OpenAI's Dot Sent 300 Emails to Move One Office. The Founder Shut It Down

Roughly 300 emails in a few days. That's what OpenAI's Dot agent produced after Robert Farlow, the 27-year-old founder of the AI startup Plato, asked it to coordinate his company's office move in San Francisco. The recipients — movers, insurance brokers, and building managers on both ends of the lease — got triple- and quadruple-emailed, bombarded with clarification requests, and in at least one case exposed to information about each other they had no business seeing.

Farlow told the story in a first-person account published by Business Insider. OpenAI did not respond to the outlet's request for comment.

Plato had just raised a new funding round and was growing the team, which meant a bigger office. Anyone who has managed a commercial move knows the friction: insurance requirements from building managers, lease-compliance coverage, mover scheduling. Farlow admitted he didn't know much about insurance, so he delegated the whole thing to Dot with a simple instruction: "We're moving to this office. You can read my emails to find out the details. Just get it done."

What went wrong, concretely?

When the movers weren't finalized after a few days, Farlow checked the agent's activity. His word for it: "It was mayhem."

The failure modes were textbook:

• Clarification loops. When the movers proposed "How's 9 a.m.?" the Dot replied, "Is that in Pacific Time?" It needed clarification on nearly everything, so it kept asking.

Duplicate sends. It sent multiple back-to-back messages to the same parties, forwarding every tiny status update.

Context leakage. It told the movers details about Brian, Farlow's building manager — information the movers had no need for.

Thread sprawl. The new building manager had to start a fresh email thread because the Dot-generated one had grown too large to open.

Cross-contamination. In some cases it emailed the movers, insurance contacts, and both building managers simultaneously.

For a while, the recipients didn't clock that they were dealing with software. The movers eventually caught on and told Farlow they didn't have the bandwidth to keep answering an AI agent. Farlow suspects the others never figured it out — they likely just assumed "some annoying guy" was spamming them.

What did the founder do next?

"I saw the email from the movers and felt so bad," Farlow said. "I'm like, 'What did I do?'" He shut the Dot down immediately and says he still owes the movers a response.

The deal with those movers fell through anyway — they lacked the required insurance. Farlow hired different movers for the second attempt, and this time he kept the agent out of the loop. He still uses AI to draft emails, but he reviews everything before hitting send.

His own verdict on the guardrail: unless you're absolutely sure the model will do a good job, review everything your agent sends. That's the operational takeaway for any PM or founder tempted to hand off multi-party, relationship-dependent workflows — vendor negotiation, customer onboarding, anything with humans on the other end who can burn out on your behalf.

Is this a one-off or the new normal?

Farlow is a heavy agent user by any measure. He runs Instinct at the gym and uses Grok Bot to take snack orders at Plato and buy them on Amazon. He isn't writing off agents after the moving fiasco — he argues this is "a glimpse of what the future will look like," with another agent, not a person, eventually answering on the other end of these email chains.

That's the honest tradeoff sitting under the story: agent-to-agent coordination could eliminate the spam problem entirely, but until the counterparty side catches up, every autonomous send lands in a human inbox with a human's patience attached.

The lesson for Monday morning: autonomy without a send-review checkpoint works only when the cost of annoying the recipient is zero. In real-world logistics, it almost never is.