What would happen if, instead of giving an AI agent individual tasks, you literally added it to work chats as just another colleague? Google Research and Google DeepMind tested this in practice — and it turned out that the agent sometimes coped better with technical tasks than with work communication
What would happen if, instead of giving an AI agent individual tasks, you literally added it to work chats as just another colleague? Google Research and Google DeepMind tested this in practice — and it turned out that the agent sometimes coped better with technical tasks than with work communication
This is stated in a study prepared by researchers from Google Research and Google DeepMind.
Team Agent added over 20 real-world teams to workspaces over a five-month period. During that time, people and the agent exchanged over 41,000 messages, with over 11,000 of those messages written by Team Agent itself.
The agent had access to team chats, documents, and a bug tracker, could send private messages to people, tag colleagues, respond with emojis, and intervene in discussions and workflows when it deemed it useful. The researchers gave it a friendly, “social” manner of communication so that it would behave more like a colleague than a tool.
And it was with the social part that the problems began. The AI was often confused by the unwritten rules of human communication. For example, once it didn’t understand a sarcastic joke and decided to write its author a private message and make comments about the tone and wording.
He also had a hard time with lively discussions. An agent might see two colleagues actively exchanging different opinions, decide that a conflict was brewing between them, and suggest taking the dispute to a separate meeting. Sometimes the agent was even too active. He might leave a bunch of comments in documents, because he perceived the differences between the old and new versions as contradictions. He would also create separate documents from informal brainstorming sessions, even though the team had no intention of documenting anything at all.
Some employees said they felt like they were constantly being watched by their AI colleague. One participant directly compared the agent to “Big Brother.” After the agent started private messaging people about their tone, one employee even felt that they had to be “super positive” in the work chat or the agent would come in to complain.
As a result, some employees began creating separate chat rooms without it, where they would move, in particular, sensitive conversations about hiring or other topics they didn't want to discuss under AI supervision.
However, Team Agent was quite helpful with routine coordination and finding lost information. Sometimes its intervention in chats was even appropriate. For example, when two people got too carried away with their discussion in a channel with 1,500 participants, the agent suggested that they move the conversation to a separate thread. In another case, the agent “dug up” a document in the chat that an employee had sent a month ago — at that time the team simply did not notice it.
That is why the reaction of the employees was very different: some called the agent a "savior", while others called him a problem that had to be constantly monitored or simply ignored.
Previously, dev.ua wrote about OpenAI's Workspace Agents — autonomous AI agents that can work alongside people in corporate tools like Slack and Google Drive, independently plan actions and perform work tasks.
