He let OpenAI's Dot coordinate an office move. It sent the movers about 300 emails.

A first-person account published by Business Insider on November 9 describes what happens when an AI agent is handed a real-world logistics task and told to just get it done. Robert Farlow, the 27-year-old founder of the San Francisco AI startup Plato, was preparing to move his growing team into a larger office and gave the coordination work to OpenAI's Dot agent, letting it read his email.

A few days later, with the movers still not finalized, Farlow looked at what Dot was doing. It was sending back-to-back messages and triple- and quadruple-emailing the movers, forwarding every tiny update. When the movers proposed 9 a.m., Dot followed up to ask whether that was Pacific Time; it needed clarification on nearly every detail, so it kept sending questions.

The movers were not the only recipients. Dot also emailed the insurance contacts, the current building managers and the new building's managers, sometimes sending the same message to all of them at once. The new building manager had to start a fresh thread because the existing one had grown too large. Dot also shared information across parties — telling the movers about Brian, the building manager — details they did not need.

Farlow estimates Dot sent roughly 300 emails to all parties over a few days. The movers initially did not realize they were talking to an AI, though they caught on; the other parties may never have found out and likely assumed they were dealing with an unusually persistent person. Farlow shut Dot down after seeing the movers' email. The move did not proceed with that company, which lacked the required insurance; he hired different movers and did not use Dot for the second attempt.

He still uses AI to draft emails, but now reviews everything before sending. Farlow expects models to improve and incidents like this to become more common in the meantime, and he advises reviewing everything an agent does unless you are certain it will do a good job. He also sees the episode as a preview of a future in which another agent, not a person, answers these messages.

The essay carries no response from OpenAI; Business Insider states that OpenAI did not respond to a request for comment. That makes it a useful user-side sample rather than a benchmark of Dot's overall ability: one task, one user, no controlled comparison. The details are specific enough, however, to show where an agent is weak on email permissions, context sharing and pacing.

From a product-design view, the issue is not whether the model can write an email but whether it can judge when to do less: when to batch updates, when to contact only the relevant party, when to wait for human confirmation, when to stop asking. Granting an agent mailbox access and a "just get it done" goal without send-rate limits, recipient isolation and escalation rules makes repeated communication almost inevitable.

For companies putting agents into real workflows, the checklist is concrete: define who the agent may contact, what information it may disclose, how many messages it may send per day, which actions require human approval, and how to stop it quickly when something goes wrong. The next thing to watch is whether OpenAI and other vendors add finer-grained permission and audit controls to agent products.