded from the regression when at least three earlier days remain, because the report is typically pulled mid-day and its partial total would otherwise drag the trend down. It is still plotted as an actual.
  • The trend indicator (rising / falling / flat) compares the slope's effect over the horizon against the noise level, so small fluctuations read as flat.

The implementation lives in src/lib/forecast.ts.

Team Insights

Per-user metrics, models, and budget forecasts.

What it shows

  • A per-user table with total AI Credits, active days, run rate, and a projected month-end total, with search, sorting, and pagination.
  • A spend distribution across users and per-user month-end projections against an optional budget.
  • Usage cohorts (power, heavy, typical, light, and near-zero users) with the average and median spend per user in each.

How to read it

  • Set a budget to flag users projected to exceed it by month-end.
  • Use the cohorts to see how concentrated spend is among your heaviest users.

Key assumptions

  • Each user's run rate is their average AI Credits per active day, falling back to a simple mean when there are too few days to fit a regression.
  • Cohorts rank active users by AIC gross cost; near-zero users have less than $1 of AIC cost in the report.

Model Breakdown

AI Credit usage and trends broken down by model.

What it shows

  • A stacked area chart of daily AI Credit usage per model over time.
  • A per-model table with totals and trends, expandable for detail.

How to read it

  • Identify which models drive the most spend and whether their usage is rising.

Cost Center Rollup

Roll up AI Credit usage and budgets by cost center.

What it shows

  • A stacked area chart of daily AI Credit usage per cost center.
  • A per-cost-center table with each center's share of spend, user and model counts, run rate, and projected month-end total.

How to read it

  • Compare cost centers by their share of total spend and month-end trajectory.

Key assumptions

  • Rows without a cost center are grouped under (no cost center); share is each center's portion of total AI Credits in the report.

Spike Detection

Flag days with anomalous usage above the trend.

What it shows

  • An actual usage line plotted against an expected line and an expected range band.
  • Spike markers on days whose usage rises anomalously above the trend.
  • A table of flagged days with the top contributing users and models for each.

How to read it

  • Investigate flagged days to find the users or models behind an unexpected jump.

Key assumptions

  • The expected range is derived from the usage trend and its variability; days above the band are flagged as spikes.