How We Use AI to Replace Employees

Colin and Brent discuss how they are using AI to write code and do marketing for their software companies.

Apple Podcast / Spotify / Google Play / Overcast

Want to get more breakdowns? Subscribe to my free newsletter here:

Enroll in Colin's "How To Buy a Small Business" Course & Community

Hire top 1% remote talent in Latin America.

Sell your SaaS at

[00:00:00] Colin Keeley: All right. Hello and welcome back. This is Colin Keeley here,

[00:00:02] Brent Sanders: and I'm Brent Sanders,

[00:00:04] Colin Keeley: and we are two guys buying and building wonderful internet companies.

[00:00:07] Brent Sanders: Yeah. And then it's probably been a couple weeks since the last time we recorded and wanted to get something down on, on, on tape. We can't really say on tape o on, on record of kind of what's been going on and what we've been up to. I guess one thing that has been going on is I feel like our team has been getting a lot better, right?

Like things are gelling, at least on the scout side of things, except everybody's kind of running into, personal issues. Everyone has stuff, but it's been nice to see everybody kind of help each other out. So, you know, we have a team that's, there's some people in the Philippines, people in South America, people here in Cleveland Austin.

Right. Like, and things have been happening with the team personally that, you know, people have to take days off. And it's been really nice to see everybody kind of chipping in extra time in order to, to, you know. Make their schedule more flexible, which has been nice to see. It makes as you put a team together, it's nice to see like everyone kind of gelling together and helping each other out and covering for one another.

And that's probably like the only big update on, on the portfolio side of things other than just things do continue to grow. I, I don't have the numbers in front of me, but I think I saw your last update. Everything's kind of still incrementally growing, moving upwards, and progressing well.

[00:01:24] Colin Keeley: Yeah, well it's, so we do investor updates once a month, which is pretty high level.

It's just like the financials, the MRR, and like basically bullet points on what was done last month, what's working and what's not. Which investors seem to really appreciate highly recommend doing that. It's good like accountability just generally. But yeah, and people issues. So we have a number of employees at the burn companies and then at Jules, it's like a lot of what we do is staffing.

So technically we are employing these people and we're like, you know, making some profit on top of that. And at scale it's just like endless random issues come up just for humans. So it's like, there's car accidents, there's deaths and families you know, people are always trying to get better jobs.

You have like almost extortion of your like clients where it's like, I'm gonna leave unless you give me an a pay increase. But it, it is just endless people issues. I see.

[00:02:15] Brent Sanders: Yeah. Yeah. How involved are you in that? 'cause you know, as I, I think you mentioned on our last part, it's like people keep asking about, oh yeah, I see you guys are doing the staffing thing.

I'm like, I guess like I don't really do it. That that's more Collin's bag. That's your, your thing. I mean, it seems like it's growing well, right? Like from what I've heard from you we talked about it a little bit last week. How is that going?

[00:02:34] Colin Keeley: Yeah, it's going well and it's growing. As far as how involved I am.

I have a good person that kind of runs recruiting now, so he's always hiring and testing new recruiters. And then I have a person that basically runs sales. Mm-Hmm. So between the two of 'em, you know, they interact and they basically run the business and I kind of keep on top of things and I watch the, you know, cash payments and then, you know, occasionally I'll tweet it out or I'll post to my newsletter.

But that's really the extent of my involvement and probably the extent that kind of going forward is like when we have extra capacity, I'll send a few tweets and they'll like, keep 'em busy for a time. Nice. Yeah, it's, you know, gotten to the point that it's self-sufficient. I'm certainly not out there recruiting anyone or talking with clients.

[00:03:16] Brent Sanders: Sure. Cool. Yeah, yeah. Sorry, go ahead.

[00:03:21] Colin Keeley: Do you wanna talk we've both been noodling on like AI assisted, you know, things. You on the tech side and me on the marketing side. You wanna talk about some of that stuff?

[00:03:30] Brent Sanders: Yeah. So I was going to, you know, when we were last. When I was in last, in Austin, I visited you, we were kind of noodling on this idea.

We talked about it with a couple people around, like, you know, is SaaS gonna get killed? Because AI can just generate new SaaS products entirely. And, and I think I touched on this on a prior po, we were talking about GPT Pilot. And so I went deep into that and I was gonna record a whole like YouTube of it.

