AI spend intelligence from a raw card feed

The problem

Finance teams see AI tools as truncated, inconsistent card descriptors spread across many cardholders.

Using a generated feed of 1,900 transactions from 600 employees over 6 months, this project answers three questions: which vendor a charge belongs to, whether a vendor is priced by seat or by usage, and when month-end spend can be called.

Headline numbers

• 231 distinct descriptors resolved to 15 vendors at 100% purity against ground truth; five vendors are recovered by trigram clustering alone.
• Seat, usage or hybrid pricing classified from the time series alone: 14 of 15 correct.
• At day 5 the 80% interval on month-end spend is plus or minus 100%; by day 12 it narrows to plus or minus 34%.
• Plain run-rate was the best point estimate; a weekday pacing model did worse at three of five cutoffs.

What it means

Metered AI spend does not behave like a seat subscription, so budget controls built for predictable recurring amounts stop working without failing loudly. The useful forecast is a calibrated interval, and it says metered spend cannot be called before roughly day 12.