0:00 Okay so this is actually my seventh time recording this today. It has gone wrong in every way imaginable. There's a million bugs and I think I've finally fixed them all, but if I haven't, you'll see them because this is the one. One shot if it doesn't go well. This is going to be a shit show and it is what it is. 0:19 All right, so this is a prototype where I am going to explore a couple of things. One is how our agent can learn how to mimic given representatives voice and tone and preferences and style so that they can send better emails on their behalf. 0:37 The next thing is how the natural language interface of chat can be shifted into more of specific direct interaction UI in order to actually work on email sequences and then the handoff from chat into a different surface where your sequences will end up living. 1:01 So let's start by saying like, I want to see 10 VPs of Sales and sales. 1:11 So, regular chat so far, it's going to prospect 10 VPs of Sales for me. 1:19 And as it does that I can, you know, quickly look through. Okay, this kind of wanted. Yes, awesome. 1:31 And then I can say okay let's build an average sequence. 1:34 So, at this point, it suggests for me like hey, I can make it sound like more like you. 1:40 If we do a quick voice exercise you want to do that so I can either skip it and just write the email or I can say let's do it so let's do that. 1:48 It creates a series of mini games for me that don't necessarily feel like chat are embedded here feel pretty light and easy to do, and will give the AI more and more signal about what I sound like. 2:06 So these are pretty easy like select anything you might say I might say happy to chat, say outbound worth the chat sounds good. I don't really use rev up say wanted to reach out that space, I don't really say that space. 2:19 So I think a few things that I'm likely to say, a select things you would never say so I would never say synergy. 2:28 Probably would never say like, touch point, don't really like circling back. 2:34 That's fine. 2:37 Now I can build an email so build an email that feels most like you so hey saw that Acme is is growing the BDR. Do I just do my dash Hey Maya congrats on the VR expansion like what feels most like me I guess like this one. 2:53 And then I get the next page so like the first month usually decides whether reps inherent clean motion or noise. 2:59 Probably say this. 3:04 And then how do I close like a close worth comparing notes next week or open to something happy to send. If this is on your mind, like worth comparing notes feels most like me. 3:13 But now I get to say hey you noticed the VP of sales viewed your profile after the team announced a new outcome benchmark. So, what do I do on LinkedIn do I send like a quick message like game I would love to chat about stuff. 3:29 I mean there's a reason I'm not in sales. 3:33 Another exercise so you're reaching out because the BDR team is hiring quickly and account characterization might get noisy. So how would you, how do you title this email so we're trying to get different signals and both how you write your emails, as well as how you title 3:49 them because those are slightly different things. 3:52 So, lots of hires. 3:59 Okay, we have another mini game of like hey, how do you write your emails like do you write them kind of terse and short long. Do you make them really wordy and short, like, where, what kind of feels most like you. 4:18 What kind of personalizations do like using so like happy weekday like no actually never say that I do use their industry, they do use hiring signals, something they posted is great mutual connections not really my style. 4:35 LinkedIn message. Hello. 4:58 How I think about things. And you'll notice that as I'm doing this just thumbprint is filling out. So, at the moment there's about 26% accuracy, obviously this is finger in the air and this isn't a real model. 5:13 But it's enough for a rough tone of guests not really reliable voice matching yet. And I can click on this and they can say hey here's the voice prompt I have like Joel likes to keep the copy tight expect specific personalization has explicit no go language and 5:28 approachable, and it got it through just like a few clicks for me. And obviously this is like low accuracy, but we can keep going and say like okay make this yours so this is something that has way more signal if I actually write out like we help revenue 5:44 teams improve outcome deficiency like how would you say this that's a bit different. 5:49 I can again go through a bunch more of these like am I witty or earnest and my playful or serious. Let's say I'm like, you know, probably playful and earnest. 6:01 What are your top three group points for unify so again I could if you've seen the previous prototype, you know that we actually do have a lot of information about your company through your domain scraping and whatever we've managed to find. 6:18 Like, okay, you're actually the expert on this, what is your strongest proof point. So, as we learned recently, obviously deliverability, I can add a second best third best that this is giving the model and idea of what I would usually put front and center. 