WEBVTT
00:00:00.959 --> 00:00:07.839
If you're still asking whether to buy or build AI, yeah, you're already behind.
00:00:08.080 --> 00:00:30.399
The question was a hot topic and made a ton of sense in the early days of generative AI in 2023 when rag pipelines and AI chatbots were cutting edge, and your company had to make a choice to either build something fairly complex on new technology or try to take business advantage of a consumer chatbot.
00:00:30.640 --> 00:00:36.960
But the buy versus build question doesn't make a ton of sense in 2026.
00:00:37.200 --> 00:00:47.600
That's because AI agents take actions inside your business software, compile your data while you sleep, and complete real workflows and outputs without much input from you.
00:00:47.840 --> 00:00:51.920
So your company's AI decision is no longer a single choice.
00:00:52.159 --> 00:00:53.840
It's actually at least four.
00:00:54.079 --> 00:00:57.920
That's because AI lives in at least four different layers of your company.
00:00:58.079 --> 00:01:04.959
I mean, the model itself, the workflows it runs, your data in context, and the business software it plugs into.
00:01:05.120 --> 00:01:07.680
And for each layer, you have four choices.
00:01:07.840 --> 00:01:16.719
You could build it yourself, buy it from a vendor, partner with someone who has the expertise, or wait until the category settles down a bit.
00:01:17.040 --> 00:01:21.120
But get one wrong and you wind up being in technical debt.
00:01:21.359 --> 00:01:28.159
Get another wrong and you lock yourself into a vendor, you can't escape or miss out on a new line of revenue that your competitors all seized.
00:01:28.319 --> 00:01:29.840
It's easier said than done.
00:01:30.000 --> 00:01:36.560
Yet I think this conversation is one of the most basic conversations being overlooked in boardrooms everywhere.
00:01:36.719 --> 00:01:42.400
That's why today we're walking through every layer, every choice, and exactly when to make each one.
00:01:42.640 --> 00:01:46.799
So you can stop guessing and start owning the right pieces.
00:01:47.040 --> 00:01:52.879
So this is everyday I, everyday AI, and this is part of our start here series.
00:01:53.040 --> 00:01:56.480
But before we go any further, let me just paint the big picture.
00:01:56.799 --> 00:02:00.480
So AI is different than it was.
00:02:00.560 --> 00:02:13.840
And I think back in the you know 2023, 2024, when companies were still deciding if AI was for them, the decision really boiled down to okay, are we going to use a consumer chatbot?
00:02:13.919 --> 00:02:20.400
Right back then it was, you know, BARD, the early days of Claude, you know, Copilot and Chat GPT.
00:02:20.479 --> 00:02:22.080
There was no business context.
00:02:22.240 --> 00:02:31.840
So you either had to say, okay, well, we're gonna try this thing and maybe risk it, or we're gonna build a pretty expensive uh, you know, custom rag setup, right?
00:02:31.919 --> 00:02:35.039
We're gonna build this rag pipeline and bring in all of our data.
00:02:35.199 --> 00:02:52.000
And uh, a lot of times that was a very expensive, long and tedious process, but this is all shifted because AI now, off the shelf, AI is made for businesses, and it's not just giving answers via a chat bot, it's executing real work inside of the actual software that you use.
00:02:52.159 --> 00:02:58.560
So um, new McKinsey study actually said that 88% of organizations regularly use AI in business.
00:02:58.800 --> 00:03:02.400
I'm always side note, I'm always like confused by these stats.
00:03:02.479 --> 00:03:04.960
I'm like, what are the other 12% people doing?
00:03:05.199 --> 00:03:23.199
Anyways, uh, you know, but this is a big shift because it's completely changed the narrative versus uh the you know traditional buy versus build, uh, which has always been a cornerstone of any tech innovation inside of any big companies.
00:03:23.439 --> 00:03:26.800
So now it's not just two choices.
00:03:26.960 --> 00:03:33.840
You have to decide when you're going to own, when you're gonna rent, partner with someone, or if you're just gonna wait.
00:03:34.000 --> 00:03:36.800
So on today's show, stick around and here's what you're gonna learn.
00:03:36.879 --> 00:03:42.719
You're gonna know why the 2023 builder by question fundamentally broke uh in the 2026 economy.
00:03:42.879 --> 00:03:47.439
Uh, you're gonna know the four AI stack layers you should own, rent, partner on, or delay.
00:03:47.680 --> 00:03:51.599
Uh, you're gonna know how to make the right call on each layer without those expensive mistakes.
00:03:51.759 --> 00:03:59.680
And I'm gonna lay out at the very end the three-week sprint plan to understand the build by partner or wait.
00:03:59.919 --> 00:04:02.719
All right, welcome to everyday AI.
00:04:02.800 --> 00:04:04.800
And this is the Start Here series.
00:04:04.960 --> 00:04:10.719
This is your essential podcast series to both learn the AI basics and to double down on your knowledge.
00:04:10.879 --> 00:04:20.480
That's because after literally 750 plus episodes of everyday AI, I never had a good answer when someone was like, I'm new here, where do I start?
00:04:20.560 --> 00:04:22.319
Well, you start with the start here series.
00:04:22.480 --> 00:04:25.600
I think these are best if you go about them in order.
00:04:25.680 --> 00:04:29.199
Uh, so make sure you go to starthirseries.com and you can do exactly that.
00:04:29.279 --> 00:04:32.879
That's going to give you free access uh to our inner circle community.
00:04:32.959 --> 00:04:46.879
And then you can go click on the start here series um space there and go listen, watch, read every single uh Start Here series episode in order, as well as a Spotify playlist that we keep updated there.
00:04:47.439 --> 00:04:54.240
So this is our uh series in the volume number 24 in the Start Here series.
00:04:54.319 --> 00:04:59.360
Um, or sorry, that was last episode was uh Start Here series, volume 24.
00:04:59.439 --> 00:05:07.839
Uh so we talked on that one open source AI 101, why local models, cheap APIs, and AI agents change everything.
00:05:08.079 --> 00:05:12.720
And today uh we are going over volume 25 of the start here series.
00:05:12.800 --> 00:05:15.120
So, yeah, it's best if you listen to these in order.
00:05:15.199 --> 00:05:19.680
You know, you can knock them out in a long road trip, uh, long weekend, something like that.
00:05:19.759 --> 00:05:26.160
But we're going over build, buy, partner, or wait the four-layer AI stack decision framework for 2026.
00:05:26.319 --> 00:05:32.160
All right, so why are we tackling this?
00:05:33.279 --> 00:05:48.319
As crazy as it sounds, I I talked to a pretty big company uh a couple of weeks ago, and I remembered talking with them originally in 2024, and they weren't investing a lot of time, money, and resources into the build side.
00:05:48.480 --> 00:05:52.240
And at the time, I nicely said, I don't think this is a good idea.
00:05:52.399 --> 00:05:55.120
Uh, the space is moving too quick, right?
00:05:55.199 --> 00:06:04.879
Because essentially you had ChatGPT come out in November 2022, and then eventually, as it got updated, as it got better, there started to get some business momentum, right?
00:06:04.959 --> 00:06:05.600
It was kind of weird.
00:06:05.680 --> 00:06:07.439
It was kind of like social media, right?
00:06:07.519 --> 00:06:11.040
Before Facebook had pages and it was just profiles, right?
00:06:11.120 --> 00:06:18.000
A lot of businesses were like, wait, should we just, you know, kind of go in the gray area here and uh, you know, create a profile, right?
00:06:18.079 --> 00:06:19.600
Because we can connect with all these people.
00:06:19.759 --> 00:06:26.399
And I think that the very early days of like late 2022 and early 2023 were kind of like this for businesses, right?
00:06:26.480 --> 00:06:45.120
They saw these tools, and at first it was just more of you know shadow IT and you know, employees were using it, and you know, a lot of decision makers didn't know what was going on, but eventually, right, in 2023 and 2024, businesses had a decision to make because at the time, large language models were not made for business, right?
00:06:45.199 --> 00:06:54.079
Even though businesses were using them and the companies were starting to wake up to that to be like, wait, we can make a ton of money instead of selling one seat for$20.
00:06:54.639 --> 00:06:59.600
We can send one, we can sell 1,000 seats, but the infrastructure was not there.
00:06:59.759 --> 00:07:02.000
Um now it is, right?
00:07:02.079 --> 00:07:07.920
If you look at uh one of the um benchmarks I love most is GDP Val.
00:07:08.399 --> 00:07:09.279
It's very simple.
00:07:09.439 --> 00:07:15.759
It's a model's ability to create economically valuable work, going head to head against a domain expert, right?
00:07:15.839 --> 00:07:23.680
And where uh you know essentially the model and the domain, the humans get the exact same information and they have to produce an artifact.
00:07:23.759 --> 00:07:26.240
And then judges will look at those and say which one's better.
00:07:26.399 --> 00:07:35.360
So, right now, today's best large language model, uh, GPT 5.5, uh, can tie or do better than 85% of experts, right?
00:07:35.439 --> 00:07:39.839
So that's a big shift here, and that also changes those decisions, right?
00:07:39.920 --> 00:07:48.399
That's why the decision to build or buy made sense, and it was a binary decision back in 2023 and 2024, but not so much anymore.
00:07:48.720 --> 00:08:01.040
Because not only has agentic AI ushered in a completely new era of what's possible, it's actually made the decision much more difficult.
00:08:01.519 --> 00:08:08.639
Um, and that capability puts some serious uh, you know, software creation power into almost everyone's hands.
00:08:08.720 --> 00:08:15.600
And that's one of the hardest parts because even building something two years ago was incredibly difficult.
00:08:15.839 --> 00:08:24.000
Building something now, something that your team, your department, your entire organization can use is actually straightforward.
00:08:24.160 --> 00:08:30.240
It may not be simple in theory, but it sometimes looks straightforward enough that you're like, oh, let's do this.
00:08:30.399 --> 00:08:34.480
We don't even need to partner or wait because we can build this ourselves.
00:08:35.759 --> 00:08:38.320
But everything's kind of changed this month.
00:08:38.639 --> 00:08:41.039
Uh, you know, at least over the past six weeks.
00:08:41.200 --> 00:08:44.480
There's been this big shift toward uh headless AI.
00:08:44.559 --> 00:08:48.399
We actually covered that in a recent start here um episode.
00:08:48.480 --> 00:08:51.679
So make sure you go go back a couple of episodes and listen to that one.
00:08:51.759 --> 00:08:55.919
But essentially, things have really changed very recently.
00:08:56.080 --> 00:09:00.080
So yeah, uh, right now, you know, it's May 2026.
00:09:00.159 --> 00:09:04.240
Maybe you're listening to this in June or July or December, right?
00:09:04.399 --> 00:09:11.039
So this could be not as timely, but there's been some some recent happenings that I think are worth uh talking about.
00:09:11.200 --> 00:09:14.000
So one would be kind of this incumbent shift.
00:09:14.159 --> 00:09:17.039
So you've had these large enterprises such as SAP.
00:09:17.200 --> 00:09:23.039
They've launched an autonomous enterprise where they have 50 plus domain assistants orchestrating 200 plus specialized agents.
00:09:23.200 --> 00:09:25.919
Uh, you have now the lab as consultancy.
00:09:26.080 --> 00:09:30.639
So uh I would say Anthropic and OpenAI made big plays, right?
00:09:30.720 --> 00:09:37.440
Uh OpenAI a little bit more in terms of investment with deployed co uh, but you know, and Google as well.
