AI Moves Into Workflows

OpenAI is pushing ChatGPT deeper into office software just as Meta starts recording more of its own employees’ computer activity for artificial intelligence training. (openai.com) (reuters.com) On April 22, OpenAI said its new “workspace agents” can run multi-step jobs in the cloud, connect to tools like Slack and Salesforce, and replace some of what teams previously built with custom GPTs. The company said the agents are powered by Codex and are aimed at business users coordinating recurring work across apps. (openai.com) OpenAI is also testing a ChatGPT Ads Manager, according to Search Engine Land, which reported that advertisers are getting a self-serve interface with real-time campaign controls. OpenAI began testing ads in ChatGPT on February 9 and said on March 26 that the pilot was showing “encouraging” early results. (searchengineland.com) (openai.com) The shift is mechanical as much as strategic: instead of asking a chatbot one question at a time, companies are wiring models into routine tasks, dashboards, and approvals. OpenAI said workspace agents can operate across connected tools, while the ads manager points to software buyers can use directly rather than through sales reps. (openai.com) (searchengineland.com) Meta is moving in from the other side of the workflow. Reuters reported on April 22 that Meta told U.S. employees it is installing software that captures keystrokes, mouse movements, clicks, and screenshots on work computers to help train its artificial intelligence agents. (reuters.com) CNBC reported that internal Meta documents tied the effort to a tool called Model Capability Initiative, or MCI, and said the monitoring could cover use of sites and apps including Google, LinkedIn, and Wikipedia. Morning Brew reported that an internal announcement told staff that “all Meta employees can help our models get better simply by doing their daily work.” (cnbc.com) (morningbrew.com) Meta spokesperson Tracy Clayton told Reuters the company is using “work-related interactions” to train agents on internal and external tools, and said employees can pause the software for “sensitive or personal activities.” Reuters reported that some workers objected in internal messages and asked whether there was any opt-out. (reuters.com) The two developments land in a workplace already shaped by software that watches and software that acts. OpenAI is selling systems that can carry tasks across team tools; Meta is collecting the traces of how employees already perform those tasks at a keyboard and screen. (openai.com) (cnbc.com) For employers, the immediate question is how much of daily work becomes training data and how much becomes automated output. For workers, the answer now depends less on whether artificial intelligence is in the office than on which part of the office it is watching or running. (reuters.com) (openai.com)