Why are AI agents designed with approval checkpoints for sensitive actions?

AI agents are designed with approval checkpoints, particularly for sensitive actions like payments or account changes, to implement a 'human-in-the-loop' model that prioritizes security and user control. This approach ensures that while the AI can navigate apps, draft purchases, or prepare bookings autonomously, it cannot finalize these actions without explicit user confirmation. The primary reasons include preventing financial loss, protecting sensitive data, and maintaining user trust. For example, in banking apps, transfers already require confirmation, and this principle is now being extended to AI-driven actions across various services. By requiring approval at critical junctures, companies mitigate risks associated with errors or unauthorized actions, ensuring that users retain ultimate control over important decisions. This design also aligns with privacy protections, as sensitive data can remain on the device rather than being sent to external servers, reducing exposure to potential breaches.

📖 Read the full article: Why companies like Apple are building AI agents with limits