00:09:37.600 --> 00:09:47.120
So essentially these big AI companies are now putting out uh, you know, front-end, I think they're calling them, you know, front-end uh engineers uh into actual companies.
00:09:47.200 --> 00:09:50.159
And they're creating also consulting companies at the same time.
00:09:50.320 --> 00:10:01.120
Uh so even what your company, especially if you're at an enterprise company, there might have been things where you were trying to build uh or buy it, but it might make sense now to partner, right?
00:10:01.279 --> 00:10:14.799
If you can have, as an example, a consulting company that is built or owned by Enthropic or OpenAI or Google come into your organization, that might completely change what you thought was, oh, well, we're gonna build this or we're gonna wait.
00:10:14.960 --> 00:10:16.480
Well, maybe not anymore.
00:10:16.799 --> 00:10:21.039
And now there is the down market squeeze, uh, right.
00:10:21.120 --> 00:10:28.720
And I think that this will become um even more apparent, probably uh toward the beginning of quarter three.
00:10:28.799 --> 00:10:34.000
Uh, I think Anthropic was first to the punch here, but Anthropic launched Claude for small businesses.
00:10:34.159 --> 00:10:39.360
Uh, this was late last week with 15 pre-packaged agentic workflows.
00:10:39.519 --> 00:10:41.440
I think this is going to become the standard.
00:10:41.600 --> 00:10:48.159
I think you're going to see it across Microsoft Copilot, uh, Google Gemini, and uh across OpenAI as well.
00:10:48.399 --> 00:10:50.960
And you might be wondering, well, what the heck is that?
00:10:51.039 --> 00:10:52.879
And well, why does it matter?
00:10:53.200 --> 00:10:56.480
So uh let's take finance as an example.
00:10:56.639 --> 00:11:02.720
Uh, you might have had to buy a completely, you know, separate finance product three to six months ago.
00:11:02.799 --> 00:11:13.039
Uh, maybe you might have hired uh someone to come in, but maybe if you're a smaller organization, uh, you know, I don't know, an SMB with you know a thousand employees, uh, something like that.
00:11:13.200 --> 00:11:20.720
So um maybe you didn't have the seven-figure budget uh to go out and really build something two or three years ago.
00:11:20.799 --> 00:11:27.279
Uh, but now you might have something like a pre-packaged uh skill inside of anthropic.
00:11:27.519 --> 00:11:35.039
That's one area I think that they've uh just really have owned that vertical of having these high value agents, right?
00:11:35.200 --> 00:11:44.000
Which are essentially a series of pre-packaged uh skills and plugins uh that are essentially fine-tuned for specific purposes, right?
00:11:44.080 --> 00:11:53.840
So, as an example, small businesses, or you know, if you're a medium-sized enterprise uh trying to figure out the best, you know, finance AI to use.
00:11:53.919 --> 00:11:58.639
Maybe you were going through specific vendors that were not part of the big four, right?
00:11:58.720 --> 00:12:02.480
Which is uh Microsoft, Google, Anthropic, and Open AI.
00:12:02.639 --> 00:12:10.559
But maybe now you are looking at those uh because I think they are gonna start to uh verticalize and start to really essentially fine-tune, right?
00:12:10.720 --> 00:12:18.080
It's kind of like I had this prediction, uh, I think this was my 2023 predictions for 2024, maybe, uh, right.
00:12:18.159 --> 00:12:25.759
But I talked about uh eventually there's gonna be thousands or tens of thousands or even millions of small language models.
00:12:25.840 --> 00:12:42.960
And I think ultimately, at least now, what that's going to be is skills and in pre-packaged uh agent packs, essentially, that are models which are more or less fine-tuned via markdown files uh to very specifically carry out certain tasks.
00:12:43.120 --> 00:12:45.360
So the rules have all completely changed.
00:12:45.440 --> 00:12:49.039
So billing with AI also got a lot easier.
00:12:49.440 --> 00:13:00.480
And I think that's also really changed the conversation over the past couple of quarters because you know, before when it was buy, build, partner, or wait, it was hard to build, right?
00:13:00.639 --> 00:13:19.120
Unless you were a Fortune 500-esque level, uh, right, doing uh a couple billion dollars, 10, 20 billion dollars in revenue, it wasn't always feasible or possible via a talent uh per uh uh a talent perspective to build anything internally.
00:13:19.200 --> 00:13:20.639
And that's pretty much change.
00:13:20.879 --> 00:13:34.480
Uh, because uh you look at different stats, but you know, some stats say that uh up to 51%, um, so more than half of code that gets pushed out today is AI generated.
00:13:34.559 --> 00:13:41.440
Uh, but a recent retool study found that only 8% of AI builders use AI generated code without making changes.
00:13:41.679 --> 00:13:51.039
So a lot of companies, when they saw this, you know, this uh you know vibe coding wave actually get serious, uh right now we call it uh agenc engineering, right?
00:13:51.200 --> 00:13:54.320
It's like when vibe coding went to college and and graduated.
00:13:54.480 --> 00:13:59.200
Now we call it agentic engineering um or you know, agentic software engineering.
00:13:59.360 --> 00:14:05.200
Um, the problem is a lot of companies thought that they could own something, right?
00:14:05.519 --> 00:14:15.279
You know, uh anthropics claude code really uh I I I guess exploded this um this sector in late 2025 and early 2026.
00:14:15.360 --> 00:14:20.399
And a lot of companies that were early adopters thought that they could, well, they could build it and they could own it themselves.
00:14:20.559 --> 00:14:32.799
And then, well, they realized that if they didn't actually know what they were doing, and if they actually put these uh you know tools or systems into production, maybe they were, you know, dime smart dollar dumb.
00:14:32.960 --> 00:14:38.399
And they maybe solved one problem, but created a couple more uh with some expensive tech debt.
00:14:40.000 --> 00:14:44.399
So this is how we need to frame these decisions now.
00:14:44.639 --> 00:14:46.559
It's not a single choice, right?
00:14:46.720 --> 00:14:51.120
You don't just decide build by partner or wait for your entire organization.
00:14:51.440 --> 00:15:00.720
Because uh, if you are implementing AI from the front office uh to the back admin to every employee in between, there's a lot of different layers.
00:15:00.799 --> 00:15:01.600
So we talked about that.
00:15:01.679 --> 00:15:02.559
There's the model.
00:15:02.879 --> 00:15:04.480
What model are you using?
00:15:04.639 --> 00:15:06.399
The workflow, all right.
00:15:06.480 --> 00:15:15.679
Uh, and that's not just the it's not just the harness that a model uses, but it's how are you actually using that model in your day-to-day workflow?
00:15:15.919 --> 00:15:18.159
Then there's number three, your data.
00:15:18.799 --> 00:15:39.840
Okay, that's bringing your dynamic data in and having a sound data strategy, making sure that you have uh clean, structured data that is dynamic, and then your business software or connecting it uh both on the input side uh so you can grab data from that software that you need, but also to write it, uh, right.
00:15:40.000 --> 00:15:49.919
Because I think even nine months ago, most software integrations were one-way streets, they're bidirectional now, right?
00:15:50.000 --> 00:15:59.200
It is uh most, it's kind of almost the status quo for most, uh kind of if you're looking at it as simple as connectors or apps, right?
00:15:59.279 --> 00:16:04.159
Different uh different big four call them, different things, but they're bi-directional now, right?
00:16:04.240 --> 00:16:05.039
Read and write.
00:16:05.200 --> 00:16:14.799
It's no longer good enough to just be able to read your emails dynamically or to read your calendar and triage between your email and your calendar and your storage, right?
00:16:14.879 --> 00:16:21.919
Now you need to be able to write, you need to be able to create a uh a calendar out of your AI system, uh a calendar entry.
00:16:22.000 --> 00:16:25.679
You need to be able to create a new file um agentically, right?
00:16:25.840 --> 00:16:27.840
On demand, on a schedule, on a cron.
00:16:28.240 --> 00:16:35.600
Um, but that's not it's not happening um because it's getting more and more complex and the updates are coming out faster.
00:16:35.840 --> 00:16:38.720
So you need to really be intentional.
00:16:38.879 --> 00:16:50.320
So if you are in the boardroom, if you are helping make these decisions on AI implementation, this isn't one size fits all, right?
00:16:50.720 --> 00:16:53.360
You don't say we're gonna build this, we're gonna buy it, and that's it.
00:16:53.679 --> 00:16:56.159
You do have to look at it in those four separate layers.
00:16:57.840 --> 00:17:12.559
But obviously in the long run, you want to if you can, if you have the foundation, if you have the talent, if you have the data, you always want to skew toward owning if you can't, right?
00:17:13.119 --> 00:17:15.519
There's there's no doubt about that.
00:17:15.680 --> 00:17:19.680
But it's not always feasible or possible to own.
00:17:19.839 --> 00:17:23.599
But when we talk about owning or building something, right?
00:17:23.759 --> 00:17:35.119
So kind of using those two terms interchangeably, uh, you should be building for proprietary processes for internal context and workflows that are unique to how you operate, right?
00:17:35.200 --> 00:17:45.279
So you shouldn't be as a lot of companies were wasting a lot of money in 2023, uh, you know, trying to essentially build their own large language model, which go back and listen.
00:17:45.440 --> 00:17:48.160
I said all along, I'm like, that's not smart.
00:17:48.319 --> 00:17:58.160
Uh, you know, maybe you got some nice gains in, you know, in 2023, but then you were left with tech debt or you had to kill the project.
00:17:58.720 --> 00:18:06.240
So uh that same retool study found that 35% of AI builders replaced a purchase tool with a custom one.
00:18:06.480 --> 00:18:08.400
So this shift is happening, right?
00:18:08.559 --> 00:18:10.079
I was even talking about this recently.
00:18:10.240 --> 00:18:11.440
I had a lot of software.
00:18:11.519 --> 00:18:15.759
Uh you know, I'm a small business, but I just replaced it, right?
00:18:15.839 --> 00:18:22.400
Like if I'm using a tool and I'm like, ah, some things are good, but some things I don't like, I just build it, right?
00:18:22.559 --> 00:18:23.759
That's not gonna be the case.
00:18:23.839 --> 00:18:32.559
And I'm not saying that, you know, you're gonna have a bunch of SMBs, you know, being like, all right, I'm out, uh, Zoho, CRM, I'm building my own.
00:18:32.799 --> 00:18:37.759
It's not feasible or or realistic, although it is possible.
00:18:37.920 --> 00:18:48.799
Um, but I think you have to look at building those common workflows yourself, where you can actually um have a long-term competitive advantage.
00:18:48.880 --> 00:18:58.400
So when you're thinking about the things you should be building, when you are thinking about the things you should have ownership on, think about your proprietary processes, uh, that internal, right?
00:18:58.480 --> 00:19:01.680
I always call it first company, uh, first company data.
00:19:01.759 --> 00:19:15.200
Uh, so it's not just what shows up in the spreadsheets, it's what shows up in your subject matter experts' minds, what's in your uh your SOPs, uh, what's in your meeting notes.
00:19:15.359 --> 00:19:18.240
Those are the things that you want to build ownership around.
00:19:18.319 --> 00:19:18.799
Uh, right.
00:19:18.880 --> 00:19:37.039
And a lot of times you'll be porting those off to other places, but those processes, even something as simple, something as simple as having a meeting assistant and being able to take those transcripts and instantly share that knowledge with your organization, that shouldn't be locked up behind a certain vendor.
00:19:37.200 --> 00:19:38.960
That should be a process you can own.
00:19:39.119 --> 00:19:42.960
And yes, there are local models that are easy enough, right?
