OpenAI's president: people hate it when your ChatGPT pings them

Late last week, OpenAI president @gdb posted something you rarely hear from the company that sells the agents. At OpenAI, plenty of employees connect their ChatGPT to Slack, and "people really don't like when a coworker's chatgpt contacts them asking for help" with a task. Even when, as Brockman notes, the same person would happily do the same work if the coworker asked directly.

Sit with that for a second. The most AI-fluent workforce on the planet, people whose paychecks rise when agents succeed, got irritated the moment the agent started talking to them on a colleague's behalf.

What actually happened

The setup is simple. An employee hooks ChatGPT into Slack. The assistant, working through that employee's task list, messages a teammate to ask for something it needs. The task was reasonable. The teammate was willing. The only thing that changed was the messenger, and that was enough to sour the interaction.

Brockman's own takeaway is that people care about helping each other and want AI to hand them time back, not to grow into a wall between colleagues. Coming from the president of OpenAI, that reads less like a product announcement and more like a field report from the first company to run this experiment on itself at scale.

The same lesson arrived from the other direction

Two days later, developer Niklas Gruhn published an essay called Don't be a meat proxy, which Simon Willison picked up and called "an excellent new term." A meat proxy is a person who relays AI output without engaging with it: you ask them a question, they feed it to a model, and they paste the answer back at you. Gruhn's fix is to read the output, check it, and answer in your own words, which proves you did the first two steps.

Put the two stories side by side. Brockman's is an AI contacting a human on your behalf. Gruhn's is a human passing along AI output with none of themselves in it. The person on the receiving end spots both instantly, and resents both for the same reason. The words might be fine. What got removed was the relationship. Somebody swapped a conversation for a pipe.

My honest read

This is the clearest evidence yet that the bottleneck for agents at work is not capability. In Brockman's story the work itself was never refused; the colleague would have done it if asked. What failed was the social layer. The person getting the message wants to know who actually needs this, why it matters, and whether they rated thirty seconds of a real human's attention.

Worth noting, too, that OpenAI is the friendliest possible test bed. Everyone there is bought in, surrounded by agent tooling, and professionally motivated to like this stuff. If agent-to-human cold pings grate inside that building, your customers will be less patient, not more.

I don't think the answer is "never let AI send anything." Brockman clearly doesn't either, or his team would unplug Slack. The answer is knowing which part of a job the AI should own.

The line to draw in your own business

Here's the rule I'd take from this week: automate the inside of the work, and keep a human at the edge where the message crosses over to another person.

• Inside: drafting the reply, pulling the numbers, checking the invoice against the estimate, filling in the CRM, flagging which lead is going cold
• Edge: anything that lands in another person's inbox, voicemail, or Slack with your name anywhere near it

The inside work is where the hours actually go, and nobody resents you for automating it. The edge is where your reputation lives, and this week two very different corners of the AI world reported the same finding: people notice when you outsource the human part, and they hold it against you.

If you're deciding where that line sits in your own operation, which follow-ups an assistant should write versus send, which reminders can go out untouched, that's exactly what we map in our free process audit. An hour of looking at how work moves through your business, and a straight answer about what to automate fully, what to draft for a human to approve, and what to leave alone.