Subscribe
Share
Share
Embed
Hosted by Runpoint Partners’ founders Sam Gaddis (tech entrepreneur & AI builder) and Matthew Hall (PE operator & growth strategist), Runpoint Podcast strips the hype from artificial intelligence and shows you how to turn it into concrete business results—fast.
Matthew Hall (00:01)
I love saying, how's it going?
Sam Gaddis (00:04)
Man, it's going great. I feel like things are coming together. I feel like ⁓ people are realizing that AI works and it's working for us. It's working for our clients and we're starting to see that all over the news. So it's great.
Matthew Hall (00:18)
Well,
Cool. All right. Well, we are building off of what we did during our last podcast, tidying up a little bit, leaning into the take economy, as they say. We've got six topics to talk about today that cover basically all of the news and most interesting things that have happened in the last couple of weeks in AI. We've got a shot clock. So we're going to talk about each for two minutes, and I'm going to run through each of these topics and see what we agree on.
What kind of new things our audience should be paying attention to? What is real and what's bogus? You ready to get into it? All right.
Sam Gaddis (00:53)
Let's do it.
Matthew Hall (01:05)
First up, we've got a new study that dropped. We spoke a few weeks ago about the infamous MIT study that showed that 95 % of AI projects fail. But there's a new study now from Wharton, an equally reputable school and study that shows that actually 75 % of firms are seeing positive ROI with their AI initiatives. Less than 5 % report negative ROI and 46 % of leaders using AI daily. I think there was another bullet in this study.
about how almost everybody is increasing their budgets next year for AI initiatives. So my question to you, Sam, two minutes on the clock, is who's right? Is it Wharton? Is it MIT? Doesn't matter.
Sam Gaddis (01:53)
Yeah, I love this. I'm so glad that we're talking about this. I, in my talks, will reference that MIT study that went viral two months ago. And when I did the talk ⁓ at Capital Factory the other day, I did the show of hands thing, how many people have done AI projects that have failed? By the way, the title of the talk was why 95 % of AI projects fail. And nobody raised their hands. And I think, one, these people are liars. And two,
It's actually nowhere close to 95%. And that was the wonderful thing about this new warden study. It finally felt like it reflected the reality that we're seeing every day. Yes, these projects do fail. There are necessary experimentations that have to happen to make this stuff work. We cannot say definitively what works and what doesn't. But I actually was talking to the authors about this ⁓ on LinkedIn.
and
It's just the case that executives especially, middle managers kind of disagree, but I think there's a bit of maybe potentially some fear underlying that, but executives agree across the board that their Gen.ai experiments are paying off for the most part. So yeah, I think this is awesome. It's a great study. I read the whole thing. We actually just published a,
a TLDR version of this on our website so if people want to check that out they can go see the highlights and also get a link to the study.
Matthew Hall (03:30)
Yeah, I think if you need a study to sway your decision-making on if AI is worth investing in, you're done. That's my take. You can clip that for social. This is so clearly a technology that is valuable, that will change how work is done. You gotta figure out how to do your business or else you're gonna get left behind. All right, let's move on to question two. Reset the shot clock.
I'll set you up a little bit here. So news of the past week is that ChatGPT released their new browser, which they call Atlas. And ⁓ what it basically is is a skin of Chromium, the same underlying technology that Chrome is built on. ⁓ And it has a few interesting embedded features. No matter what web page you're looking at, you could open up a chat window on the right and ask ChatGPT about what it's seeing.
You can also ⁓ start any engagement ⁓ with a ChatGPT conversation. You can highlight text, it rewrite things for you in emails, and then you can do agent mode, which is what ChatGPT was able to do previously, but it takes control of the browser, can actually ⁓ access the mouse, can click on stuff, can move around. So question to you, Sam, is...
Does this matter? This is a combination of features that kind of existed previously, but now in their own product. And it seems bold. There's a lot of noise about this that they're going after Chrome and Google, which is huge business. But, you know, I'm not sure where adoption is going. What will we know if this is working? Do you think this is a story worth paying attention to?
