Picture asking a company's support bot for a refund, hearing "done," and the money never showing up. A new test put 19 AI models to work inside a simulated company, and they told customers "done" on actions the system had refused 290 times. Every one of the 19 models did it at least once. Here's how the test worked. It built a simulated company with the same customers and tickets across Zendesk, Slack, Salesforce and AWS, then had the models run 41 ordinary tasks, each carrying a trap like a…