Like here's my experience and here's what to expect. But I found it to be really boring. Like 90% of what I was doing was waiting for it to type. Code really slowly. But yeah, I'd love to share my experience from a high level of like GPT pilot evaluation. So I used it for something really simple. We have had an existing contact form that's used on Scout that people embed in their site and that's already been built.

And I was like, okay, what if we wanna replace this? I wanted to add some features and I had a very clear spec. And it was as as follows, it was like build a con, a a, a node based, node JS based contact form that hits an end point. So there's no server, it's just a front end thing. Write it in view js, make it so I can embed it into anyone's website or like a WordPress site and you know, have a short version and a long version.

And you know, I was very detailed so. Lemme back up a step. So the way GPT pilot works is essentially uses GPT-4 or any of the open ai or you can actually plug in other models to it, other large language models into it, such as your own well actually I'm not so sure if I do your own, but you can plug in a bunch of different ones.

I used open AI 'cause it's easiest and already have an account and what it does is it breaks the. Development process into roles, and it kind of sets up the first role, which is like the product manager. And they gather the requirements and ask the questions, and then that role passes it to the software architect.

And the software architect comes up with, well, it should be written in this, with that and that. And then they pass it to sort of the code monkey, and then their code monkey passes it to the code reviewer. And then finally it kind of comes back to you as, Hey, I wrote this code. Can you try it out? And so there's these checkpoints in which you kind of.

Run through. So I'll get to the, the really, really short version of it. And I, and I think like the one point that they put in their new, in their read me in this project. So first of all, like the idea of segmentation is, is smart. That seems like a really novel use of like, I'm gonna use GPTs individually 'cause they're good, more concentrated.

But until there's like a GI like this tool, this tool set is just not going to be better. And so the, the short version is. It's more expensive and takes longer to use this tool to build something like a contact form than it would be to use sort of like lower end cheaper labor on like Upwork. It certainly is faster to get started, but I'll, I'll go through a couple of the quick pitfalls.

So we immediately started writing code. I had a, you know, I spent time writing a spec, which I'd have to do with a, a human engineer. Anyways, I write this spec, I give it to GPT. I am using specifics, but I missed maybe one. I said view js, right? And in view js, there's multiple version. There's View one, two, and three, and I believe current version is three.

Right? I, I could be wrong off by one, but like it immediately started writing this, you know, unbeknownst to me using View version two, and it starts layering more stuff. And I'm like, oh, once I did the first code review on my end, I started looking at the requirements. I'm like, ah, shit, this thing, it's, it's written a bunch of code in the wrong version.

And that's kind of like, as you see with GPT, it, it's using stuff that it's been trained on. So it was trained on the older version. So like you have to know these things. And so there's already this level of expertise where even if I was less technical and I didn't really know the difference between view versions, it was gonna build something that immediately out of the gate.

Is going to have deprecated parts, security issues, like, and mind you, it's just a contact form. It's not a big deal, but it just immediately came to me as like, oh, this is, this is like an easy way to waste a lot of time and money. So I wanna talk about the other part of this, which is you use a lot of API requests in this.

So you have five, four or five roles that are kind of kicking messages back to each other and engaging themselves. And using tokens and you are charged by the tokens. By the end of this project I ended up with a contact form that was like barely usable like I could have written. I I spent an almost about an hour and a half working back and forth, letting it, right, reviewing its work.

And I was doing something else while it was doing, 'cause I'd wait for the prompt to come back. So I don't wanna say, I'm like staring at it watching. But I'm confident I could have either a, written it myself in an hour and a half using, without, you know, bumping into like, Hey, I know that I shouldn't use the old version of UJS.

And that was just one bump of many that kind of came up. And then I'm ending up with something that's like, you know, I, I went back, got us use Ujs three, the latest version, and, and it, it kind of got through it, but it still was kind of semi-functional and kept having to go back and. So I guess the sum up here is it's slower and more expensive than what I would say if you went to the Philippines or maybe Eastern Europe to say, Hey, I wanna get this done.

And you know, it may be like over the calendar time it would've taken less or taken longer, but it definitely cost us, I mean, I racked up. Just in, in this short process, at least 20, $30 of, of fees against the API to get this thing going. So it was, it was a little disappointing. That being said, it was also a little chilling.

It definitely had that moment of like, oh my God, this thing is smart enough that it's talking to itself and working out problems. But, you know, overall grade I would give it as like maybe a d If it was a human delivering that code result or that product, I would not have hired them and I definitely wouldn't have paid them $30 to, to build a, a silly contact form.

