1. The user (you) needs to provide it in the prompt for the call. I have ideas for being able to build memory context for each user so you wouldnt have to re-provide those details each time you want to make a call and instead pull it from memory bank, but thats not implemented yet.

2. So far I have actually never been called out for not being a human. I think thats in part because these people probably get a lot of pretty crazy calls and by comparison a slightly robotic over-the-phone voice has a lot of forgiveness. When you know, its easy to hear though. I did build it with Elevenlabs TTS to make the voice as natural as possible. Models have gotten much faster and so latency feels good, not outstanding

3. So this IS very interesting, I have to refine the system prompt extensively to get it to the right balance of objection handling without being weird and obviously non-human. It is very clearly on the side of the human, in fact in the NYT case it was initially told no but pressed ahead anyway. "I understand that, how do we move forward" "any other way we can resolve this" that kind of thing. Better still, customer service has a very high abr for hanging up the phone,so the persistency is very effective

Yes I think the API holds real potential. The space is busy, but I think there could be room for being the simplest possibel way to make a phone call programmatically. There would be less customization than a Vapi or Bland, but it would take 10 minutes to implement. Basically anywhere engineers experience automated workflows breaking because the real world requires a phone call. Voice probably wouldn't be core to these users business though, int hat case they'd want more control than we would offer.

Let me know if you have any issues trying it! Very curious to see how folks end up breaking it. We have a discord you can join here (https://discord.gg/2tn2ugXu) if you have any problems to troubleshoot (or email me!)

My friend and I hacked together Pamela over the new year to make those customer service calls on your behalf and report back with a summary of what happened (https://www.thisispamela.com/).

You prompt it with who to call and what you want to accomplish, provide any relevant context, and it places the call. You can monitor progress via live transcripts or listen in. The agent can navigate phone trees (“press 1 for…”) or speak directly to a human when needed.

I now use this for everything from calling up my local super popular bakery to reserve a croissant before they run out, to calling the New York Times to request a refund when they moved me from the promo rate to their regular rate (it worked!).

This is early and imperfect. Some calls fail, and some situations still require the account holder to intervene. But it already works often enough to be genuinely useful.

(There’s also an API underneath this for anyone who wants to experiment programmatically: https://docs.thisispamela.com/)

Comments URL: https://news.ycombinator.com/item?id=46873719

Points: 2

# Comments: 2

Not bad tackling ocean acidifcation to boot