00:19:43.039 --> 00:19:45.599
I built this for myself, not terribly hard, right?
00:19:45.680 --> 00:19:47.759
So your organization should be looking at those.
00:19:47.839 --> 00:19:51.599
Those processes are ones that you should own, right?
00:19:51.839 --> 00:19:53.759
Your proprietary data.
00:19:54.079 --> 00:19:58.880
Um, and I have a little graphic here uh on my screen.
00:19:59.039 --> 00:20:01.920
You know, you should be building what compounds.
00:20:02.559 --> 00:20:12.240
So, not just the proprietary data, but something that can help you create a durable moat and something where lock-in would destroy the value, right?
00:20:12.319 --> 00:20:13.279
Think of something like this.
00:20:13.359 --> 00:20:24.160
Think of something like uh, you know, if you're trying to use, you know, Zoom and Fireflies AI uh and Google Meet all at the same time in your organization.
00:20:24.480 --> 00:20:25.920
Well, that's not good.
00:20:26.000 --> 00:20:42.240
You either, well, need to build a process at least where you can have complete ownership of that and it's not segmented and fragmented across these different buckets, or just build your own solution because that at the baseline, think of what compounds.
00:20:42.319 --> 00:20:54.400
And I think it's people's time, it's your subject matter expertise, and it is that IP, it is that first company uh data that is essentially, I think, the fuel for reasoning models.
00:20:54.480 --> 00:20:56.640
That's what you should be building around.
00:20:57.359 --> 00:21:03.039
Next, buy when the workflow is common and already in your software, uh, right.
00:21:03.279 --> 00:21:09.599
So you should buy common, regulated, audit-heavy workflows that already sit inside of your existing platforms.
00:21:09.759 --> 00:21:10.319
Uh, right.
00:21:10.400 --> 00:21:25.759
So this is obviously going to be different from sector to sector, uh, but outside of the big four, there's probably good reasons to buy certain software, you know, AI-powered software that fits within those workflows, right?
00:21:25.839 --> 00:21:38.079
And maybe that SAP um is a great uh example, maybe something if you're a startup, uh you know, on the entrepreneur scene, maybe it's something like Slack, right?
00:21:38.160 --> 00:21:49.440
Uh just because when you're talking about buying workflows inside of regulated or audit heavy um or existing platforms, it doesn't mean that you have to think these, you know, large enterprise, right?
00:21:49.519 --> 00:21:51.759
It could be something as simple as ClickUp.
00:21:51.839 --> 00:21:55.680
It could be something as you know, simple as Slack, right?
00:21:55.839 --> 00:21:59.839
So you could be doing a lot of these things via third party uh tools.
00:22:00.480 --> 00:22:25.359
But if or sorry, via the the big four providers, but if your workflows are so ingrained right in this enterprise, uh in certain enterprise uh ecosystems, well, maybe that's where you buy, maybe that's where you yeah, spend that extra money on you know the the the Slack AI, spend that extra money on the clickup AI, second braid, whatever they're calling it nowadays.
00:22:25.519 --> 00:22:26.079
The same thing.
00:22:26.240 --> 00:22:36.799
Maybe you spend that extra money on uh on the sales force, whatever it is, wherever your team's uh data is locked into your day-to-day processes.
00:22:37.519 --> 00:22:40.240
So when you think about uh, I'm kind of saying zone two.
00:22:40.400 --> 00:22:45.519
So zone one is build what compounds, zone two is buy what protects, right?
00:22:45.599 --> 00:22:52.960
So those workflows that are, like I said, common, regulated, mission critical, and or require that strict audit log.
00:22:53.119 --> 00:23:04.799
So yeah, if you're somewhere in uh financial uh something HIPAA that needs to be HIPAA compliant, maybe you can't get that specific workflow uh off the shelf from some of the big four.
00:23:04.880 --> 00:23:06.319
So buy what protects.
00:23:06.480 --> 00:23:11.359
Number three, partner when the cost of failure is too high to go alone.
00:23:11.519 --> 00:23:11.920
All right.
00:23:12.000 --> 00:23:20.640
So, you know, if this is something in cybersecurity, regulated decisions, physical operations, critical infrastructure, that's when you should be looking at partnering.
00:23:20.799 --> 00:23:21.119
Okay.
00:23:21.839 --> 00:23:27.119
A good partner also needs to share the execution, uh, execution risk.
00:23:27.519 --> 00:23:33.920
But some AI categories maybe are still a little bit too unstable uh to commit to.
00:23:34.160 --> 00:23:37.680
So uh partnering is important.
00:23:38.000 --> 00:23:44.720
So I'm not just saying, oh, like on the legal side, we need to go get, you know, Harvey or whatever, right?
00:23:44.880 --> 00:23:50.640
Or, you know, Claude did come out with some uh, you know, some impressive uh legal agents.
00:23:50.720 --> 00:23:57.039
Um, same thing with uh Microsoft Copilot, they just rolled out some support uh for some legal agents as well.
00:23:57.200 --> 00:24:21.599
But when you think about zone three here, partnering where failure is too expensive, if the workflow touches physical operations specifically, uh, right, or highly complex domain data that is extremely valuable, um maybe, you know, you this is kind of the like the thing I think of this um, you know, when you partner when failure is expensive.
00:24:21.839 --> 00:24:25.359
I kind of think of this as the on-prem versus off-prem, uh, right.
00:24:25.440 --> 00:24:33.680
So if there's certain things that you would just do on-prem, whether it's, you know, from a data storage uh or whether just, you know, physical operations, right?
00:24:33.839 --> 00:24:37.119
We can't move this operation, you know, somewhere else.
00:24:37.200 --> 00:24:39.680
Well, because it needs to be physically located here.
00:24:39.839 --> 00:24:44.319
Um, I think that's where partnering is a big reality.
00:24:44.559 --> 00:24:48.720
Um, so last but not least, wait.
00:24:49.039 --> 00:24:49.359
All right.
00:24:49.519 --> 00:25:00.559
And I will preface this by saying that waiting can be both the best decision that your company makes and the absolute worst decision that your company makes.
00:25:00.799 --> 00:25:01.920
Let me explain that.
00:25:02.559 --> 00:25:12.720
Again, not to keep beating the dead horse, all right, but I'm gonna go pick on the uh the initial uh you know, rag pipelines, right?
00:25:13.119 --> 00:25:17.440
Uh don't get me wrong, rag's not dead, right?
00:25:17.519 --> 00:25:19.519
It's just it's it's just changed, right?
00:25:19.680 --> 00:25:25.519
But if you go back to quarter one of January 2023, and probably quarter two as well, right?
00:25:26.319 --> 00:25:31.039
There was a lot of companies who were on the edge, right?
00:25:31.279 --> 00:25:34.240
Not edge devices, but on the cutting edge of AI.
00:25:34.400 --> 00:25:43.359
And they were in theory making good decisions by saying, hey, this large language model, this generative AI thing, we're gonna get on the train, right?
00:25:43.519 --> 00:26:01.359
We're gonna invest heavily instead of using a consumer chatbot, which at the time, right, you didn't have the ability to connect your, you know, in five clicks, you couldn't bring in your emails, you couldn't bring in your SharePoint, your OneDrive, your box, right?
00:26:01.519 --> 00:26:02.240
All these things.
00:26:02.319 --> 00:26:03.440
It was impossible.
00:26:03.599 --> 00:26:07.759
You couldn't, it wasn't an option in the off-the-shelf consumer chatbot.
00:26:07.920 --> 00:26:11.839
So, you know, a lot of bigger companies had to make this investment.
00:26:12.000 --> 00:26:18.319
And at the time, it might have seemed like a smart idea, even though go back and listen, I was saying don't do it.
00:26:18.559 --> 00:26:22.559
Um, this is one of those times when you could have waited.
00:26:22.640 --> 00:26:45.759
Uh, because even to do something like that correctly and to train your entire organization on taking advantage of okay, now all of a sudden we, you know, ragged up our uh data to, you know, I don't know at the time, what was it, GPT 3.5 or GPT-4 or uh the original Claude 2 or Claude 3, right?
00:26:45.920 --> 00:26:52.960
Um number one, the models weren't good enough to use in production, anyways, uh, without some great prompt engineering.
00:26:53.119 --> 00:27:04.240
Uh but number two, I think there's a big uh divide between non-agentic AI and agentic AI.
00:27:04.400 --> 00:27:04.960
Uh right.
00:27:05.200 --> 00:27:21.440
So when the you know the AI was essentially just super smart, next token completion, you know, the fact that these companies uh really wanted to go all in, it's brave, it's courageous.
00:27:21.519 --> 00:27:30.640
And I think you have to make some of those decisions, but sometimes the brave and courageous decision is to wait, especially now, knowing the pace and the rate of innovation.
00:27:30.799 --> 00:27:31.359
Uh, right.
00:27:31.519 --> 00:27:40.079
So, as an example, maybe you're just waiting for a certain industry-specific software uh that everyone in your industry uses.
00:27:40.160 --> 00:27:47.440
Let's say you're in construction and everyone uses, you know, something to you know create bid shells for RFPs, right?
00:27:47.599 --> 00:27:52.319
Um, and you're just waiting because there's two pieces of software everyone uses.
00:27:52.480 --> 00:27:55.839
So you're like, okay, do we wait or do we build?
00:27:56.640 --> 00:28:03.119
You have to understand the risk versus the reward versus the rate of innovation, right?
00:28:03.279 --> 00:28:12.960
So, as an example now, with not a lot of technical expertise, companies can spin up a model context protocol server.
00:28:13.279 --> 00:28:33.599
And even if there uh wasn't a bridge uh between enterprise software that is crucial for your day-to-day versus the AI operating system that your company or department uses, well, now with model context protocol, you can build that, you know, maybe in an hour, right?
00:28:34.079 --> 00:28:41.839
Uh, so this concept of waiting, there's no one size fits all, but you have to be able to wait when vendors change monthly.
00:28:42.000 --> 00:28:50.240
You have to be able to wait if the ROI is unclear, uh, or you might have to wait until the category is a little more stable.
00:28:50.480 --> 00:28:51.279
A good example.
00:28:51.359 --> 00:28:53.119
Yeah, here's another one, y'all.
00:28:53.359 --> 00:28:57.680
Not saying I was right, maybe I was wrong, but I'd say I was right right now.
00:28:57.920 --> 00:29:01.920
You notice, don't talk a lot about open claw on this show, right?
00:29:02.079 --> 00:29:08.480
I got a lot of emails, people being like, oh, Jordan, you could be making so much money, you know, teaching people open claw.
00:29:08.559 --> 00:29:09.039
And I said it.
00:29:09.119 --> 00:29:11.359
I'm like, hey, open claw is great, right?
00:29:11.680 --> 00:29:15.599
It is the most successful open source project of all time.
00:29:15.839 --> 00:29:24.400
Nvidia CEO Jensen Wong, right, said it's the best thing since sliced bread, isn't what he actually said, but he called it the best piece of open source software ever.
00:29:24.640 --> 00:29:30.079
Um, and a lot of companies and a lot of very visible leaders have invested into open claw.
00:29:30.160 --> 00:29:34.240
And I've been saying since the beginning, do I have an open claw?
00:29:34.400 --> 00:29:34.720
Sure.
00:29:34.880 --> 00:29:35.839
Do I use it?
00:29:36.000 --> 00:29:37.200
Not much, right?
00:29:37.359 --> 00:29:43.920
I use it enough to know how it works and to you know see the struggles and the successes other people are having.
