OpenAI's president admits people resent AI agents that insert themselves into workplace relationships

OpenAI's President Concedes the Problem With AI Agents Is Not the AI

At OpenAI, employees connect ChatGPT to Slack. When one person's ChatGPT messages a coworker to ask for help with a task, that coworker resents it. The telling detail: they would happily do the exact same work if the human colleague asked directly.

That observation comes from Greg Brockman, OpenAI's president and co-founder, whose remark was collected and published by software developer and AI commentator Simon Willison on his weblog 1. Brockman framed the reaction as evidence that people care deeply about human relationships at work and want AI to reduce their workload without inserting itself between them and the people they collaborate with.

The remark lands at an awkward moment for OpenAI. The company introduced workspace agents in ChatGPT in April 2026, describing them as shared, reusable agents that teams configure once and deploy across their connected tools 2. Through the ChatGPT Agents app on the Slack Marketplace, those agents can be tagged by name in channels, for example as @sales-helper, or configured to act automatically in response to incoming messages 3. In other words, the specific behavior Brockman describes coworkers disliking is one OpenAI's own product is designed to enable.

Here is why this matters.

The entire AI agent model assumes that workplace relationships are transactional pipelines that can be optimized by inserting software between people. A colleague asks for a favor. The agent routes the request. The recipient completes the task. Efficiency goes up. The flaw in that logic is that a request from a human carries social information the agent cannot transmit. When a colleague asks for help directly, they signal that they know what you do, that they value your contribution, and that they are spending their own credibility to make the ask. An agent relaying that request through Slack strips away every one of those signals. The recipient is no longer helping someone they know. They are servicing a ticket that software generated.

The shift from being personally asked to being automatically notified is where the friction lives. Brockman is describing a reaction rooted in how people build trust at work, not in how well the agent performs the task. No amount of product polish smooths over that difference.

Independent research points in the same direction. A May 2026 study in Harvard Business Review by researchers at Boston Consulting Group and Boston University found that when organizations place AI agents on their org charts as "employees," the framing produces unintended consequences. The authors questioned whether agents should be positioned as organizational members at all 4.

The technology keeps improving. The social structures it runs up against are not adapting at the same pace.

The AI industry is betting heavily on agents as the next computing platform. Every major lab is racing to build autonomous systems that operate across tools and take action on behalf of users. The pitch is throughput: fewer meetings, faster handoffs, less manual coordination. What Brockman's observation reveals is the cost that pitch does not account for. When an agent inserts itself between two colleagues, it can optimize the transaction while eroding the trust that made the transaction possible in the first place.

This suggests the real adoption ceiling for AI agents in the workplace may be social, not technical. Companies that figure out how to give agents enough social context to make their requests feel human, or how to keep them invisible to the colleagues on the receiving end, will clear a hurdle that more compute and better benchmarks cannot solve. The ones that do not will keep hitting the wall Brockman described. No amount of model improvement will make people enjoy being summoned by a coworker's chatbot.