Sam Gaddis (05:13)
No, I get why they did this. I think that something like this is definitely part of the future, but this was a dud for me. Everything that this purports to do is still easier and better to do in the chat GBT UI. There's just very few scenarios where you interacting with chat within the website will take you far enough to get you the answers that you want. ⁓
the responses are throttled such that it's not reliable for any sort of scraping. If you're trying to do that kind of thing, if you want to go, for instance, read all of the API docs, it's not as good as simply going to chat, GPT and dropping the link and asking questions there. So I think this will be where things go, but Chrome is going to do it as well. And Safari will do it eventually. Maybe who knows what Apple is going to do. So, yeah, I just don't, I, I've not found the value from this and
doesn't get me excited at all. I'd much rather be in chat chippy tea.
Matthew Hall (06:20)
I'll take a slightly different tack on this. I do think it's exciting, but it's a starting point. And you never know with these new products if they're things that OpenAI is actually going to actively invest in, or if this was just sort of a ploy, a PR stunt, something they're going to abandon. I don't know if this is a cornerstone of their business or just something that they're trying. Right now, it can't make use of all the context in the browser.
And that's something we talked about a few weeks ago, which is I really thought that lock-in was going to come from whichever LLM had access to all of your personal docs and all of your history. Your browsing history is obviously massively important context-wise, but the context windows aren't good enough and the caching is not good enough for it actually to be good at that sort of stuff. It's way too much. It overloads it. And so none of that's useful. I do think it's immediately useful in ⁓ code debugging and previewing what you're building of having the chat GBT.
you can write specs based on best practices about how you can improve UI. So I am using a little bit in development, or at least trying to.
Sam Gaddis (07:22)
That's interesting. I hadn't tried that. I'll have to try that.
Matthew Hall (07:26)
⁓ All right, next up.
News story here isn't really so much of a news story as it is a tweet from one of our favorite guys, Wade Foster from Zapier, CEO of Zapier. And he, in a few different podcasts, talked through this framework that they use for identifying and ⁓ choosing new AI initiatives and how to run them. So basically it's five steps. It's really simple. Pick one process close to revenue or your biggest cost, break it into three to five steps.
Automate the boring parts with a workflow. And there's an important distinction here of what's a workflow. So that's just classic deterministic automation stuff that Zapier has been great at for a decade. Four is insert an agent only where judgment is needed. So separate what's automation from what is agentic. The agent is doing something where it needs judgment or needs access to a tool that the automation doesn't have. And then five, keep a human in the loop, remove, remove as confidence grows. So.
I read this, it resonated with me because it's a lot like what we've been telling our clients to do. ⁓ And I've seen a few different versions of this. And what gets me excited is it seems like best practices are actually starting to coalesce, to codify that these, that AI workflow automation is not some unknowable that like you're going to have to go into the forest and figure out for the first time. So my question to you then Sam is what are these best practices, the commonalities you're seeing among companies that are getting this right?
Sam Gaddis (08:57)
Yeah, this is what I was thinking of when I said earlier that it feels like things are starting to gel. Well, this in the Wharton study. To me, what this reflects is a growing recognition that the way to make these tools work best is to break the process down into its most atomic level. By atomizing whatever the business workflow is, it becomes essentially obvious what you can.
do with AI and what requires a human in the loop. It's actually, it's old school. It's like franchising with the E-Myth is the book that I think of. Can you break up the, did you ever read that book? It was all about how to take any business and essentially ⁓ convert it into a franchise model, which is to say, create protocols for every critical business workflow, which is to say, write it down.
And that's a huge part of what we do with clients. spend so much more time trying to understand what's happening at an atomic level ⁓ than I do coding. And the other thing that this makes me really think of is
When you break all that out, there are most often things that you can automate with deterministic workflows. I would say 90 % of it can often be automated with deterministic workflows and 10 % get an LLM. And then actually there's human in the loop to some parts of that as well. So I think a thing that this focus on agents sometimes misses is the best work that we do.
often has no AI in it or has a very light AI touch. But the reason that it's still really interesting to do is because AI coding enables us to do these things that were previously not possible, just because it's really fast.
Matthew Hall (10:58)
Yep. Yep. My, my cherry on top of that is the thing I like best about this is so pick the workflow and it actually matters. So actually prioritize meaningfully in terms of the workflows that you want to do automate as much as you can. Cause the automation part's easy. The automation is the deterministic stuff and then pick one use of AI to start with like a single use of an agent within that workflow. I know most companies, once they've done the good work of, you know, documenting the whole workflow.
