Ep 860: Managing the AI Capability Gap: AI Is More than Ready. Most Companies are Not (Start Here Series Vol 19)

Welcome to the Everyday AI Podcast. My name is Jordan Wilson, and for the past 3 and a half years, we've put out more than 800 episodes. Yet one of the most common questions I get, I didn't really have an answer for. Where do I start on the Everyday AI Podcast? And that's why we started the Start Here series. And with fall now back in full swing, the Everyday AI Podcast is going back to school and playing back the entire Start Here series. From front to back. We've hit pause on our normal Monday to Friday programming to run back our most popular series ever for the next 30 days. We made the Start Here series for beginners and AI champions alike. So whether you're just trying to get a grasp on large language models or grappling with the best coding harness for multi-agentic workflows, the Start Here series covers it all. Plain language, no jargon, and easy to follow along each day. So make sure to subscribe to the podcast and Check back each day for new insights. Day by day, the series is a culmination of spending more than 10,000 hours covering generative AI over the past three and a half years. So you don't want to miss a single episode of the Start Here series. Let's get into it. There's been a lot of talk lately about these AI models that are scary good, and even more chatter about how many of the brightest minds in AI and all the tech AI CEOs feel that we've already reached artificial. artificial general intelligence and are racing toward superintelligence. But what does that matter to you or your company? Because right now those are kind of just theoretical issues. It's kind of like when you see an email on a Friday and you look at it and you're like, yeah, that's a Monday problem. But do you know what's scarier than some private AI model that's scary good or artificial superintelligence? The AI capability gap. That's the scariest of all. Why? Because it's very real. And unlike that Friday afternoon email that you're maybe going to tackle on Monday, the AI capability gap is a today problem. And if you don't address it head on, it's going to slowly stall your company's growth. And if your company isn't growing, the AI capability gap will undoubtedly stop your company dead in its tracks. And this is not an exaggeration. And the timing here is specifically urgent. So we're going to hit rewind today and explain the capability gap in AI and why AI is racing ahead of your company. And we're going to show you how to slow it down a bit and catch it. All right, you ready? This is part of our Start Here series on Everyday AI. Let's get into it. So if you are new here, well, let's just get to the big picture. AI capability right now is outpacing business adoption, and it has changed recently, and the pace is too fast for any of us right now. Frontier AI, and when we talk about actual capabilities, this is not an exaggeration. Frontier AI models, when used correctly, and that's the big asterisk here, they now match or exceed human professionals on most defined knowledge work tasks. And there's a massive gap that persists between what AI could in theory automate today and what organizations are actually using it for, which is often just topical, right? There's a McKinsey study that said only about 6% of organizations are generating meaningful profits despite widespread AI adoption. And the bottleneck is not actually the AI models. You could have made that argument maybe mid-2025, but not anymore in 2026. It's actually About organizational adoption, workflow design, and training and education. So here's what we're going to tackle on today's show. So stick with me for the next— I'm going to try to make this one 25 minutes. We'll see. All right. Stick with me for the next 22-ish minutes and you're going to learn the benchmark proving that AI outperforms human experts and it's not even close. You're going to know more about Anthropic's new-ish study revealing most knowledge workers barely touch AI's true automation potential. You're going to learn why the AI capability gap only started to show itself over the past few months. It's actually kind of new. And what the top 6% of AI-performing companies actually do that everyone else completely ignores. All right, welcome to Everyday AI, and this is our Start Here series. This is the essential podcast series to both learn the AI basics and for AI experts to double down. All right, I started this thing because after 750+ episodes, everyone always said, Jordan, great podcast, right? When they found it, they're like, where do I start? I have no clue. And I'm like, I also have no clue. So that's why I started the Start Here series. Uh, it's best if you are brand new here, start in order. All right. These are faster, uh, you know, podcasts, usually about 25 to 35 minutes. But if you listen on 2x, they're even half that, right? Uh, but this is a great way to listen from 1 to now volume 19. And then also make sure to go to starthearseries.com. That's going to give you free access to our exclusive inner circle community. You can't find access Right now, any gap report card. All right, so a little bit more on that at the end. But trust me, you are going to want to repost today's episode. I'm just saying. All right. And if you missed our