Microsoft’s Copilot chief explains AI’s product gap and agent risks

Andreou also highlighted the challenges of agent autonomy, warning that irreversible actions—like deleting files or contacting others—require human oversight. His rule of thumb is to classify tasks as one-way (irreversible) or two-way (reversible) doors, ensuring agents ask for permission when crossing boundaries. Meanwhile, Microsoft’s Copilot now offers an Auto mode to route tasks to the best model, acknowledging that no single lab’s model excels at everything. Andreou also noted that AI’s value risks centralizing in a few companies if product design doesn’t keep pace with model advancements.