Try to stick AI in all five steps of the workflow and every single thing they do. Have one mega agent that can do everything. I think, you know, automate what you can that's repeatable and then have a single agent do something, make that great and then move on. Do not pass go until the workflow is better with just automation and maybe in a single agent doing something before you're layering on a bunch of
Sam Gaddis (11:49)
Yeah, or make it readily apparent within a week whether this is just not something that AI is good at and kill it. That's also a ⁓ highly valuable thing to figure out.
Matthew Hall (11:54)
Right.
All right, next up, this was a great article that we will put in our show notes. I think everybody should read in Creator Economy, Behind the Craft, I think is the name of the blog. And it's just about the state of AI coding assistance, that there are so many coding assistants and it seems like the leader changes weekly, not just in terms of the frontier model. So, you know, the new Sonnet 4.5 or the new Codex or whatever tick the lead.
But then you've got these, you know, the actual technology, the software for coding. So here's a list of some of them right now. You've got cursor, you've got cloud code, you've got warp, Klein, Devin, factory, Bolt, Replet, V zero, lovable, Figma, base 44. They're popping up constantly. And one fascinating thing here is that GitHub copilot, even though I don't know anybody that uses it in sort of the vibe coding community.
is still dominating in terms of usage and money because they have distribution, because they've got all the relationships with enterprise software. And so it's just this nice reminder that we live in a different world than most enterprise employees here who use the tools that their system gives to them that they have approval for. And I'm sure there are shadow ⁓ usage of a lot of these other tools as well, but it's just hard. It's a really difficult task for a procurement organization or an IT organization about which horses to back.
in the AI assisted coding space right now. My question to you though, Sam, if we acknowledge that, know, kind of the ideal scenario is that everybody has access to as many trials and tests that they want and they can change weekly, which is what we do. We are very lucky in that, have the max plans for everything and can change constantly. What would a mid to large size enterprise IT or how would you consult them? What tools should they stick with and what should they potentially experiment with?
Sam Gaddis (13:52)
I would answer this in two ways. think you can answer it for the organization writ large. And for that, you do have to settle on one tool because it's very difficult to either stomach the license fees or just manage the deployment of different types of software that do the same thing. ⁓ So I think, yeah, maybe cursor if you're talking about coding agents or general, even using that.
as a generalized AI prototyping tool is great, but.
I actually think that the right approach for this is to find that operator engineer that we talk about all the time, that business person who is passionate enough about AI to learn how to code and code and vibe code well, and figure out a way to give them carte blanche to use whatever tool that they need. ⁓ Give them
a license to try different things and constantly switch if they desire to do so and have that one person be the vanguard. And maybe the objective is to ultimately settle on a tool. But for me, this, have to use Claude code and codex. We can't do what we do effectively without using both of those tools because one will get nerfed one day and you have to use the other one. And you also, it's, it's much better if you can play them off of each other.
So I can't imagine doing something without that. That's what I would say. The other thing that I'll say about this is it is fascinating that people are using GitHub Copilot. I've actually run into this with our clients some, and I guess we're so in it that we don't realize that distribution overpowers the fact that it's terrible. At least that's my understanding of it, is terrible compared to something like clogged code.
Matthew Hall (15:46)
Yeah.
My advice to individuals would be try, see if you can't make the Claude code $200 plan create surplus value for you, kind of no matter who you are. Because I think once you figure out the value there, it's worth so much more than that, to me personally, and I think to pretty much everybody that uses it constantly. And it makes you so much more productive and useful.
I think the boring marketer is one of these tweets. He uses cloud code and cursor and that's pretty much it. That's pretty much the same for me too. Those are the two I use. I like being able to switch in cursor to other models when I want to and I have credits and I I like being able to, if I'm looking at a larger code base, I've got better version control and I've got better editing and approval flows and all those sorts of things. But I think that's the combo. ⁓ I think you can, you pretty.
You grow out of lovable in Replet and ⁓ in Ramp pretty quickly. You know, I think those are good places to start. All right. Yeah.