last Start Here series in volume 18, and this was episode 750, we went over the vibe coding boom, why vibe coding isn't going away and how it's both good and bad. So make sure to go check that one out. But now let's talk about how you can actually manage the AI capability gap and why AI is more than ready and companies are not. All right. And I'm not the only one talking about this. I've been talking about this now since I think late 2025. But a lot of super smart people are starting to dive in deep on the AI capability gap. I really like what Jack Clark said. He is the Anthropic co-founder of Anthropic. And here's what he said in a post a couple of weeks ago. He said most of AI progress has this flavor. If you have a bit of intellectual curiosity and some time, you can very quickly shock yourself with how amazing capable modern AI systems are. But you need to have that magic combination of time and curiosity. And otherwise, you're going to consume AI like most people do, as a passive viewer of some unremarkable synthetic slop content, or at best just asking your LLM of choice how to roast a turkey and keep it moist, or Tonybox lights spinning but not playing music What do I do? And all the amazing advancements are mostly hidden from you. All right. So that's what Jack Clark said in a kind of a viral post that he had a couple of weeks ago on X. And I think this is very telling. And the 2 things that he talked about, I think are very true. All right. To start, well, first, to even realize the AI capability gap, let alone close it, you have to be extremely curious and The reason why I say that is because if you work how you've been working for the last few decades, you're not going to discover, or your team, your organization is not truly gonna discover that capability gap because AI is not meant for the average knowledge worker, right? That's why I talk about all the time about how I hate, absolutely hate the concept of upskilling because AI is not something that you sprinkle on top of a seasoned knowledge worker, right? Like myself, I've been working full-time for 20-ish years, right? I can't just sprinkle AI on the top. I have to, I have to unlearn, right? So you do have to have the time and you have to be curious on, hey, what does my role look like if I completely started over? And if I kind of forgot everything that got me to the point I am in my career, and that's what you have to do. And you also have to have a lot of time and you have to devote a lot of time to it. Olivia Moore from a16z talked about, she said OpenAI dropped a state of enterprise reports across a million customers. This was a couple of months ago, but her response, she said, The gulf between AI power users and everyone else is wide. The 95th percentile user spends 6— sends 6 times more messages than the median, with coding, writing, and analysis showing the biggest gaps. So yeah, those people that are power users, well, because they actually understand the capability gap, they're using AI all the time. This is why myself— and I'm not saying this is like some weird flex, I'm saying this because I am part of this, you know, power user group— when I say that I use AI for 10 to 12 hours a day, Right? That's how much I'm working. You know, it's not like I'm working 16 hours a day. I am only working in AI every single day, no matter where I am. I'm using AI from beginning to end in literally every single step in between, and it has completely reshaped my workflow. Kevin Roose, New York Times columnist, said this, and I really like how he put this. He said, I follow AI adoption pretty closely, and I have Never seen such a yawning inside-outside gap. People in SF, San Francisco, are putting multi-agent cloud storms in charge of their lives, consulting chatbots before every decision, wireheading to a degree only sci-fi writers dared to imagine. People everywhere are still trying to get approval to use Copilot in Teams if they're using AI at all. It's possible that the early adopter bubble I'm in has always been this intense, but there seems to be a cultural kick takeoff happening in addition to the technical one. Not ideal. And I'll say from my experience, I think there's 5 main reasons that the AI capability gap has really been exposed in 2026. And number one, the models of what they can actually do, right? I'm just going to roll through these things really quickly. So number one, what AI models can actually do. Number 2, what business leaders think AI can do. Number 3, AI literacy. Number 4, human skill set. And number 5, AI access. Now let me break those 5 things down a little bit more in depth. So number 1, what AI models can actually do. A year ago, right? So I don't know when you're listening to this Start Here series episode, you know, could be in April 2026. You might be listening to it at the end of 2026. I don't know. Right. But a year ago, in the beginning of 2025, You could, for the most part, understand what AI models are capable of. Today, you absolutely can't. And it is literally getting to the point of science fiction, right? All of these math problems and science problems that have been plaguing researchers for decades are now being solved. I look at them and I have no clue, right? I read the papers and I'm like, okay, I don't have any clue. I can follow and understand it a year ago. What AI