AI spend, governed like a budget line — 540,431 LLM calls across 4 teams, priced at published model rates. One page, one story: where the money went, and the one week it went wrong.
The story: spend is calm for three months — then Engineering runs an agent experiment on the most expensive model and burns
$2,384 above its Apr–Jun monthly average (July total: $3,219) before anyone notices. Press ▶ Play and watch it happen.
Total AI spend · 6 mo
$0
Across 540,431 calls · Apr–Sep 2026
Spend so far (playback)
$0
Cumulative to Apr 1, 2026
July spike · excess cost
$0
Engineering hit 268% of its $1,200 cap in July
Avoidable premium spend
$0
10,500 short calls ran on the top-tier model
Daily AI cost — the calm, then the spike
Total daily spend, all teams. The shaded band is the flagged window: 11 days (Jul 10–24) where Engineering ran at 5.9–6.7× its trailing 28-day median. Drag the timeline or press play.
Daily cost Spike window · Jul 10–24 Peak day · Jul 16 — $285 total
Apr 1, 2026September 2026
Budget burn by team
September 2026 — monthly spend vs cap. Follows the timeline above.
Cost per ticket resolved · Sep 2026
The headline KPI: each team's monthly AI spend divided by the tickets it resolved.
Support resolves a ticket for $0.028 — Engineering costs 22× more per ticket ($0.624). The gap is model mix and context size, not waste — but now it's visible and priced.
Where the money goes · model mix
Share of the $12,673 total spend, by model (all teams, 6 months).
One model — the most expensive tier — eats 57¢ of every AI dollar, while doing plenty of short, routine work.
Inside the spike · Engineering, July 2026
Daily cost on the 11 flagged days vs a trailing median of ≈ $37/day. Detector rule: daily cost > 3× trailing 28-day median.
An agent experiment at 3.5× normal volume on the priciest model. The fix isn't “stop experimenting” — it's a per-experiment budget key and this alert running daily, not at month-end.
FINDING 01
One experiment cost $2,384 before anyone noticed
Engineering's July spend hit $3,219 — 268% of cap and ~3.7× a normal month. Eleven flagged days, all one experiment. A daily spike alert would have caught it on day one.
FINDING 02
The premium model is doing routine work: ~$812 avoidable
10,500 calls under 3,000 input tokens ran on the top-tier model at $1,015; priced at mid-tier they'd cost $203. Routing small calls down a tier saves ~$135/month, no workflow change.
FINDING 03
Support will breach its cap next month — from growth
Support closed September at 96.3% of its $340 cap with cost per ticket flat at $0.028. That's healthy adoption hitting a static budget: review the cap, don't freeze the spend. Failed/retried calls also burned $382 over 6 months.