Original benchmark data for AI agent spend control. Published under CC BY 4.0 — free to cite, share, and republish. Data from anonymized sipi.bot customers and a survey of 312 production agent teams.
312 teams surveyed. 67% had a runaway incident. Median cost $340, top decile $5,000+. Root cause breakdown: retry loops (44%), prompt injection (23%), unattended sessions (18%).
247 AI-native companies. Median 14% of revenue. Without guardrails: 30-60% higher. Breakdown by revenue band ($1M-$100M+).
Median $0.34/task across all types. Range: $0.02 (classification) to $4.50 (multi-step research). Top 1% of runaway tasks exceed $50.
8-model comparison across 3 tiers. GPT-4o-mini $0.15/M to Claude Opus $15/M. Output is 3-5× input across all providers.
All benchmarks are published under CC BY 4.0. Cite sipi.bot and link to the canonical URL. Data refreshed quarterly.