6:34 And then when it has enough confidence, and maybe sounding like me, it starts writing things out like Maya, I do not want to keep poking at this at time. 6:47 So this is like it trying to write a thing as if it's me and so it's either it doesn't sound like me at all, it's like looking in the mirror can say you know it's a little off, actually still. 6:58 I can take away some of the confidence. So, I can also add like a bunch of stuff so I can drag things in if I want to so let's say drag in like a nooks battle card. 7:09 And then it can start getting like just through just through like the things that I upload it can get accuracy and signal much faster. So let's say I go over to my documents here and I want to upload like all of these files, battle cards, you know, 7:26 like building whatever. So, I can give like, you know, we can make all of this like feel very fun and feel like they're just, you know, shooting through this thing without really thinking. 7:37 And before they know it we actually like by the time they write their first email. 7:42 We actually have a pretty good idea of what they like. 7:45 And we can like, you know, incentivize this by making this like ridiculous and make it on fire or something I don't know give them hero titles. 7:55 More of these again like the idea is this can be endless really more of this like what's the most effective way you stand out like what's your secret sauce, basically. 8:07 And then at some point it's like hey do you want to speed this up, and some ways to speed it up or like paste in your best emails. 8:15 This is something you've written on LinkedIn or blog post, you can connect your mailboxes so we can actually scrape all of your mail and that'll get us to like 100% pretty fast. 8:26 But you can see this is like constantly evolving. 8:30 As it learns new thing like tense playful and earnest like prefers email structure, which words I like which words I don't, you know, product proof points, it starts building an idea of who I am and how to talk like me. 8:45 So at this point, it's like, all right, there's diminishing returns you've reached like 60% accuracy. 8:53 And I can now roughly mimic your style but at this point, if you want full accuracy, we're going to have to work on real emails. So let's go forward and kind of build this email together that I wanted for those features of sales. 9:05 So it maintains context of like what I wanted what I was doing so even if I forgot about this. It still remembers what we're doing here. 9:14 And so this is the email screen where over here we have the template it's built for me. Again, it's built in everything that it learned about me and my voice from the previous step so I don't really have to read teach it all of this stuff that just has that in memory. 9:31 So ideally, this is a thing that doesn't make the same mistake twice when I when I corrected. 9:38 Now there's a few things here to see one is first as you notice right this is a real UI embedded in chat like the games. 9:49 And I can start working on this without kind of moving to a different place, saving a sequence, building into different steps, adding snippets, I try to make this as lightweight as possible. 10:00 So here's my template. Here's an example and I can go through a bunch of the folks are like in the real world, these would be people from this list up here. 10:10 But that translates whatever's here into this. 10:13 Now, as everyone knows, AI snippets have been a kind of a mess, not super intuitive. Lots of like, you know finicky steps to get to manage them. So I was like, okay, we have, you know, these right we have variables that we can easily add. 10:31 So I'll just extend that and like when the thing that you wrote is not matched to a specific variable, make that an instruction for AI and if you say hey go research this thing. 10:43 That's what it goes and does that cost some credits and cost some credits we can we can show that in some way but I can create these things pretty easily and like in a very self explanatory way. 10:57 You can see like mention a recent hiring books like sorry tool is hiring BDRs. 11:02 So when VP lead teams are scaling without clear signals the usual failure mode is reps working the same account, but I don't really love this paragraph. 11:12 So I'm going to remove it. 11:14 This regenerates all of these actually regenerate so if I want to jump through them. I can see those things. And you might notice these like to regular brackets down here. 11:26 And so another thing I've been thinking about is this idea of like how to connect AI copywriting and the things that we want to be described verbatim so like exactly the way I wrote them don't change anything. 11:43 But also on the subject is like spintax so like, do I have to say hi hey hello, or is that a thing that the AI can just take care of for me. So I don't ever have to think about it. 11:56 All of the things here are actually like kind of par for the course the things I put in. 12:02 Or for the course all of these things are up for grabs right so like things inside the purple brackets are like okay this is all you, you go do this use my voice but like you decide what to say here. 12:15 Things here are like, you can rephrase this lightly but this is kind of what it should say, and things within these brackets are static. 