00:29:44.160 --> 00:29:55.680
But I knew at the time, I'm like, this is one of those times when it's better to wait because I knew that the big four are eventually going to have their versions, uh, right.
00:29:55.759 --> 00:30:01.279
And I'm not talking like NVIDIA's version of OpenClaw Nemo Claw, where it's essentially open claw in a more secure environment.
00:30:01.440 --> 00:30:06.400
I'm saying no, they're gonna have ChatGPT is gonna have its own version of OpenClaw, right?
00:30:06.480 --> 00:30:12.319
Claw is gonna have its own version of OpenClaw, and that's obviously where all of the big players have been trending.
00:30:12.400 --> 00:30:25.359
So I don't know, is it worth it to get, I don't know, 500 employees on an open claw, and all of a sudden they set all these things up, but their heartbeat's not working, their sole MD files corrupted, cron jobs aren't crowning, right?
00:30:25.440 --> 00:30:30.240
Now all of a sudden, I don't know, by being on the cutting edge, maybe you were able to take advantage of it.
00:30:30.400 --> 00:30:37.839
Maybe you were able to squeeze uh you know, the juice while the lemons were were ripe and no one had lemonade and everyone was thirsty.
00:30:37.920 --> 00:30:50.640
Maybe you had that opportunity, but probably I would venture to say that 90% of you know enterprises that invested heavily in open claw, the learning curve is so steep with just large language models.
00:30:50.720 --> 00:30:51.200
Uh, right.
00:30:51.359 --> 00:30:54.880
I think this is one of those where, well, the ROI was kind of unclear.
00:30:55.039 --> 00:30:56.000
Yes, it's exciting.
00:30:56.079 --> 00:30:58.799
And that's one of the problems with uh with AI.
00:30:59.039 --> 00:31:00.480
Everything you see is exciting.
00:31:00.640 --> 00:31:02.480
Everything you see, it's like MAD FOMO.
00:31:02.640 --> 00:31:05.519
Oh my gosh, if we're not using this right away, we're gonna go out of business.
00:31:05.680 --> 00:31:07.279
That's not the right way to look at it.
00:31:07.440 --> 00:31:14.960
You have to understand what's possible today, what's coming next, and then where do we build, buy, partner, or wait.
00:31:15.039 --> 00:31:20.079
And sometimes the bold and courageous thing might be being first or it might be waiting.
00:31:20.240 --> 00:31:23.920
But there's no uh, you know, one size fits all answer on that.
00:31:24.079 --> 00:31:29.359
But waiting while competitors redesign their workflows, though, that's drift.
00:31:29.519 --> 00:31:29.920
All right.
00:31:30.079 --> 00:31:36.240
So if you see everyone else in your space is doing something, yeah, maybe your strategy is to zig while everyone else zags.
00:31:36.400 --> 00:31:45.200
But if you can see that there's a product market fit with a certain early AI technology in your space, maybe that's not the time you wait then, right?
00:31:45.279 --> 00:31:46.400
Maybe that's the time, you know.
00:31:46.480 --> 00:31:52.319
As an example for OpenClaw, there's probably great uh enterprise use cases for it until everything else caught up.
00:31:52.400 --> 00:31:57.759
And now, I mean, I don't know, Codex is essentially a way more stable OpenClaw, right?
00:31:57.920 --> 00:32:04.720
Obviously, uh, you know, uh OpenAI kind of aqua hired uh, you know, the the sole creator of OpenClaw.
00:32:04.880 --> 00:32:11.759
So we're seeing a lot of the best pieces of it get implemented in a much more secure, robust, and enterprise-friendly way.
00:32:11.920 --> 00:32:20.319
All right, so uh so this is kind of zone four, is the danger of wait because it's also the capabilities gap, uh, right?
00:32:20.400 --> 00:32:32.160
So you need to look at the choice to wait on a case-by-case basis, uh, because it's not only the capability of the technology, it's also the learning curve of the people who have to use it.
00:32:32.319 --> 00:32:34.319
So I think the open claw is a good example there.
00:32:34.480 --> 00:32:43.039
Now it's much easier to go inside of something like Codecs, which if your company uses ChatGPT, Codex is so easy to see and understand.
00:32:43.200 --> 00:32:53.519
Or, you know, the um the workspace agents inside of um, you know, ChatGPT business plans, which are essentially like simple little open claws, right?
00:32:53.599 --> 00:32:54.960
But these are kind of the same things.
00:32:55.039 --> 00:33:03.599
They're things you can, you know, message uh from your phone and can connect to all the data sources and you can run them on a schedule, right?
00:33:03.759 --> 00:33:05.680
But sometimes there is the danger of waiting.
00:33:05.759 --> 00:33:14.160
But you do have to look first at the capabilities gap uh and the uh the knowledge gap and the training gap, right?
00:33:14.319 --> 00:33:18.000
So is your company still in the chatbot era?
00:33:18.240 --> 00:33:22.799
If so, you have to understand that capability gap is going to be a steep hill to climb.
00:33:22.960 --> 00:33:46.799
But if you are using for the most part, if your company is using models agentically, if you're already there, maybe waiting doesn't make as much sense because it could only take you, you know, I don't know, a quick, you know, three-hour training and you know, weekly check-ins to uh kind of close the capability gap so your team can actually take advantage of the latest and greatest in AI.
00:33:46.960 --> 00:33:47.359
All right.
00:33:47.519 --> 00:33:53.359
So here's the four wrong calls before we get to the one, two, three week plan.
00:33:53.680 --> 00:34:01.279
So uh wrong call number one, being wrong in the build creates technical debt that your team might have to maintain forever, right?
00:34:01.599 --> 00:34:08.079
Yeah, I know there's still companies out there maintaining their uh their rag chat bot from 2023.
00:34:08.480 --> 00:34:09.840
Uh wrong buy.
00:34:10.000 --> 00:34:13.840
That can create vendor lock-in that you can't exit without a major cost, all right?
00:34:14.000 --> 00:34:16.079
Don't there's too much competition up there.
00:34:16.159 --> 00:34:19.599
Don't get locked into a contract for two or three years, right?
00:34:19.760 --> 00:34:22.559
These vendors are gonna sell you something like there's no other option.
00:34:22.639 --> 00:34:23.599
There's always option.
00:34:23.679 --> 00:34:24.000
All right.
00:34:24.159 --> 00:34:30.880
The wrong partner, number three there, that can create dependency on someone whose business is also changing.
00:34:31.039 --> 00:34:44.639
Uh, working with the right partner is sometimes trickier than the uh the buy versus build, uh, because you have to then understand a third party's uh complete front to back.
00:34:44.719 --> 00:34:52.480
But you know, your partners should be extremely niche uh and like we talked about earlier, you know, um in those certain areas.
00:34:52.559 --> 00:35:00.480
And then last but not least, the wrong weight decision that can create a capability gap that competitors can quietly close first.
00:35:00.639 --> 00:35:09.119
So here is your one, two, three week blueprint, all right, to go through and understand your build by partner or weight decisions.
00:35:09.280 --> 00:35:12.000
So week number one, you're gonna audit and score.
00:35:12.159 --> 00:35:15.280
Week number two, you're gonna execute the extremes or the edges.
00:35:15.440 --> 00:35:18.719
And then week number three, you're gonna you're gonna govern and hedge.
00:35:18.880 --> 00:35:19.679
Here's what I mean.
00:35:19.840 --> 00:35:24.239
Uh so uh do have this on a screen here, but I'm just gonna read it all off.
00:35:24.320 --> 00:35:28.400
But you can always go watch the video version on our website at your everydayai.com.
00:35:28.559 --> 00:35:30.000
So week one, audit and score.
00:35:30.079 --> 00:35:38.320
You're gonna pick 10 workflows, you're gonna score them on an ownership map, and then you're gonna name one accountable human uh owner for each.
00:35:38.400 --> 00:35:44.880
All right, that's where you always need to start before you even go down the four different layers and the build by partner weight.
00:35:45.039 --> 00:35:50.079
You have to have an accurate inventory of what AI is actually being used for.
00:35:50.239 --> 00:35:53.679
Uh all right, and then week two, you need to execute the extremes.
00:35:53.920 --> 00:36:00.239
So you need to build one low risk or high differentiation, uh high differentiation internal workflow.
00:36:00.400 --> 00:36:10.320
Then you need to buy one packaged workflow inside your existing, you know, ERP, CRM, whatever the case, and then kill projects with no ROI.
00:36:10.480 --> 00:36:21.440
So essentially, first you're gonna audit and score, then you're gonna fit uh you're gonna find a low-risk, high uh high value internal workflow, and you're essentially gonna A-B test it, right?
00:36:21.519 --> 00:36:25.280
And you're gonna say, okay, what does this look like if we build it ourselves?
00:36:25.440 --> 00:36:28.320
And what does it look like if we buy something else?
00:36:28.400 --> 00:36:32.559
You need to measure your human costs and your external costs for the buy side.
00:36:32.719 --> 00:36:38.480
And then last but not least, week three, uh, you need to compare the build versus buy performance.
00:36:38.639 --> 00:36:43.760
And then for the buy side, you need to renegotiate contracts for data export and model portability.
00:36:43.840 --> 00:36:46.639
Uh, because again, think of something like skills, right?
00:36:46.800 --> 00:36:48.880
Skills are modular, they're open source.
00:36:48.960 --> 00:36:52.639
You can take those, you know, as an example from vendor to vendor.
00:36:52.719 --> 00:36:58.159
So, you know, if you are doing something on the buy side, you need to make sure that number one, you're not locked in.
00:36:58.239 --> 00:37:09.519
But number two, that whatever um that you are investing, uh, you know, aside from the money, but your resources, your data, your education, you have to make sure that it's modular before you get in too deep.
00:37:09.679 --> 00:37:17.519
And then last but not least, in week three, make sure you establish a weekly review cycle in ongoing learning as well.
00:37:17.760 --> 00:37:18.400
All right.
00:37:18.800 --> 00:37:26.800
That is a wrap for the build by partner or wait, the four-layer AI stack decision framework for 2026.
00:37:26.960 --> 00:37:31.280
And y'all, this just like everything else, it all moves quickly.
00:37:31.360 --> 00:37:34.320
So, this is not a, you know, if you're listening to this, sorry.
00:37:34.400 --> 00:37:42.639
If somehow you're listening to this in 2027, you might just want to forget everything I said, uh, FYI, and instead just keep up with the daily podcast.
00:37:43.039 --> 00:37:51.920
Because, like everything, sometimes advice, sometimes the latest and greatest has a shelf like has a shelf life, just like the models that we all use.
00:37:52.000 --> 00:37:53.119
So I hope this was helpful.
00:37:53.280 --> 00:37:56.639
If so, do me a favor, go to starthirseries.com.
00:37:56.800 --> 00:37:59.519
That is going to get you free access to our community.
00:37:59.599 --> 00:38:01.440
That's the only way you can get in right now.
00:38:01.599 --> 00:38:16.800
And then make sure you go straight to the Start Here series and check out the Spotify playlist that has every single volume of the Start Here series ready for you to consume, whether it's in text, reading the email newsletter that went with it, watching the video, listening to the podcast.
00:38:16.880 --> 00:38:18.400
It's all there in one place.
00:38:18.480 --> 00:38:23.119
And you can also network and connect and ask questions with other people who are going through the same process.
00:38:23.280 --> 00:38:24.400
So, thank you for tuning in.
00:38:24.480 --> 00:38:27.440
I hope to see you back tomorrow and every day for more everyday AI.
00:38:27.519 --> 00:38:28.639
Thanks, y'all.
