In 2025, Salesforce researchers put leading AI agents through realistic CRM work. They succeeded about 58% of the time on single-turn tasks and around 35% once the task became a multi-turn conversation. And a benchmark is still tidier than a real company. A working company is messy in specific ways. The same customer sits in the CRM twice, and the Slack escalation contradicts the ticket. The APIs throw rate limits and vendor error codes that most test setups skip. Asking a model to invent a…