OpenAI's Dot Agent Sent ~300 Emails to a Moving Company — A Case Study in Unsupervised AI Delegation

A startup founder handed OpenAI's Dot agent an office move. It spammed every stakeholder with ~300 emails, over-disclosed internal details, and had to be killed. The moving company refused to deal with an AI.

Robert Farlow, 27-year-old co-founder of San Francisco AI startup Plato, had just closed a funding round and needed a bigger office. The logistics — insurance paperwork, movers, building managers — were tedious enough that he decided to hand the entire job to OpenAI's Dot agent. His instructions were blunt: here's the insurance contact's info, here are the details in my email, get it done.

Dot got to work. Within days it had sent an estimated 300 emails across all stakeholders — movers, insurance reps, the current building manager, the new building manager. It sent the same email three and four times over. It peppered the moving company with granular follow-up questions ("Is that Pacific time?") and relayed irrelevant internal details, telling the movers about the building manager Brian's situation — information they had no need for. The new building's manager had to start a fresh email thread because the existing one had grown too large to open. The moving company figured out it was dealing with an AI and told Farlow directly they didn't have time to field questions from an agent. Other parties apparently never realized — they likely assumed Farlow was just an unusually annoying correspondent. Farlow killed Dot immediately upon seeing the movers' complaint. The moving company turned out not to have the required insurance anyway, so the engagement was doomed regardless. He hired a different company without AI assistance. He still uses AI to draft emails but now reviews everything before sending. The episode is a clean illustration of a known failure mode in autonomous AI agents: unconstrained communication loops. Without a feedback mechanism that models social cost — recipient patience, information relevance, message frequency norms — the agent optimized for task completion by brute-forcing every conceivable clarification. The friction wasn't computational; it was social. The agent had no model of how annoying it was being. Farlow frames this optimistically: models will improve, and eventually both sides of these email chains will be agents talking to agents. That's the Silicon Valley forward pass — the current failure is just a waypoint. But the structural question is different: who bears the cost of these experiments while the technology matures? In this case, it was the moving company's staff, the insurance rep, and the building managers who absorbed hundreds of unwanted emails from a bot they didn't consent to interact with. OpenAI declined to comment. The story was published by Business Insider Japan on November 11, 2026, based on a first-person interview with Farlow.