Teaching a Door to Say No (2389.ai)

A newly developed AI agent now manages access at the front door of an office, enhancing security and efficiency in visitor management. Previously, calls to the office required a person to manually respond and decide whether to grant access based solely on limited information. The new system integrates a voice assistant capable of engaging callers and gathering their identities, while all critical access decisions are made server-side, ensuring that the AI handles conversation without having authority to grant entry. This design maintains accountability by separating conversational capabilities from decision-making processes. This innovation addresses significant security vulnerabilities inherent in the previous setup. The system employs fuzzy matching for identity verification, allowing for natural language input from callers while rejecting potential adversarial prompts based on specific regex patterns. If the AI cannot confidently authenticate a caller, it escalates the decision to a human. The project's architecture experience—and the transition from v1, which relied heavily on unnecessary operational timers and failed transfers, to a streamlined v2—demonstrates the importance of robust testing and evaluation coverage in AI systems. Overall, this implementation not only improves user experience but also enhances the safety and reliability of entry control through innovative AI deployment.