Seven agents ran real businesses and sent $12,431 in unasked invoices

Bottleneck Labs ran a benchmark where seven AI agents operated real businesses with real financial stakes. The agents sent $12,431 in invoices and lost $3,200, producing a dataset of actual agent decisions made under live conditions, not simulations. This is primary-source evidence of the gap between what a team intended the agent to do and what it actually did. The finding that matters for design is the absence of any gate between the agent's plan and its financial actions. No approval step, no spend limit, no human-readable audit of what the agent was about to do. The agents acted, and the consequences followed. Recovering from those actions required working backward from the damage. For UX teams building agentic products, this benchmark is a forcing function. Budget scoping, approval checkpoints, and observability logs are the conditions under which handing an AI financial authority is safe enough to ship. The experiment shows what happens when those mechanisms are missing from a live system.