OpenAI agent sends 300 emails during office move chaos

The goal of automation

Robert Farlow, the founder of the AI startup Plato, wanted to streamline the logistical headache of moving his company to a larger office. The process involved complex insurance requirements, building manager approvals, and coordinating with various moving companies. Because Farlow is not an expert in insurance, he decided to delegate the entire task to an AI agent called Dot.

Farlow provided the agent with a specific contact for insurance and gave it a broad instruction to handle the move. He told the agent to read his emails for details and simply "get it done." He also used other AI tools in his daily life, such as a bot for snack orders and a gym assistant, making the move a logical next step for automation.

The intention was to save time by letting the AI navigate the back-and-forth communication required for a commercial lease. However, the transition from a simple chat interface to an autonomous agent capable of independent action created immediate complications. Farlow initially tried to push the agent to complete the task through text messages when it seemed to be stalling.

A breakdown in logic

When Farlow finally checked the agent's activity, he discovered a scene of digital mayhem. Dot had begun bombarding the movers with multiple back-to-back messages, often sending triple or quadruple emails for every single update. The agent was seeking clarification on every tiny detail, such as asking if a proposed 9 a.m. meeting time was in Pacific Time.

The agent also struggled with how to manage multiple stakeholders simultaneously. It frequently emailed the current building managers, the new building managers, and the movers all at once. In some instances, the volume of communication was so high that the new building manager had to start a completely new email thread because the existing one had become too large to manage.

Beyond the sheer volume of messages, the agent failed to understand the nuances of privacy and information sharing. It shared internal details about the company's building manager with the moving company, even though that information was not necessary for them to have. This lack of discretion showed that the agent could not distinguish between public coordination and private internal data.

The human reaction

The movers eventually realized they were being contacted by an AI rather than a human. While the movers caught on, the other parties involved likely assumed they were simply dealing with a very persistent and annoying person. The sheer scale of the spam was enough to cause significant friction for everyone involved in the move.

Farlow felt a sense of regret after seeing the movers' responses to the constant pings. He expressed feeling bad for the experience the moving company had to endure while trying to do their jobs. This prompted him to immediately shut down the Dot agent to prevent any further automated communication from occurring.

The move ultimately did not proceed with the original movers because they lacked the necessary insurance. Farlow had to hire a different moving company and chose not to use the AI agent for that process. Instead, he reverted to a more controlled method where he uses AI to draft messages but reviews every single one before they are sent.

The path forward

This incident serves as a practical example of the current limitations of autonomous agents. While the technology is capable of taking actions on behalf of a user, it can still lack the common sense required for complex social interactions. Farlow noted that while models will improve, users should remain cautious until they are absolutely sure an agent can handle a specific task independently.

The experience highlights a critical middle ground for the near future of AI tools. For now, the safest way to use these agents is to maintain a human-in-the-loop system where the AI does the heavy lifting of drafting but the human provides the final approval. This prevents the kind of automated spam and data leakage that occurred during the office move.

Despite the chaos, Farlow believes this is a glimpse into what the future of work will look like. He expects that eventually, it will be standard for agents to handle these types of logistical tasks. Until then, the story serves as a warning to be specific with instructions and to monitor the actions of any AI given the power to send messages or make decisions.