Sam Gaddis (16:55)
Yeah, I use cursor for
Git management and you know, when I need to actually get into the code, but it's that plus cloud code and Codex.
Matthew Hall (17:05)
All right, next up. ⁓
This is another great tweet by a person I've never met before named Aaron Budman. And he talks about how people talk about AI in terms of, you know, improving the work they're already doing. And yes, it is effective marginally at making experts better at what they do, you know, 10 to 15 % faster, you know, more complete in terms of something you're already an expert at. But what people don't talk enough about is all of the things that
AI can get people to try, net new, new work for people, ⁓ that they never would have had the confidence to go down if it wasn't the fact, it wasn't for the fact that AI can sort of get you to mediocre in any skill or any knowledge, you know, under the sun and how incredibly valuable that is, you know, for individuals and businesses. This is something, you know, I know you believe strongly in, he uses it in terms of like.
It's not even a percentage increase. It's infinity percent better because it's something you never would have done before. It was null and now it's something. And so I want to ask you, what are the top three, let's call it, ⁓ examples of newly viable work? It's the category we say here, of work that you wouldn't have done before AI that now you are doing with AI.
Sam Gaddis (18:32)
Okay. The first thing that comes to mind is this podcast itself. I love sitting here talking to you. I hate writing YouTube descriptions or LinkedIn posts for this. I just could not care less about doing that kind of work. It's so tedious. It's so boring. And so genuinely there's no chance that I would participate in this podcast if we had to actually write that stuff. ⁓ It makes outsourcing that much easier too.
And it also helps our preparation for this. mean, would you vibe code this tool that has the countdown and everything? We wouldn't do that. Another example is, ⁓ as you know, we are restructuring our entity for run point and doing the tax efficient planning for that before we make any decisions, doing the, even a lot of the paperwork for the operating agreements.
Matthew Hall (19:10)
Absolutely.
Sam Gaddis (19:32)
I'm not going to talk to a lawyer. I'm not going to talk to a CPA about this. We're just going to do it. And you know, we would have, we would actually have to do that, but it would probably take us a year to finally get to it and it would cost thousands of dollars. And frankly, it would probably be done worse. So those are the kinds of things where it just, I don't even think twice about diving into something that I don't have expertise in now. Like,
tax efficiency. I'm doing my will right now. I'm about to get married and so I'm doing my will and it's just yeah I can do a will no problem that kind of stuff.
Matthew Hall (20:09)
Exactly. That's the,
I just don't say no to stuff anymore with clients. You know, it's like, cause it's just going to take so little time to get something for them to react to, you know, and it used to be, feel like I used to really prioritize, you know, I only going to focus on the important things in my job and I'm going to try to say no more. But now it's like, Hey, if we need to get the ball rolling on something, hand. Yeah. I'm going to get it done. I'll get something up in front of you tomorrow because I can use AI to do it. Even if it's something I don't know anything about. Yep.
Alright.
Last question here, fascinating story ⁓ from Bain originally, I think, reported on this. so what happened was a company was in the process of acquiring a software company. And in the two weeks before the deal got done, the acquiring company had somebody there who rebuilt the primary software of the acquisition target.
in two weeks, showed it to the sponsors for this acquisition and they called off the deal completely. You know, kind of proving that what they were buying, the software they were buying was not nearly as valuable as what they were marking it as if it could be rebuilt by somebody in two weeks. So apparently, yeah, and it worked better. Now, you know, I think we can... ⁓
Sam Gaddis (21:28)
And it worked better, by the way.
Matthew Hall (21:36)
I'm sure there will be people that nitpick this, right? That's like, yeah, I'm sure it worked for one, know, that's not actually scalable, all sorts of stuff, all that kind of stuff. But it's a stark reminder that the value of software is shrinking, you know, it is approaching, it is on a steep clip, right? And I think when we talk to companies about who are in kind of the investment space,
How should companies think about valuing acquisition targets these days if software is approached, the value of software is falling off the cliff? What are the most and what are the valuable things that people should care about? How are the dynamics of those things changed because of AI?
Sam Gaddis (22:17)
I mean, it's changing rapidly and it's changing exponentially. The moats now have very little to do with software competency, have very little to do with team size, and are almost entirely centered around distribution. ⁓ Probably in some cases, ⁓ regulatory capture if you've got existing contracts there. And then data, obviously, if you've got a big data set.