So anyways, that's the, the sort of quick and dirty result of it. I, I was debating kind of recording and, and kind of publishing and editing all this stuff, but it really was a mind numbingly boring process. So I don't think it's, it's interesting enough to publish to YouTube.

[00:09:56] Colin Keeley: I mean, it's interesting because this is the worst it's ever gonna be and it's already Mm-Hmm.

Somewhat compelling and increasing at like an amazing rate.

[00:10:05] Brent Sanders: Yes. So the thing that I don't, so somebody asked point blank answer. As of right now, until we get to a GI SaaS, businesses are safe. But even still, I think there's going to be a fair amount of, like, you're gonna get to a junction. And you're gonna make a decision, or it's going to make a decision, it's gonna make the wrong one.

And if as you work backwards, the cost of that decision being wrong, the closer that decision was to the beginning of the project, the more costly it's going to be to go back and change it, the more time that's gonna be wasted. So, it, it is interesting and compelling. I'm excited, you know, for the future when we can leverage it, but for right now, the way I leverage it is the way I recommend to you to leverage it, which is like.

Give it the right questions and don't expect it to do too much. So writing my code is, is for me. That's off limits. I don't want it writing my code. There's just more, there's just a lot more errors and and hallucinations versus like, Hey, how do I do this specific thing and can we work through a problem together?

And then I'll apply it and be sort of the thing transposing the eventual solution into my code.

[00:11:17] Colin Keeley: Nice. Yeah, it's been the most effective for me trying to analyze data and not like giving it data and say, analyze this. It's like, how do I make a new column that analyzes, you know, this big list of transaction data? Um Mm-Hmm. So it makes you basically like it Excel or Google Sheet Wizard when I, you know, and certainly not, or wasn't before.

And that would've taken me forever looking at help docs and trying to figure out, you know, certain things like that. And then you can just keep saying like, Hey, break it down more for me. Like, this is still too confusing. What does each piece here mean? So that's been great. One mutual friend of ours who runs a, a growing vertical software company has been generating a ton of AI articles and he says it's just been, , great for growing his SEO and getting free trials and everything.

And so he gave me like his rough playbook and so I've been testing it out. He uses byword to generate these articles. So it's by and it costs like $3 or so for each article. Mm-Hmm. And you can easily generate, I mean, 30 articles, high quality ones in an hour. And so there's different ways of approaching it.

Like the most devilish way is to look at your biggest competitors and their best SEO pages and just grab all of those URLs and you could give it to Byword and say like, generate better articles of these, and it'll go check out those articles and write, you know. Headline, sub headline, all that stuff. And then it will also interlink everything, interlink all your articles and Nice queue them up to be published to Webflow and also generate a new, unique image.

So I've done it maybe like 30 to 50 across a few different sites, and you already see 'em ranking relatively quickly on Google Search Council. So you see them like it's really good for generic things, so a lot of stuff around dog walking is pretty solid, but it doesn't have like super unique data or anything like that.

Mm-hmm, our success stories or like customer testimonials. , those are still best, done by people, but I guess we'll see whether it actually delivers results. It's already delivering clicks, and I assume some of those clicks are worthwhile and everyone seems to be doing it. So I don't think Google's gonna, , slap down every website in existence.

It's just probably the, the new world we're living in and there's like some kind of arbitrage opportunity over the next few years of publishing a bunch of articles. And as long as you have, you know, high enough domain authority, you're gonna be ranking. And there isn't that much competition for like some of these kind of longer TL keywords.

[00:13:41] Brent Sanders: Yeah. Yeah, it's interesting. I mean, I would like to think that all of this AI stuff generally is trashed. Like I see a lot of content on LinkedIn that people are, I can tell, are generating, or like heavily leveraging it and it's like you just lose their voice. But if you don't know the person or you know, it's, I think for that kind of content.

I also don't think people are really reading these articles. I think they're skimming them, bullet, pointing them, trying to find that call to action, trying to find a specific answer to a specific question, and. You know, I, we've all seen trash articles written by humans, so it's, I'm sure there's a way to tell, oh, this is likely generated based on like the, the cadence or the, the writing style or whatever.

But you know, is it just gonna create a bunch of trash? Is the, is concern, right? That's, I guess, Google's concern. They start ranking trash articles. But from your perspective, is this, does it read like trash? Does it read bad? Does it read useless, or is it. You know, it still adds value.