00:00:00.959 --> 00:00:07.839
If you're still asking whether to buy or build AI, yeah, you're already behind.
00:00:08.080 --> 00:00:30.399
The question was a hot topic and made a ton of sense in the early days of generative AI in 2023 when rag pipelines and AI chatbots were cutting edge, and your company had to make a choice to either build something fairly complex on new technology or try to take business advantage of a consumer chatbot.
00:00:30.640 --> 00:00:36.960
But the buy versus build question doesn't make a ton of sense in 2026.
00:00:37.200 --> 00:00:47.600
That's because AI agents take actions inside your business software, compile your data while you sleep, and complete real workflows and outputs without much input from you.
00:00:47.840 --> 00:00:51.920
So your company's AI decision is no longer a single choice.
00:00:52.159 --> 00:00:53.840
It's actually at least four.
00:00:54.079 --> 00:00:57.920
That's because AI lives in at least four different layers of your company.
00:00:58.079 --> 00:01:04.959
I mean, the model itself, the workflows it runs, your data in context, and the business software it plugs into.
00:01:05.120 --> 00:01:07.680
And for each layer, you have four choices.
00:01:07.840 --> 00:01:16.719
You could build it yourself, buy it from a vendor, partner with someone who has the expertise, or wait until the category settles down a bit.
00:01:17.040 --> 00:01:21.120
But get one wrong and you wind up being in technical debt.
00:01:21.359 --> 00:01:28.159
Get another wrong and you lock yourself into a vendor, you can't escape or miss out on a new line of revenue that your competitors all seized.
00:01:28.319 --> 00:01:29.840
It's easier said than done.
00:01:30.000 --> 00:01:36.560
Yet I think this conversation is one of the most basic conversations being overlooked in boardrooms everywhere.
00:01:36.719 --> 00:01:42.400
That's why today we're walking through every layer, every choice, and exactly when to make each one.
00:01:42.640 --> 00:01:46.799
So you can stop guessing and start owning the right pieces.
00:01:47.040 --> 00:01:52.879
So this is everyday I, everyday AI, and this is part of our start here series.
00:01:53.040 --> 00:01:56.480
But before we go any further, let me just paint the big picture.
00:01:56.799 --> 00:02:00.480
So AI is different than it was.
00:02:00.560 --> 00:02:13.840
And I think back in the you know 2023, 2024, when companies were still deciding if AI was for them, the decision really boiled down to okay, are we going to use a consumer chatbot?
00:02:13.919 --> 00:02:20.400
Right back then it was, you know, BARD, the early days of Claude, you know, Copilot and Chat GPT.
00:02:20.479 --> 00:02:22.080
There was no business context.
00:02:22.240 --> 00:02:31.840
So you either had to say, okay, well, we're gonna try this thing and maybe risk it, or we're gonna build a pretty expensive uh, you know, custom rag setup, right?
00:02:31.919 --> 00:02:35.039
We're gonna build this rag pipeline and bring in all of our data.
00:02:35.199 --> 00:02:52.000
And uh, a lot of times that was a very expensive, long and tedious process, but this is all shifted because AI now, off the shelf, AI is made for businesses, and it's not just giving answers via a chat bot, it's executing real work inside of the actual software that you use.
00:02:52.159 --> 00:02:58.560
So um, new McKinsey study actually said that 88% of organizations regularly use AI in business.
00:02:58.800 --> 00:03:02.400
I'm always side note, I'm always like confused by these stats.
00:03:02.479 --> 00:03:04.960
I'm like, what are the other 12% people doing?
00:03:05.199 --> 00:03:23.199
Anyways, uh, you know, but this is a big shift because it's completely changed the narrative versus uh the you know traditional buy versus build, uh, which has always been a cornerstone of any tech innovation inside of any big companies.
00:03:23.439 --> 00:03:26.800
So now it's not just two choices.
00:03:26.960 --> 00:03:33.840
You have to decide when you're going to own, when you're gonna rent, partner with someone, or if you're just gonna wait.
00:03:34.000 --> 00:03:36.800
So on today's show, stick around and here's what you're gonna learn.
00:03:36.879 --> 00:03:42.719
You're gonna know why the 2023 builder by question fundamentally broke uh in the 2026 economy.
00:03:42.879 --> 00:03:47.439
Uh, you're gonna know the four AI stack layers you should own, rent, partner on, or delay.
00:03:47.680 --> 00:03:51.599
Uh, you're gonna know how to make the right call on each layer without those expensive mistakes.
00:03:51.759 --> 00:03:59.680
And I'm gonna lay out at the very end the three-week sprint plan to understand the build by partner or wait.
00:03:59.919 --> 00:04:02.719
All right, welcome to everyday AI.
00:04:02.800 --> 00:04:04.800
And this is the Start Here series.
00:04:04.960 --> 00:04:10.719
This is your essential podcast series to both learn the AI basics and to double down on your knowledge.
00:04:10.879 --> 00:04:20.480
That's because after literally 750 plus episodes of everyday AI, I never had a good answer when someone was like, I'm new here, where do I start?
00:04:20.560 --> 00:04:22.319
Well, you start with the start here series.
00:04:22.480 --> 00:04:25.600
I think these are best if you go about them in order.
00:04:25.680 --> 00:04:29.199
Uh, so make sure you go to starthirseries.com and you can do exactly that.
00:04:29.279 --> 00:04:32.879
That's going to give you free access uh to our inner circle community.
00:04:32.959 --> 00:04:46.879
And then you can go click on the start here series um space there and go listen, watch, read every single uh Start Here series episode in order, as well as a Spotify playlist that we keep updated there.
00:04:47.439 --> 00:04:54.240
So this is our uh series in the volume number 24 in the Start Here series.
00:04:54.319 --> 00:04:59.360
Um, or sorry, that was last episode was uh Start Here series, volume 24.
00:04:59.439 --> 00:05:07.839
Uh so we talked on that one open source AI 101, why local models, cheap APIs, and AI agents change everything.
00:05:08.079 --> 00:05:12.720
And today uh we are going over volume 25 of the start here series.
00:05:12.800 --> 00:05:15.120
So, yeah, it's best if you listen to these in order.
00:05:15.199 --> 00:05:19.680
You know, you can knock them out in a long road trip, uh, long weekend, something like that.
00:05:19.759 --> 00:05:26.160
But we're going over build, buy, partner, or wait the four-layer AI stack decision framework for 2026.
00:05:26.319 --> 00:05:32.160
All right, so why are we tackling this?
00:05:33.279 --> 00:05:48.319
As crazy as it sounds, I I talked to a pretty big company uh a couple of weeks ago, and I remembered talking with them originally in 2024, and they weren't investing a lot of time, money, and resources into the build side.
00:05:48.480 --> 00:05:52.240
And at the time, I nicely said, I don't think this is a good idea.
00:05:52.399 --> 00:05:55.120
Uh, the space is moving too quick, right?
00:05:55.199 --> 00:06:04.879
Because essentially you had ChatGPT come out in November 2022, and then eventually, as it got updated, as it got better, there started to get some business momentum, right?
00:06:04.959 --> 00:06:05.600
It was kind of weird.
00:06:05.680 --> 00:06:07.439
It was kind of like social media, right?
00:06:07.519 --> 00:06:11.040
Before Facebook had pages and it was just profiles, right?
00:06:11.120 --> 00:06:18.000
A lot of businesses were like, wait, should we just, you know, kind of go in the gray area here and uh, you know, create a profile, right?
00:06:18.079 --> 00:06:19.600
Because we can connect with all these people.
00:06:19.759 --> 00:06:26.399
And I think that the very early days of like late 2022 and early 2023 were kind of like this for businesses, right?
00:06:26.480 --> 00:06:45.120
They saw these tools, and at first it was just more of you know shadow IT and you know, employees were using it, and you know, a lot of decision makers didn't know what was going on, but eventually, right, in 2023 and 2024, businesses had a decision to make because at the time, large language models were not made for business, right?
00:06:45.199 --> 00:06:54.079
Even though businesses were using them and the companies were starting to wake up to that to be like, wait, we can make a ton of money instead of selling one seat for$20.
00:06:54.639 --> 00:06:59.600
We can send one, we can sell 1,000 seats, but the infrastructure was not there.
00:06:59.759 --> 00:07:02.000
Um now it is, right?
00:07:02.079 --> 00:07:07.920
If you look at uh one of the um benchmarks I love most is GDP Val.
00:07:08.399 --> 00:07:09.279
It's very simple.
00:07:09.439 --> 00:07:15.759
It's a model's ability to create economically valuable work, going head to head against a domain expert, right?
00:07:15.839 --> 00:07:23.680
And where uh you know essentially the model and the domain, the humans get the exact same information and they have to produce an artifact.
00:07:23.759 --> 00:07:26.240
And then judges will look at those and say which one's better.
00:07:26.399 --> 00:07:35.360
So, right now, today's best large language model, uh, GPT 5.5, uh, can tie or do better than 85% of experts, right?
00:07:35.439 --> 00:07:39.839
So that's a big shift here, and that also changes those decisions, right?
00:07:39.920 --> 00:07:48.399
That's why the decision to build or buy made sense, and it was a binary decision back in 2023 and 2024, but not so much anymore.
00:07:48.720 --> 00:08:01.040
Because not only has agentic AI ushered in a completely new era of what's possible, it's actually made the decision much more difficult.
00:08:01.519 --> 00:08:08.639
Um, and that capability puts some serious uh, you know, software creation power into almost everyone's hands.
00:08:08.720 --> 00:08:15.600
And that's one of the hardest parts because even building something two years ago was incredibly difficult.
00:08:15.839 --> 00:08:24.000
Building something now, something that your team, your department, your entire organization can use is actually straightforward.
00:08:24.160 --> 00:08:30.240
It may not be simple in theory, but it sometimes looks straightforward enough that you're like, oh, let's do this.
00:08:30.399 --> 00:08:34.480
We don't even need to partner or wait because we can build this ourselves.
00:08:35.759 --> 00:08:38.320
But everything's kind of changed this month.
00:08:38.639 --> 00:08:41.039
Uh, you know, at least over the past six weeks.
00:08:41.200 --> 00:08:44.480
There's been this big shift toward uh headless AI.
00:08:44.559 --> 00:08:48.399
We actually covered that in a recent start here um episode.
00:08:48.480 --> 00:08:51.679
So make sure you go go back a couple of episodes and listen to that one.
00:08:51.759 --> 00:08:55.919
But essentially, things have really changed very recently.
00:08:56.080 --> 00:09:00.080
So yeah, uh, right now, you know, it's May 2026.
00:09:00.159 --> 00:09:04.240
Maybe you're listening to this in June or July or December, right?
00:09:04.399 --> 00:09:11.039
So this could be not as timely, but there's been some some recent happenings that I think are worth uh talking about.
00:09:11.200 --> 00:09:14.000
So one would be kind of this incumbent shift.
00:09:14.159 --> 00:09:17.039
So you've had these large enterprises such as SAP.
00:09:17.200 --> 00:09:23.039
They've launched an autonomous enterprise where they have 50 plus domain assistants orchestrating 200 plus specialized agents.
00:09:23.200 --> 00:09:25.919
Uh, you have now the lab as consultancy.
00:09:26.080 --> 00:09:30.639
So uh I would say Anthropic and OpenAI made big plays, right?
00:09:30.720 --> 00:09:37.440
Uh OpenAI a little bit more in terms of investment with deployed co uh, but you know, and Google as well.
