Are you ready to be an AI Boss?

In the words of Andrej Karpathy (former Sr. Director of AI at Tesla and the brains behind self-driving cars) - it is the decade of AI Agents.

If you’re not building agents, or preparing for their deployment, you should be. But despite what you may be hearing, they’re not taking the form of a digital colleague anytime soon, rather, they will take the form of a human-supervised semi-automated software implementation.

In Microsoft’s recent Work Trend annual report, they raise the concept of an AI Boss. It refers to the human that is a “manager” to a set of AI Agents as they attempt the tasks they were engineered to perform. However, I want to challenge that idea.

Let me give you a hypothetical: You’re an AI Boss of multiple agents, one is designed to handle online customer refunds; there are hundreds of conversations a day, you read logs of a few for quality control. One night, users discover a loophole in the logic that allows them to convince the agent to give them refunds ad infinitum, and your company loses hundreds of thousands of dollars.

If this was a human agent, the person who gave out all these refunds would be fired. But, in their current state, AI Agents cannot be accountable, they are pieces of code being executed without the capacity to learn on the job (yet).

In the AI Boss paradigm, you’re the one accountable for your AI going rogue. This would also imply that you should have seen this coming and corrected for it somehow.