[00:14:41] Colin Keeley: So you can make sure it reads well.

So you could dictate the writing style. So like for ours, I did professional engaging in academic, and so that actually reads quite well. It doesn't look like it's AI generated at all. I mean, it looks like it was written by some SEO expert 'cause it's like mm-Hmm. A beautiful headline, sub-headline, like the perfect format to make everything super readable.

And then it's like, does this stuff deliver value? And I think at some of the you know, simple, high level you know, topics, it certainly does like anything related to definitions or like, I don't know how do you create a dog walking flyer And you could have five steps in there. It, you know, someone is actually looking at how do you create a dog walking flyer to like grow your business and is probably, is relatively helpful.

It'd be more helpful, it got more detail than like, went through the actual process, but it doesn't have that, you know, advanced capabilities. But I think it's enough to give people like, huh, the scalp thing is like, maybe interesting. I'm gonna poke around a little bit more and not like, oh, this is complete junk.

I, you know, was tricked into clicking this link.

[00:15:41] Brent Sanders: Yeah, right, right. Cool. Well, I mean, it's been, it, it's such a cool topic and as you point out like it, this is the worst it's gonna be. I am really bullish on, obviously it as a, you know, category, but I'm really bullish on the idea of fine tuning or building your own custom models.

And I, you know, I've been doing that in other domains, less so large language models, but like, and it gets back to this problem of like, Hey, you need the data to train the model. And so organizing your data and thinking about how you're gonna train a model, you need to have a statistical significance.

You're gonna have, you know, thousands upon thousands of records to say, this is the, this is the prompt, this is the response, or this is the thing and this is what I expect to be predicted. So, you know, I think as we look at deals, as we look at new businesses, that is. Or even our own businesses. That's always kind of an interesting thing of like, okay, well what information do we have?

What information could we use to train and are we structuring it in a way that it's accessible? So that's always kind of a fun thing to be thinking through.

[00:16:47] Colin Keeley: Yeah. That would be the compelling argument for like all the vertical market software companies. Not that they're just gonna be recreated in AI and be super competitive, but like what kind of unique data do they have in that you could leverage with AI models?

[00:17:01] Brent Sanders: Yeah. Yeah, I think it's, that's, that's where I'd like to spend more time in the, in the coming year.

[00:17:10] Colin Keeley: That was sweet. I know you gotta run, you know, quickly here. Anything else you wanna cover quick?

[00:17:14] Brent Sanders: No, no. We'll talk about, I, I did a book report, I publish it and we can talk about on the next one, but wanted to get something on, on record, on tape.

I have this one quick story since we have like a minute left here. So I started getting into tennis and I got into it because my parents were moving outta my childhood home and like giving away a bunch of stuff. And on the pile was a few tennis rackets. I was like, I'll salvage these and you know, maybe I'll get into tennis someday. So I started getting into it more and more. I've been playing with these rackets and getting 'em like re strong and everything. And I took a lesson in Austin and the guy's like, you have a weird swing. Let me see your racket. And it turns out I've been playing with a racket design for old ladies for like two years now.

[00:17:58] Colin Keeley: And it's

amazing big

swing where basically like pull back because it's designed to give old women more power so they could get the ball over the court. And so yeah, he handed me a men's racket. And a men's racket is like, I don't know, two to three times the weight and way thinner. It looks like, you know, a child's toy versus like a real tennis racket.

[00:18:15] Brent Sanders: So have you, did you, did you buy a, a men's racket?

[00:18:21] Colin Keeley: I do, I now am playing with a men's racket. And it's amazing. Good. The game is so much easier. Yeah, it's,

[00:18:26] Brent Sanders: I bet, I bet That's great to hear. Yeah. How does it, is it the strings that are giving it more sort of like bounce?

[00:18:33] Colin Keeley: It's both, the strings are designed differently, but like you can mix and match strings and then the weight is different.

And then like the actual, I mean, the racket is, I think stiffer and substantially heavier.

[00:18:45] Brent Sanders: Nice.

[00:18:46] Colin Keeley: It's probably like three times the weight. Yeah, it is wild. But anyway, that was just my little bit.

[00:18:51] Brent Sanders: Cool. Well hey, it was good to chat. Thanks for listening to everybody and talk to you guys soon.

[00:18:58] Colin Keeley: Yeah. Take care.


Continue reading

Subscribe now.

Get the occasional update with helpful guides and insights.