12:22 So you're essentially like freezing this, the AI has no access to change this. It has to write it this way, every time at this place in the email. And I think this concept can make sequence writing just like very very easy, where we can like use 12:40 AI to look at this and be like, okay, this is this has like an agent request that is going to go out and ask a question that's going to cost the credit, and we can figure that out. 12:52 But anyway, I made some changes a little paragraph I can say like hey instead of hi, instead of cheers I can be like best. 13:00 And you'll see here as well. Okay, I wrote best, but it doesn't always appear here sometimes it says this time he did. 13:08 Okay, I made some changes let's integrate them. 13:13 And so it's continuing to learn about my preferences you see this like signal over here, still filling up my fingerprint. I can still see this at any given time. 13:23 And now it says like you made a bunch of changes. I'm not just going to change that I want to learn about why you made these. So there's a feedback loop here where every email that you generate is always going to be better than the last. 13:35 Hey explanatory preamble, not really a thing that's my thing like I want to lead with a pain point, and let's do it shorter opening before the core message like set two of these things let's say a real, and I can add whatever else I want and I can say okay let's go to 13:49 the next step. 13:51 And you can see like, well you didn't really see there but if you noticed, the thumbprint fills in much faster when it's working on real outreach that I'm willing to stand behind. 14:03 The other thing here is like this email score. And so this is using some of the docs that Ethan put together about how do you write good emails and like cited a lot of different sources different ways of thinking. 14:17 Basically the model is putting all of that together and saying hey, is this a good email is this not a good email. I can also hear seeds making a pretty dumb mistake of just like putting the company name here inside the static and you can see it actually does respect it so let me get 14:34 rid of this. 14:38 I really can't remove this. Okay, cool. I mean if I if I couldn't then that was a, that was the thing you would have had to deal with. But let's let's just say remove that I'm going to integrate some more changes now before we move forward but that was definitely not possible 14:54 Okay, I don't want it to actually learn anything that you see this, this, this point higher. Now I can say okay, it put it back. This is the kind of stuff you're gonna have to deal with in this demo, but it wouldn't otherwise. 15:08 I can say it looks good. And so now it's created a sequence draft for me. 15:13 I can look through this email and I can say okay other than this bullshit. 15:19 I can say review each email before sending. I can add a call here, LinkedIn response a reminder after three days. I can do like another email if I want to. 15:31 That goes out automatically I can either review this or not review this, but this also just like it kind of knows it uses the same stuff. 15:41 And it knows the context of the previous email of what we're trying to get at, I can change this to however many days I want to change days to month, whatever it's all of this, the idea is it feels really lightweight. 15:54 And then when I'm ready I can either save and enroll those 10 people, so it still has context here. I can save, just save the sequence so let's do save and enroll, you know, first this looks good. 16:05 So I don't have to say everything looks good but I can enroll unless nothing here it is a draft so let's say okay wait three days remind me wait nine days and just do the automatic email I can add a call if I want to. 16:19 So I'm going to click save and enroll those people. And so now that saving the sequence, the chat naturally move to the side, because now it's my time to be in the sequence itself and you can see like okay here's my sequence I set up here's the copy 16:37 wait three days wait nine days. I have the enrollees here which match, you know, these folks up here. 16:44 So that's kind of taking care of this for me and now I have this like, push this down. Now I'm in this space where I can like see analytics potentially where I can see like different people status here and it also is going to appear on the sequence 17:00 So like that's how this thing becomes real and how chat stays with you so I could ask things about the sequence if I want to if I'm on the sequence page, it's going to have context for that. And I'm going to be able to ask about the analytics here, 17:17 send us to more people enroll more people either via the UI, or via chat. And so both of these things the ideas, whatever you can do with the UI, you can always do a chat. It's just sometimes going to be much nicer and much more efficient to use the UI for this. 17:34 So this is kind of my vision for how we go from prospecting the identify job through chat into this engaged job, whether it's within chat or outside of chat, depending on where we think the best place to do it is, but also allowing to do it anywhere. 17:56 So I hope you liked it. Sorry that this took literally 500 hours to get out. But I hope I hope this is directionally where we think we're going, and we can cook on this real fast. 18:10 All right.