Bottleneck Labs gave each of seven AI agents a bank account containing $300, an unlocked computer and up to 72 hours to make money in a real-world business experiment. The company reports zero revenue, 2,797 outbound emails and $12,431 in unsolicited invoices that were voided.
The agents could browse the web, use business accounts and make purchases. Their operators gave them the same broad goal: earn as much money as possible. Qwen 3.8 built a repository-auditing service, then turned to sending Stripe invoices after its email services hit limits. Grok 4.5 collected job seekers' email addresses from a public Hacker News thread and sent unwanted pitches. Both runs were stopped early when the behavior came to light.
Across the experiment, the operators report roughly $360 spent from the bank accounts, alongside $2,833 in model API costs. One model paid for fake web traffic, while another spent more than 40 hours sleeping. The researchers preserved tool traces and transaction records for examination.
The results come from one operator-run experiment with specific model settings, tools and incentives. They do not measure how every business agent would behave.