Currently, Athena is not able to reliably determine whether a call has been answered by a live person, voicemail, or an automated voice assistant. We understand that detecting automated systems using specific keywords or phrases is not dependable because: Greetings can be fully customized Messaging varies by carrier and device AI assistants may sound conversational There is no consistent phrase pattern As a result, Athena may remain on the line without actual human engagement, which increases unnecessary billing exposure. Could you please clarify whether there is any effective or recommended way to enable Vapi to detect that the receiving end is not a real person but an automated system (such as voicemail or an AI assistant)? Please note that voicemail detection is already properly configured using the Vapi (recommended) setting. However, even with this configuration, Vapi is still unable to effectively identify when the call is answered by an AI assistant. If there are any built-in features, integrations, or best practices to better handle this scenario, kindly guide us. Thank you.

Hi Azhar, Thank you for reaching out. Currently, there is no native solution to distinguish between AI and human responses. However, we do offer a voicemail detection framework that you can implement for your use case. If you have already implemented the voicemail detection framework and it’s still not working as expected, please share the relevant Call ID(s) so we can investigate further. We’re happy to help and look into this for you. Best regards, Oshi Raghav Customer Support Team | Vapi

You’re right that detecting AI assistants versus real humans is extremely difficult today, especially since many AI systems sound natural and don’t follow fixed voicemail patterns. Even with Vapi’s recommended voicemail detection enabled, AI assistants often appear as live answers at the telecom level, so they won’t be flagged correctly. At the moment, there’s no fully reliable built-in way to detect this instantly. The best practical approach is to handle it through call behavior, such as checking for meaningful engagement early in the call and ending it if there’s no real interaction after a short time. This helps reduce unnecessary billing, but due to how advanced AI assistants are, a 100% accurate solution doesn’t currently exist.