00:09:37.600 --> 00:09:47.120
So essentially these big AI companies are now putting out uh, you know, front-end, I think they're calling them, you know, front-end uh engineers uh into actual companies.
00:09:47.200 --> 00:09:50.159
And they're creating also consulting companies at the same time.
00:09:50.320 --> 00:10:01.120
Uh so even what your company, especially if you're at an enterprise company, there might have been things where you were trying to build uh or buy it, but it might make sense now to partner, right?
00:10:01.279 --> 00:10:14.799
If you can have, as an example, a consulting company that is built or owned by Enthropic or OpenAI or Google come into your organization, that might completely change what you thought was, oh, well, we're gonna build this or we're gonna wait.
00:10:14.960 --> 00:10:16.480
Well, maybe not anymore.
00:10:16.799 --> 00:10:21.039
And now there is the down market squeeze, uh, right.
00:10:21.120 --> 00:10:28.720
And I think that this will become um even more apparent, probably uh toward the beginning of quarter three.
00:10:28.799 --> 00:10:34.000
Uh, I think Anthropic was first to the punch here, but Anthropic launched Claude for small businesses.
00:10:34.159 --> 00:10:39.360
Uh, this was late last week with 15 pre-packaged agentic workflows.
00:10:39.519 --> 00:10:41.440
I think this is going to become the standard.
00:10:41.600 --> 00:10:48.159
I think you're going to see it across Microsoft Copilot, uh, Google Gemini, and uh across OpenAI as well.
00:10:48.399 --> 00:10:50.960
And you might be wondering, well, what the heck is that?
00:10:51.039 --> 00:10:52.879
And well, why does it matter?
00:10:53.200 --> 00:10:56.480
So uh let's take finance as an example.
00:10:56.639 --> 00:11:02.720
Uh, you might have had to buy a completely, you know, separate finance product three to six months ago.
00:11:02.799 --> 00:11:13.039
Uh, maybe you might have hired uh someone to come in, but maybe if you're a smaller organization, uh, you know, I don't know, an SMB with you know a thousand employees, uh, something like that.
00:11:13.200 --> 00:11:20.720
So um maybe you didn't have the seven-figure budget uh to go out and really build something two or three years ago.
00:11:20.799 --> 00:11:27.279
Uh, but now you might have something like a pre-packaged uh skill inside of anthropic.
00:11:27.519 --> 00:11:35.039
That's one area I think that they've uh just really have owned that vertical of having these high value agents, right?
00:11:35.200 --> 00:11:44.000
Which are essentially a series of pre-packaged uh skills and plugins uh that are essentially fine-tuned for specific purposes, right?
00:11:44.080 --> 00:11:53.840
So, as an example, small businesses, or you know, if you're a medium-sized enterprise uh trying to figure out the best, you know, finance AI to use.
00:11:53.919 --> 00:11:58.639
Maybe you were going through specific vendors that were not part of the big four, right?
00:11:58.720 --> 00:12:02.480
Which is uh Microsoft, Google, Anthropic, and Open AI.
00:12:02.639 --> 00:12:10.559
But maybe now you are looking at those uh because I think they are gonna start to uh verticalize and start to really essentially fine-tune, right?
00:12:10.720 --> 00:12:18.080
It's kind of like I had this prediction, uh, I think this was my 2023 predictions for 2024, maybe, uh, right.
00:12:18.159 --> 00:12:25.759
But I talked about uh eventually there's gonna be thousands or tens of thousands or even millions of small language models.
00:12:25.840 --> 00:12:42.960
And I think ultimately, at least now, what that's going to be is skills and in pre-packaged uh agent packs, essentially, that are models which are more or less fine-tuned via markdown files uh to very specifically carry out certain tasks.
00:12:43.120 --> 00:12:45.360
So the rules have all completely changed.
00:12:45.440 --> 00:12:49.039
So billing with AI also got a lot easier.
00:12:49.440 --> 00:13:00.480
And I think that's also really changed the conversation over the past couple of quarters because you know, before when it was buy, build, partner, or wait, it was hard to build, right?
00:13:00.639 --> 00:13:19.120
Unless you were a Fortune 500-esque level, uh, right, doing uh a couple billion dollars, 10, 20 billion dollars in revenue, it wasn't always feasible or possible via a talent uh per uh uh a talent perspective to build anything internally.
00:13:19.200 --> 00:13:20.639
And that's pretty much change.
00:13:20.879 --> 00:13:34.480
Uh, because uh you look at different stats, but you know, some stats say that uh up to 51%, um, so more than half of code that gets pushed out today is AI generated.
00:13:34.559 --> 00:13:41.440
Uh, but a recent retool study found that only 8% of AI builders use AI generated code without making changes.
00:13:41.679 --> 00:13:51.039
So a lot of companies, when they saw this, you know, this uh you know vibe coding wave actually get serious, uh right now we call it uh agenc engineering, right?
00:13:51.200 --> 00:13:54.320
It's like when vibe coding went to college and and graduated.
00:13:54.480 --> 00:13:59.200
Now we call it agentic engineering um or you know, agentic software engineering.
00:13:59.360 --> 00:14:05.200
Um, the problem is a lot of companies thought that they could own something, right?
00:14:05.519 --> 00:14:15.279
You know, uh anthropics claude code really uh I I I guess exploded this um this sector in late 2025 and early 2026.
00:14:15.360 --> 00:14:20.399
And a lot of companies that were early adopters thought that they could, well, they could build it and they could own it themselves.
00:14:20.559 --> 00:14:32.799
And then, well, they realized that if they didn't actually know what they were doing, and if they actually put these uh you know tools or systems into production, maybe they were, you know, dime smart dollar dumb.
00:14:32.960 --> 00:14:38.399
And they maybe solved one problem, but created a couple more uh with some expensive tech debt.
00:14:40.000 --> 00:14:44.399
So this is how we need to frame these decisions now.
00:14:44.639 --> 00:14:46.559
It's not a single choice, right?
00:14:46.720 --> 00:14:51.120
You don't just decide build by partner or wait for your entire organization.
00:14:51.440 --> 00:15:00.720
Because uh, if you are implementing AI from the front office uh to the back admin to every employee in between, there's a lot of different layers.
00:15:00.799 --> 00:15:01.600
So we talked about that.
00:15:01.679 --> 00:15:02.559
There's the model.
00:15:02.879 --> 00:15:04.480
What model are you using?
00:15:04.639 --> 00:15:06.399
The workflow, all right.
00:15:06.480 --> 00:15:15.679
Uh, and that's not just the it's not just the harness that a model uses, but it's how are you actually using that model in your day-to-day workflow?
00:15:15.919 --> 00:15:18.159
Then there's number three, your data.
00:15:18.799 --> 00:15:39.840
Okay, that's bringing your dynamic data in and having a sound data strategy, making sure that you have uh clean, structured data that is dynamic, and then your business software or connecting it uh both on the input side uh so you can grab data from that software that you need, but also to write it, uh, right.
00:15:40.000 --> 00:15:49.919
Because I think even nine months ago, most software integrations were one-way streets, they're bidirectional now, right?
00:15:50.000 --> 00:15:59.200
It is uh most, it's kind of almost the status quo for most, uh kind of if you're looking at it as simple as connectors or apps, right?
00:15:59.279 --> 00:16:04.159
Different uh different big four call them, different things, but they're bi-directional now, right?
00:16:04.240 --> 00:16:05.039
Read and write.
00:16:05.200 --> 00:16:14.799
It's no longer good enough to just be able to read your emails dynamically or to read your calendar and triage between your email and your calendar and your storage, right?
00:16:14.879 --> 00:16:21.919
Now you need to be able to write, you need to be able to create a uh a calendar out of your AI system, uh a calendar entry.
00:16:22.000 --> 00:16:25.679
You need to be able to create a new file um agentically, right?
00:16:25.840 --> 00:16:27.840
On demand, on a schedule, on a cron.
00:16:28.240 --> 00:16:35.600
Um, but that's not it's not happening um because it's getting more and more complex and the updates are coming out faster.
00:16:35.840 --> 00:16:38.720
So you need to really be intentional.
00:16:38.879 --> 00:16:50.320
So if you are in the boardroom, if you are helping make these decisions on AI implementation, this isn't one size fits all, right?
00:16:50.720 --> 00:16:53.360
You don't say we're gonna build this, we're gonna buy it, and that's it.
00:16:53.679 --> 00:16:56.159
You do have to look at it in those four separate layers.
00:16:57.840 --> 00:17:12.559
But obviously in the long run, you want to if you can, if you have the foundation, if you have the talent, if you have the data, you always want to skew toward owning if you can't, right?
00:17:13.119 --> 00:17:15.519
There's there's no doubt about that.
00:17:15.680 --> 00:17:19.680
But it's not always feasible or possible to own.
00:17:19.839 --> 00:17:23.599
But when we talk about owning or building something, right?
00:17:23.759 --> 00:17:35.119
So kind of using those two terms interchangeably, uh, you should be building for proprietary processes for internal context and workflows that are unique to how you operate, right?
00:17:35.200 --> 00:17:45.279
So you shouldn't be as a lot of companies were wasting a lot of money in 2023, uh, you know, trying to essentially build their own large language model, which go back and listen.
00:17:45.440 --> 00:17:48.160
I said all along, I'm like, that's not smart.
00:17:48.319 --> 00:17:58.160
Uh, you know, maybe you got some nice gains in, you know, in 2023, but then you were left with tech debt or you had to kill the project.
00:17:58.720 --> 00:18:06.240
So uh that same retool study found that 35% of AI builders replaced a purchase tool with a custom one.
00:18:06.480 --> 00:18:08.400
So this shift is happening, right?
00:18:08.559 --> 00:18:10.079
I was even talking about this recently.
00:18:10.240 --> 00:18:11.440
I had a lot of software.
00:18:11.519 --> 00:18:15.759
Uh you know, I'm a small business, but I just replaced it, right?
00:18:15.839 --> 00:18:22.400
Like if I'm using a tool and I'm like, ah, some things are good, but some things I don't like, I just build it, right?
00:18:22.559 --> 00:18:23.759
That's not gonna be the case.
00:18:23.839 --> 00:18:32.559
And I'm not saying that, you know, you're gonna have a bunch of SMBs, you know, being like, all right, I'm out, uh, Zoho, CRM, I'm building my own.
00:18:32.799 --> 00:18:37.759
It's not feasible or or realistic, although it is possible.
00:18:37.920 --> 00:18:48.799
Um, but I think you have to look at building those common workflows yourself, where you can actually um have a long-term competitive advantage.
00:18:48.880 --> 00:18:58.400
So when you're thinking about the things you should be building, when you are thinking about the things you should have ownership on, think about your proprietary processes, uh, that internal, right?
00:18:58.480 --> 00:19:01.680
I always call it first company, uh, first company data.
00:19:01.759 --> 00:19:15.200
Uh, so it's not just what shows up in the spreadsheets, it's what shows up in your subject matter experts' minds, what's in your uh your SOPs, uh, what's in your meeting notes.
00:19:15.359 --> 00:19:18.240
Those are the things that you want to build ownership around.
00:19:18.319 --> 00:19:18.799
Uh, right.
00:19:18.880 --> 00:19:37.039
And a lot of times you'll be porting those off to other places, but those processes, even something as simple, something as simple as having a meeting assistant and being able to take those transcripts and instantly share that knowledge with your organization, that shouldn't be locked up behind a certain vendor.
00:19:37.200 --> 00:19:38.960
That should be a process you can own.
00:19:39.119 --> 00:19:42.960
And yes, there are local models that are easy enough, right?
