Jason M. Lemkin used an AI agent to write personalized emails to CEOs of 10,000 companies he invested in. The AI emails outperformed matching software, took 2-3 minutes each, and identified eight competitors' executives attending an event plus 20 high-value meetings. A founder called it the best marketing outreach he'd ever received.
I had 10K write it to the CEOs of all the companies I invested in, and it went through all 8,000 people coming and assembled the best list for them to all meet. And it destroyed all the matching software we have here. Yeah, destroyed it. Destroyed it. Now, it did take like 2 to 3 minutes per email. That's a lot of compute and context. Yeah, a lot. But— and I sent it to one of my founders and it uncovered that 8 of his competitors, 8 of the entire management team of his top competitor were coming and all the adjacent folks in his marketplace. And then it summarized why maybe it wasn't worth his time to come this year. Oh yeah, no, he's like, he's like, yeah, no, he was like, you're a part of vertical SaaS is relatively lightly attended, but here are the 20 people you should meet. And he's like, that's the best marketing email I've ever gotten too. I don't know how we got on that track, but that is, um, yeah, well, it's super powerful, right?
Well, that's, that's a piece that I've seen that you and Amelia are extremely good at with working with the agents. We were talking earlier about like, instead of focusing on the how for it, focusing on defining the end state and the goal that you're going for, and also giving it context on the why of like, And that's why I think like my kind of like pet theory about like why you've been so successful with this is that you have a lot of scar tissue over like a lot of doing this is, you know what right looks like for a lot of these things. And so you're also, when you're communicating with the agent, is that you're very good at describing what right looks like and then kind of leaning in and letting it figure out of like the path that it's going to go down to achieve that kind of like right end state for it. But then you're also feeding a lot in because like you always talk about it as like training the agent for it and sort of like, hey, you can feed like the world of context into here, but also what— how should that context be looked at? It should be looked at from this perspective of choosing this end state, this goal. This is the actual problem I'm trying to solve for it. And then not getting in the way of the agent figuring out the interesting way to solve that problem.
