The useful assistant files the report before you ask for it. OpenAI says Holiday Extras saw that the day after an app launch, when VP of AI Innovation Liam Thompson got performance data he never requested. I read that as the real test for agents at work.
Thompson calls his dot Beast.
Beast noticed the launch, pulled data from Google BigQuery via a ChatGPT plugin and delivered findings unprompted, per OpenAI. Thompson leads AI use at the UK travel firm, and he said, "I was really happy when Beast said, 'Here you go,' and I didn't even have to ask him." That one incident carries the case for proactivity over chat.
David Lee, Director of Innovation and Growth, offers the second proof with his dot Brian. Lee calls Brian on his commute to preview meetings and order reports from company data, per OpenAI. He said, "By the time I get into the office and sit at my desk, the report is there and it's ready to go." Brian also flagged retiring API keys and traced affected people through email, which Lee credits as thinking ahead of him.
This sits inside a larger OpenAI pitch. Enterprise is sold as models and agents including ChatGPT Work and Codex to turn goals into finished work, per OpenAI's business page. Holiday Extras maps it plainly with Enterprise as foundation, Codex for building software, and dots for the work between. Voice is what makes Beast stick for Thompson, who says he spends at least two hours a day talking with it for briefings and recall.
The fair pushback is access. Thompson says he tried OpenClaw and Hermes but held back data, while he gave Beast his Slack, inbox, Drive and laptop because he trusts OpenAI in a company setting. His design win is striking but soft, with days of animation work down to about ten minutes by his own estimate using After Effects and Rive access. Without that level of permission, there is no unprompted report.
That caveat does not kill the claim. It clarifies it. Give one assistant real access and one painful recurring job, then judge it on what it does before you ask.