00:19:43.039 --> 00:19:45.599
I built this for myself, not terribly hard, right?
00:19:45.680 --> 00:19:47.759
So your organization should be looking at those.
00:19:47.839 --> 00:19:51.599
Those processes are ones that you should own, right?
00:19:51.839 --> 00:19:53.759
Your proprietary data.
00:19:54.079 --> 00:19:58.880
Um, and I have a little graphic here uh on my screen.
00:19:59.039 --> 00:20:01.920
You know, you should be building what compounds.
00:20:02.559 --> 00:20:12.240
So, not just the proprietary data, but something that can help you create a durable moat and something where lock-in would destroy the value, right?
00:20:12.319 --> 00:20:13.279
Think of something like this.
00:20:13.359 --> 00:20:24.160
Think of something like uh, you know, if you're trying to use, you know, Zoom and Fireflies AI uh and Google Meet all at the same time in your organization.
00:20:24.480 --> 00:20:25.920
Well, that's not good.
00:20:26.000 --> 00:20:42.240
You either, well, need to build a process at least where you can have complete ownership of that and it's not segmented and fragmented across these different buckets, or just build your own solution because that at the baseline, think of what compounds.
00:20:42.319 --> 00:20:54.400
And I think it's people's time, it's your subject matter expertise, and it is that IP, it is that first company uh data that is essentially, I think, the fuel for reasoning models.
00:20:54.480 --> 00:20:56.640
That's what you should be building around.
00:20:57.359 --> 00:21:03.039
Next, buy when the workflow is common and already in your software, uh, right.
00:21:03.279 --> 00:21:09.599
So you should buy common, regulated, audit-heavy workflows that already sit inside of your existing platforms.
00:21:09.759 --> 00:21:10.319
Uh, right.
00:21:10.400 --> 00:21:25.759
So this is obviously going to be different from sector to sector, uh, but outside of the big four, there's probably good reasons to buy certain software, you know, AI-powered software that fits within those workflows, right?
00:21:25.839 --> 00:21:38.079
And maybe that SAP um is a great uh example, maybe something if you're a startup, uh you know, on the entrepreneur scene, maybe it's something like Slack, right?
00:21:38.160 --> 00:21:49.440
Uh just because when you're talking about buying workflows inside of regulated or audit heavy um or existing platforms, it doesn't mean that you have to think these, you know, large enterprise, right?
00:21:49.519 --> 00:21:51.759
It could be something as simple as ClickUp.
00:21:51.839 --> 00:21:55.680
It could be something as you know, simple as Slack, right?
00:21:55.839 --> 00:21:59.839
So you could be doing a lot of these things via third party uh tools.
00:22:00.480 --> 00:22:25.359
But if or sorry, via the the big four providers, but if your workflows are so ingrained right in this enterprise, uh in certain enterprise uh ecosystems, well, maybe that's where you buy, maybe that's where you yeah, spend that extra money on you know the the the Slack AI, spend that extra money on the clickup AI, second braid, whatever they're calling it nowadays.
00:22:25.519 --> 00:22:26.079
The same thing.
00:22:26.240 --> 00:22:36.799
Maybe you spend that extra money on uh on the sales force, whatever it is, wherever your team's uh data is locked into your day-to-day processes.
00:22:37.519 --> 00:22:40.240
So when you think about uh, I'm kind of saying zone two.
00:22:40.400 --> 00:22:45.519
So zone one is build what compounds, zone two is buy what protects, right?
00:22:45.599 --> 00:22:52.960
So those workflows that are, like I said, common, regulated, mission critical, and or require that strict audit log.
00:22:53.119 --> 00:23:04.799
So yeah, if you're somewhere in uh financial uh something HIPAA that needs to be HIPAA compliant, maybe you can't get that specific workflow uh off the shelf from some of the big four.
00:23:04.880 --> 00:23:06.319
So buy what protects.
00:23:06.480 --> 00:23:11.359
Number three, partner when the cost of failure is too high to go alone.
00:23:11.519 --> 00:23:11.920
All right.
00:23:12.000 --> 00:23:20.640
So, you know, if this is something in cybersecurity, regulated decisions, physical operations, critical infrastructure, that's when you should be looking at partnering.
00:23:20.799 --> 00:23:21.119
Okay.
00:23:21.839 --> 00:23:27.119
A good partner also needs to share the execution, uh, execution risk.
00:23:27.519 --> 00:23:33.920
But some AI categories maybe are still a little bit too unstable uh to commit to.
00:23:34.160 --> 00:23:37.680
So uh partnering is important.
00:23:38.000 --> 00:23:44.720
So I'm not just saying, oh, like on the legal side, we need to go get, you know, Harvey or whatever, right?
00:23:44.880 --> 00:23:50.640
Or, you know, Claude did come out with some uh, you know, some impressive uh legal agents.
00:23:50.720 --> 00:23:57.039
Um, same thing with uh Microsoft Copilot, they just rolled out some support uh for some legal agents as well.
00:23:57.200 --> 00:24:21.599
But when you think about zone three here, partnering where failure is too expensive, if the workflow touches physical operations specifically, uh, right, or highly complex domain data that is extremely valuable, um maybe, you know, you this is kind of the like the thing I think of this um, you know, when you partner when failure is expensive.
00:24:21.839 --> 00:24:25.359
I kind of think of this as the on-prem versus off-prem, uh, right.
00:24:25.440 --> 00:24:33.680
So if there's certain things that you would just do on-prem, whether it's, you know, from a data storage uh or whether just, you know, physical operations, right?
00:24:33.839 --> 00:24:37.119
We can't move this operation, you know, somewhere else.
00:24:37.200 --> 00:24:39.680
Well, because it needs to be physically located here.
00:24:39.839 --> 00:24:44.319
Um, I think that's where partnering is a big reality.
00:24:44.559 --> 00:24:48.720
Um, so last but not least, wait.
00:24:49.039 --> 00:24:49.359
All right.
00:24:49.519 --> 00:25:00.559
And I will preface this by saying that waiting can be both the best decision that your company makes and the absolute worst decision that your company makes.
00:25:00.799 --> 00:25:01.920
Let me explain that.
00:25:02.559 --> 00:25:12.720
Again, not to keep beating the dead horse, all right, but I'm gonna go pick on the uh the initial uh you know, rag pipelines, right?
00:25:13.119 --> 00:25:17.440
Uh don't get me wrong, rag's not dead, right?
00:25:17.519 --> 00:25:19.519
It's just it's it's just changed, right?
00:25:19.680 --> 00:25:25.519
But if you go back to quarter one of January 2023, and probably quarter two as well, right?
00:25:26.319 --> 00:25:31.039
There was a lot of companies who were on the edge, right?
00:25:31.279 --> 00:25:34.240
Not edge devices, but on the cutting edge of AI.
00:25:34.400 --> 00:25:43.359
And they were in theory making good decisions by saying, hey, this large language model, this generative AI thing, we're gonna get on the train, right?
00:25:43.519 --> 00:26:01.359
We're gonna invest heavily instead of using a consumer chatbot, which at the time, right, you didn't have the ability to connect your, you know, in five clicks, you couldn't bring in your emails, you couldn't bring in your SharePoint, your OneDrive, your box, right?
00:26:01.519 --> 00:26:02.240
All these things.
00:26:02.319 --> 00:26:03.440
It was impossible.
00:26:03.599 --> 00:26:07.759
You couldn't, it wasn't an option in the off-the-shelf consumer chatbot.
00:26:07.920 --> 00:26:11.839
So, you know, a lot of bigger companies had to make this investment.
00:26:12.000 --> 00:26:18.319
And at the time, it might have seemed like a smart idea, even though go back and listen, I was saying don't do it.
00:26:18.559 --> 00:26:22.559
Um, this is one of those times when you could have waited.
00:26:22.640 --> 00:26:45.759
Uh, because even to do something like that correctly and to train your entire organization on taking advantage of okay, now all of a sudden we, you know, ragged up our uh data to, you know, I don't know at the time, what was it, GPT 3.5 or GPT-4 or uh the original Claude 2 or Claude 3, right?
00:26:45.920 --> 00:26:52.960
Um number one, the models weren't good enough to use in production, anyways, uh, without some great prompt engineering.
00:26:53.119 --> 00:27:04.240
Uh but number two, I think there's a big uh divide between non-agentic AI and agentic AI.
00:27:04.400 --> 00:27:04.960
Uh right.
00:27:05.200 --> 00:27:21.440
So when the you know the AI was essentially just super smart, next token completion, you know, the fact that these companies uh really wanted to go all in, it's brave, it's courageous.
00:27:21.519 --> 00:27:30.640
And I think you have to make some of those decisions, but sometimes the brave and courageous decision is to wait, especially now, knowing the pace and the rate of innovation.
00:27:30.799 --> 00:27:31.359
Uh, right.
00:27:31.519 --> 00:27:40.079
So, as an example, maybe you're just waiting for a certain industry-specific software uh that everyone in your industry uses.
00:27:40.160 --> 00:27:47.440
Let's say you're in construction and everyone uses, you know, something to you know create bid shells for RFPs, right?
00:27:47.599 --> 00:27:52.319
Um, and you're just waiting because there's two pieces of software everyone uses.
00:27:52.480 --> 00:27:55.839
So you're like, okay, do we wait or do we build?
00:27:56.640 --> 00:28:03.119
You have to understand the risk versus the reward versus the rate of innovation, right?
00:28:03.279 --> 00:28:12.960
So, as an example now, with not a lot of technical expertise, companies can spin up a model context protocol server.
00:28:13.279 --> 00:28:33.599
And even if there uh wasn't a bridge uh between enterprise software that is crucial for your day-to-day versus the AI operating system that your company or department uses, well, now with model context protocol, you can build that, you know, maybe in an hour, right?
00:28:34.079 --> 00:28:41.839
Uh, so this concept of waiting, there's no one size fits all, but you have to be able to wait when vendors change monthly.
00:28:42.000 --> 00:28:50.240
You have to be able to wait if the ROI is unclear, uh, or you might have to wait until the category is a little more stable.
00:28:50.480 --> 00:28:51.279
A good example.
00:28:51.359 --> 00:28:53.119
Yeah, here's another one, y'all.
00:28:53.359 --> 00:28:57.680
Not saying I was right, maybe I was wrong, but I'd say I was right right now.
00:28:57.920 --> 00:29:01.920
You notice, don't talk a lot about open claw on this show, right?
00:29:02.079 --> 00:29:08.480
I got a lot of emails, people being like, oh, Jordan, you could be making so much money, you know, teaching people open claw.
00:29:08.559 --> 00:29:09.039
And I said it.
00:29:09.119 --> 00:29:11.359
I'm like, hey, open claw is great, right?
00:29:11.680 --> 00:29:15.599
It is the most successful open source project of all time.
00:29:15.839 --> 00:29:24.400
Nvidia CEO Jensen Wong, right, said it's the best thing since sliced bread, isn't what he actually said, but he called it the best piece of open source software ever.
00:29:24.640 --> 00:29:30.079
Um, and a lot of companies and a lot of very visible leaders have invested into open claw.
00:29:30.160 --> 00:29:34.240
And I've been saying since the beginning, do I have an open claw?
00:29:34.400 --> 00:29:34.720
Sure.
00:29:34.880 --> 00:29:35.839
Do I use it?
00:29:36.000 --> 00:29:37.200
Not much, right?
00:29:37.359 --> 00:29:43.920
I use it enough to know how it works and to you know see the struggles and the successes other people are having.
