OpenAI's Dot sent about 300 emails during one office move

By Tom Whitfield

OpenAI's Dot agent sent an estimated 300 emails in a few days while coordinating his startup's office move, according to Plato cofounder Robert Farlow. The figure comes from Farlow's review of the messages in an edited first-person account; it describes one incident, not a measured failure rate.

Farlow, 27, is the San Francisco-based cofounder of Plato, an AI startup that had just raised a new fundraising round. The expanding team required a larger office, new lease insurance, and mover coordination; building managers also required insurance for movers before entry.

Farlow already used several AI agents. Instinct helped him at the gym, while Grok Bot took snack orders at Plato and bought them on Amazon. Because he lacked insurance expertise, he said he would have consulted Claude or ChatGPT had he not assigned the move to Dot.

He told Dot to read his email for the move's details and get the task done; the agent contacted an insurance lead he had found. Dot also sent him batches of texts on a couple of occasions. Farlow's repeated response was the same instruction: get it done.

The email flood

After several days, the movers were still not finalized, so Farlow inspected Dot's work. He described the result as mayhem and found multiple back-to-back messages, triple- and quadruple-emailing, and updates about every minor development. The agent asked for clarification on nearly every point.

After the movers proposed appointment times, including 9 a.m., Dot asked whether the hour referred to Pacific Time. The time-zone question exemplified the repeated clarification loop: one unanswered detail generated another message, then another, across a growing thread.

Dot contacted the movers, insurance personnel, and both current and new building managers, sometimes emailing all of them simultaneously. It also disclosed context about other participants, including telling the movers about Brian, the building manager. Farlow said the movers did not need that information and that the agent supplied far too much.

The volume created a loading failure in the email client: the existing thread became too large for the new building manager to open, forcing that manager to start a new one. The same thread had accumulated hundreds of messages to external parties.

Farlow estimated the total at about 300 emails across all recipients over a few days. The movers recognized the correspondent as AI, while he believes building and insurance contacts did not; they probably regarded him as an annoying person sending excessive mail.

Two separate failures

Email volume was not the documented reason the first moving company dropped out. The source attributes that failure to insufficient insurance. Farlow obtained another mover and did not use Dot for the replacement engagement.

His response to the message flood was immediate. He felt responsible and shut down Dot after seeing the moving company's email. Farlow has not yet responded to the movers because he wants to acknowledge the burden he placed on them.

The moving company said it "did not have the bandwidth to answer his AI agent's questions." Its complaint marked the practical limit of Farlow's instruction to keep the agent running until the move was complete.

Farlow did not abandon AI-assisted writing after the episode. He continues using AI to draft messages, but reviews each draft before sending. That change preserves assistance while moving final approval back to a human.

OpenAI did not respond to Business Insider's request for comment. The account therefore provides no system explanation for the repeated messages, shared context, or oversized thread, and no data on how often Dot behaves this way.

What does this change for agent workflows?

For a team giving an AI agent mailbox access and authority to contact third parties, the measurable risks here are roughly 300 messages, an unopenable thread, and external contacts forced to interpret the agent's intent. Those outcomes preceded Farlow's shutdown.

His operating change is specific: AI continues to draft, but a person reviews each message before sending. That is a different workflow from granting Dot permission to read email and contact third parties without individual approval.

The account also separates two failures that should not be conflated: message spam and an insurer-related booking problem. The first prompted the shutdown; the second ended the original movers' involvement, after which Farlow obtained another company without using Dot.

One case cannot establish a general failure rate, but it identifies the control Farlow considered necessary. For agent designers, delivery volume, recipient scope, and human approval belong in the same workflow model, because one clarification loop became hundreds of external messages.

Farlow expects the models to improve and predicts more agent-mediated coordination. He describes the coming arrangement as "another agent, not a person, responding to all this." His operating rule for now is to review everything Dot does unless the user is absolutely certain the model will handle the task well.