These are the things that are valuable, that make a company valuable. And if you have distribution, if you have data, there are probably dozens of revenue streams buried within that, that are now enabled because of AI. If you don't have distribution, you have nothing. I did ⁓ this talk last week at Capital Factory and I had so many people come up to me afterwards and say, I'm building this thing and it's a product and it's in...
whatever category, and I'm really excited about it because the product's really good. And my advice was ⁓ as politely as I could and as kindly as I could to just say that it doesn't matter if you don't have distribution. There's no world in which I would go build a product without having the distribution done first. ⁓ What, you were going to say something?
Matthew Hall (23:41)
How do you build distribution from nothing?
Sam Gaddis (23:45)
Well, these days, I mean, it depends on the category that you're in, but it's brand. It's a lot of these products, indie developers trying to build things, launch on Twitter because that person already had 50,000 followers. And, you know, it's different for every network. It's different for every industry. through partnerships, companies that do this through partnerships, that makes a ton of sense. This is the most fascinating story I saw in the last week because we live this every day. You we're constantly building things in two weeks.
Really, it's about, like the thing that I always say is, companies are not fully internalizing the fact that you can build now in two weeks with one person what used to take six months and 20 people. And that's not true for everything, but it is true for a lot of things. People can't internalize this because especially kind of like mid-level management can't internalize this because it has huge implications for
team structures, cost structures, the future of development. It's scary. I think development's gonna be fine. I think developers are gonna be fine, especially the most competent developers. But it's a scary thing to think about for sure. But VPs and executives are starting to really get this finally. It's a hard mental shift though.
Matthew Hall (25:09)
Yeah, I think it's sort of, it's almost the emperor's clothes sort of thing. ⁓ I think there's been a myth for a long time that the best product software companies are the winners. I think there's a lot of examples out there of massive SaaS companies who are not best in class in terms of the actual product experience, but they've figured out the other parts of the ecosystem.
the partner models, the distribution, the lock-ins of their contracts, the novelty about the pricing strategy. There's a million other things that kind of work for a business to actually work. The actual software is less valuable than it once was, but it's never, very seldomly the thing that is unreplicable, unless you're talking about something truly, truly novel, which is just not, so much of business has never really that. ⁓
Sam Gaddis (26:04)
Yeah,
I mean, think about the new client that we're talking to right now. How do you value SaaS companies when this is the case? The client we're talking to right now has a business that requires a big SaaS product. Let's call it a CRM. And a large CRM company quoted them $500,000 to manage their CRM.
Who in their right mind would pay $500,000 for a CRM that has basic CRM functions that you can, where you can deploy an open source tool like ERP Next and put some, one or two people on that a year and have it do exactly what you want. Like how, how do you value a SaaS company when that's the reality? I don't know.
Matthew Hall (27:00)
But I would not advise a large multinational company to roll their own CRM and roll it all out right now or to switch to the latest thing. Because there's just so much decades of data and experience and customization that has gone into their specific use case. And the switching costs are going to be massive. There's a lot of lock in there. Everybody knows how to use the old thing, all that kind of stuff. But if you're a new business, my God.
If you're a new business or you're under a couple hundred people or the switching costs aren't that high, yeah, I don't know why you would go with any of the grandfathered SaaS companies for anything.
Sam Gaddis (27:35)
Yeah, the only thing is the switching costs are high, but the license fees are extremely high too. These companies are paying millions of dollars. And so yes, we got to go, we got to spend a year documenting the workflows and building training materials. Sure. But that's one person, you know, put a team of five people on it. It's still dwarfed by the license fees.
Matthew Hall (27:52)
Yep. Yep.
I'm sorry, Sam, I have to cut you off. We're over time. So for every second that you went over, you owe me $1,000. That's the way this game worked. So, ⁓ yep, sounds fair. I thought this was great. Thanks for the time today, ⁓ Like always, we'd love more feedback from folks. If you've got new news stories we should think about, if you've got questions, if you've got feedback in the format, let us know.
Sam Gaddis (28:03)
Okay, sounds fair.
All right, we'll see you next time.
Matthew Hall (28:22)
Thanks, Tim. Bye.