00:29:44.160 --> 00:29:55.680
But I knew at the time, I'm like, this is one of those times when it's better to wait because I knew that the big four are eventually going to have their versions, uh, right.
00:29:55.759 --> 00:30:01.279
And I'm not talking like NVIDIA's version of OpenClaw Nemo Claw, where it's essentially open claw in a more secure environment.
00:30:01.440 --> 00:30:06.400
I'm saying no, they're gonna have ChatGPT is gonna have its own version of OpenClaw, right?
00:30:06.480 --> 00:30:12.319
Claw is gonna have its own version of OpenClaw, and that's obviously where all of the big players have been trending.
00:30:12.400 --> 00:30:25.359
So I don't know, is it worth it to get, I don't know, 500 employees on an open claw, and all of a sudden they set all these things up, but their heartbeat's not working, their sole MD files corrupted, cron jobs aren't crowning, right?
00:30:25.440 --> 00:30:30.240
Now all of a sudden, I don't know, by being on the cutting edge, maybe you were able to take advantage of it.
00:30:30.400 --> 00:30:37.839
Maybe you were able to squeeze uh you know, the juice while the lemons were were ripe and no one had lemonade and everyone was thirsty.
00:30:37.920 --> 00:30:50.640
Maybe you had that opportunity, but probably I would venture to say that 90% of you know enterprises that invested heavily in open claw, the learning curve is so steep with just large language models.
00:30:50.720 --> 00:30:51.200
Uh, right.
00:30:51.359 --> 00:30:54.880
I think this is one of those where, well, the ROI was kind of unclear.
00:30:55.039 --> 00:30:56.000
Yes, it's exciting.
00:30:56.079 --> 00:30:58.799
And that's one of the problems with uh with AI.
00:30:59.039 --> 00:31:00.480
Everything you see is exciting.
00:31:00.640 --> 00:31:02.480
Everything you see, it's like MAD FOMO.
00:31:02.640 --> 00:31:05.519
Oh my gosh, if we're not using this right away, we're gonna go out of business.
00:31:05.680 --> 00:31:07.279
That's not the right way to look at it.
00:31:07.440 --> 00:31:14.960
You have to understand what's possible today, what's coming next, and then where do we build, buy, partner, or wait.
00:31:15.039 --> 00:31:20.079
And sometimes the bold and courageous thing might be being first or it might be waiting.
00:31:20.240 --> 00:31:23.920
But there's no uh, you know, one size fits all answer on that.
00:31:24.079 --> 00:31:29.359
But waiting while competitors redesign their workflows, though, that's drift.
00:31:29.519 --> 00:31:29.920
All right.
00:31:30.079 --> 00:31:36.240
So if you see everyone else in your space is doing something, yeah, maybe your strategy is to zig while everyone else zags.
00:31:36.400 --> 00:31:45.200
But if you can see that there's a product market fit with a certain early AI technology in your space, maybe that's not the time you wait then, right?
00:31:45.279 --> 00:31:46.400
Maybe that's the time, you know.
00:31:46.480 --> 00:31:52.319
As an example for OpenClaw, there's probably great uh enterprise use cases for it until everything else caught up.
00:31:52.400 --> 00:31:57.759
And now, I mean, I don't know, Codex is essentially a way more stable OpenClaw, right?
00:31:57.920 --> 00:32:04.720
Obviously, uh, you know, uh OpenAI kind of aqua hired uh, you know, the the sole creator of OpenClaw.
00:32:04.880 --> 00:32:11.759
So we're seeing a lot of the best pieces of it get implemented in a much more secure, robust, and enterprise-friendly way.
00:32:11.920 --> 00:32:20.319
All right, so uh so this is kind of zone four, is the danger of wait because it's also the capabilities gap, uh, right?
00:32:20.400 --> 00:32:32.160
So you need to look at the choice to wait on a case-by-case basis, uh, because it's not only the capability of the technology, it's also the learning curve of the people who have to use it.
00:32:32.319 --> 00:32:34.319
So I think the open claw is a good example there.
00:32:34.480 --> 00:32:43.039
Now it's much easier to go inside of something like Codecs, which if your company uses ChatGPT, Codex is so easy to see and understand.
00:32:43.200 --> 00:32:53.519
Or, you know, the um the workspace agents inside of um, you know, ChatGPT business plans, which are essentially like simple little open claws, right?
00:32:53.599 --> 00:32:54.960
But these are kind of the same things.
00:32:55.039 --> 00:33:03.599
They're things you can, you know, message uh from your phone and can connect to all the data sources and you can run them on a schedule, right?
00:33:03.759 --> 00:33:05.680
But sometimes there is the danger of waiting.
00:33:05.759 --> 00:33:14.160
But you do have to look first at the capabilities gap uh and the uh the knowledge gap and the training gap, right?
00:33:14.319 --> 00:33:18.000
So is your company still in the chatbot era?
00:33:18.240 --> 00:33:22.799
If so, you have to understand that capability gap is going to be a steep hill to climb.
00:33:22.960 --> 00:33:46.799
But if you are using for the most part, if your company is using models agentically, if you're already there, maybe waiting doesn't make as much sense because it could only take you, you know, I don't know, a quick, you know, three-hour training and you know, weekly check-ins to uh kind of close the capability gap so your team can actually take advantage of the latest and greatest in AI.
00:33:46.960 --> 00:33:47.359
All right.
00:33:47.519 --> 00:33:53.359
So here's the four wrong calls before we get to the one, two, three week plan.
00:33:53.680 --> 00:34:01.279
So uh wrong call number one, being wrong in the build creates technical debt that your team might have to maintain forever, right?
00:34:01.599 --> 00:34:08.079
Yeah, I know there's still companies out there maintaining their uh their rag chat bot from 2023.
00:34:08.480 --> 00:34:09.840
Uh wrong buy.
00:34:10.000 --> 00:34:13.840
That can create vendor lock-in that you can't exit without a major cost, all right?
00:34:14.000 --> 00:34:16.079
Don't there's too much competition up there.
00:34:16.159 --> 00:34:19.599
Don't get locked into a contract for two or three years, right?
00:34:19.760 --> 00:34:22.559
These vendors are gonna sell you something like there's no other option.
00:34:22.639 --> 00:34:23.599
There's always option.
00:34:23.679 --> 00:34:24.000
All right.
00:34:24.159 --> 00:34:30.880
The wrong partner, number three there, that can create dependency on someone whose business is also changing.
00:34:31.039 --> 00:34:44.639
Uh, working with the right partner is sometimes trickier than the uh the buy versus build, uh, because you have to then understand a third party's uh complete front to back.
00:34:44.719 --> 00:34:52.480
But you know, your partners should be extremely niche uh and like we talked about earlier, you know, um in those certain areas.
00:34:52.559 --> 00:35:00.480
And then last but not least, the wrong weight decision that can create a capability gap that competitors can quietly close first.
00:35:00.639 --> 00:35:09.119
So here is your one, two, three week blueprint, all right, to go through and understand your build by partner or weight decisions.
00:35:09.280 --> 00:35:12.000
So week number one, you're gonna audit and score.
00:35:12.159 --> 00:35:15.280
Week number two, you're gonna execute the extremes or the edges.
00:35:15.440 --> 00:35:18.719
And then week number three, you're gonna you're gonna govern and hedge.
00:35:18.880 --> 00:35:19.679
Here's what I mean.
00:35:19.840 --> 00:35:24.239
Uh so uh do have this on a screen here, but I'm just gonna read it all off.
00:35:24.320 --> 00:35:28.400
But you can always go watch the video version on our website at your everydayai.com.
00:35:28.559 --> 00:35:30.000
So week one, audit and score.
00:35:30.079 --> 00:35:38.320
You're gonna pick 10 workflows, you're gonna score them on an ownership map, and then you're gonna name one accountable human uh owner for each.
00:35:38.400 --> 00:35:44.880
All right, that's where you always need to start before you even go down the four different layers and the build by partner weight.
00:35:45.039 --> 00:35:50.079
You have to have an accurate inventory of what AI is actually being used for.
00:35:50.239 --> 00:35:53.679
Uh all right, and then week two, you need to execute the extremes.
00:35:53.920 --> 00:36:00.239
So you need to build one low risk or high differentiation, uh high differentiation internal workflow.
00:36:00.400 --> 00:36:10.320
Then you need to buy one packaged workflow inside your existing, you know, ERP, CRM, whatever the case, and then kill projects with no ROI.
00:36:10.480 --> 00:36:21.440
So essentially, first you're gonna audit and score, then you're gonna fit uh you're gonna find a low-risk, high uh high value internal workflow, and you're essentially gonna A-B test it, right?
00:36:21.519 --> 00:36:25.280
And you're gonna say, okay, what does this look like if we build it ourselves?
00:36:25.440 --> 00:36:28.320
And what does it look like if we buy something else?
00:36:28.400 --> 00:36:32.559
You need to measure your human costs and your external costs for the buy side.
00:36:32.719 --> 00:36:38.480
And then last but not least, week three, uh, you need to compare the build versus buy performance.
00:36:38.639 --> 00:36:43.760
And then for the buy side, you need to renegotiate contracts for data export and model portability.
00:36:43.840 --> 00:36:46.639
Uh, because again, think of something like skills, right?
00:36:46.800 --> 00:36:48.880
Skills are modular, they're open source.
00:36:48.960 --> 00:36:52.639
You can take those, you know, as an example from vendor to vendor.
00:36:52.719 --> 00:36:58.159
So, you know, if you are doing something on the buy side, you need to make sure that number one, you're not locked in.
00:36:58.239 --> 00:37:09.519
But number two, that whatever um that you are investing, uh, you know, aside from the money, but your resources, your data, your education, you have to make sure that it's modular before you get in too deep.
00:37:09.679 --> 00:37:17.519
And then last but not least, in week three, make sure you establish a weekly review cycle in ongoing learning as well.
00:37:17.760 --> 00:37:18.400
All right.
00:37:18.800 --> 00:37:26.800
That is a wrap for the build by partner or wait, the four-layer AI stack decision framework for 2026.
00:37:26.960 --> 00:37:31.280
And y'all, this just like everything else, it all moves quickly.
00:37:31.360 --> 00:37:34.320
So, this is not a, you know, if you're listening to this, sorry.
00:37:34.400 --> 00:37:42.639
If somehow you're listening to this in 2027, you might just want to forget everything I said, uh, FYI, and instead just keep up with the daily podcast.
00:37:43.039 --> 00:37:51.920
Because, like everything, sometimes advice, sometimes the latest and greatest has a shelf like has a shelf life, just like the models that we all use.
00:37:52.000 --> 00:37:53.119
So I hope this was helpful.
00:37:53.280 --> 00:37:56.639
If so, do me a favor, go to starthirseries.com.
00:37:56.800 --> 00:37:59.519
That is going to get you free access to our community.
00:37:59.599 --> 00:38:01.440
That's the only way you can get in right now.
00:38:01.599 --> 00:38:16.800
And then make sure you go straight to the Start Here series and check out the Spotify playlist that has every single volume of the Start Here series ready for you to consume, whether it's in text, reading the email newsletter that went with it, watching the video, listening to the podcast.
00:38:16.880 --> 00:38:18.400
It's all there in one place.
00:38:18.480 --> 00:38:23.119
And you can also network and connect and ask questions with other people who are going through the same process.
00:38:23.280 --> 00:38:24.400
So, thank you for tuning in.
00:38:24.480 --> 00:38:27.440
I hope to see you back tomorrow and every day for more everyday AI.
00:38:27.519 --> 00:38:28.639
Thanks, y'all.