models can actually do today is Mind-boggling. Number 2, what business leaders think AI can do. I get to talk to a lot of smart people. It's really, right? I've had hundreds of guests on this show. I'll say maybe a handful that I've talked to, and I'm like, yes, this business leader fully understands what AI can do for the most part. And that's not their fault necessarily, because most people in their role, they have to be an expert at one Application, right? You can't be in, like, very few business leaders know what AI can do. I feel fairly confident that I'm in that category, but only because this is all I do. If I had another job, right, and I was just an AI champion at a company, you can't, like, you literally can't. You used to be able to, you can't anymore because literally every single day, right, whether you're talking about Google, Microsoft, OpenAI, Anthropic, now Meta's back in the conversation, Perplexity, et cetera, right? You literally can't have an actual job where you have to produce something for a company and still understand what AI can do. Not anymore. You used to be able to juggle that. Now, unless you're literally working 20+ hours a day, you can't do it. All right, number 3, AI literacy. And this kind of goes, you know, hand in hand with, you know, number 2, what business leaders understanding. So there's a difference between understanding AI's capabilities and then AI literacy because you have to actually be able to speak the The language foundationally, which again, most people skip over that, right? They go straight to the bells and the whistles. They go straight to clicking the button, not knowing what happens, you know, from A to X. They only want to get to Y and the result that brings on Z. Number 4 is human skillset. All right. This is different than literacy. You have to have the foundational understanding, but then you also have to have these skills, right? I know a lot about basketball. But my basketball skill set, not great anymore, right? I think I peaked in 8th grade. But you have to have that human skill set. And then last but not least is AI access. So that's access to the tools, to the technology in your role. Okay, so what can AI models actually do? Right? And I do want to spend a little bit more time on that, because that will better explain this gap that I think started manageable, and now it's very hard to manage that capability gap because it has turned into a friggin canyon. So the reason being is because AI models are better than expert humans. All right, I've talked about this, this evaluation a couple of times, but the more I follow these AI benchmarks and all of these other things, right, there's, there's dozens of them. I think there's about 5 that matter. Right? And one of them, and I think probably the most important one, is GDP Val. So this is a benchmark from OpenAI, and the reason why I think it's the most valuable benchmark to look at is because it's about creating business value. So more or less, this benchmark evaluates AI models against real professional deliverables across 44 different high GDP occupations. So these are things that you and I may do, you know, going in and doing research on a fast-moving market segment and then creating an artifact like a, you know, a pitch deck or creating a spreadsheet, something like that. These are real-world domain-specific tasks where both a human and an AI model from start to finish completed a task and then submitted their final version to a panel of experts who are experts in that field. All right. And expert judges then compared the unlabeled AI and human outputs in a blind Head-to-head pairwise comparison. And what we've seen now is, as an example, the best model for actually creating front-to-back economically valuable outputs like a human would is the OpenAI's GPT-5.4 model, and it matches or exceeds industry professionals in 83% of these evaluations, right? I remember when some of these first GDP-bound numbers came out and I'm like, oh, that's pretty impressive, right? When we were in the 30, 40%, but now it's undeniable. And I do assume that by the end of the year, that number is gonna be like in the mid-90s, right? My, like one of my predictions is it was gonna get to 80% and we're already at 80%. It was just a huge jump. That's why I keep talking about the AI capabilities are just running too fast for most business leaders to understand and keep up. And it's nearly doubled. That GP Bell score has nearly doubled in 5 months. So when I talk about the last, you know, since essentially late 2025, I'm going to talk about why. But it has been— we've seen more developments in the past 5 months than we have in the 5 years prior, at least when it comes to large language models. And it's not even close. Right. So the— in October, the best AI model scored on that was a 47%. So it hasn't quite doubled, but it's nearly doubled in about 5 or 6 months. And right now AI handles most defined professional tasks, yet most organizations barely use models in this way. It's like, did you know, right? As an example, your, you know, GPT and Claude models can do this. They can create artifacts by default. You don't have to have anything special turned on. It can just create spreadsheets, presentations, Word docs, etc., all personalized based on your company's info. All right. Next, I want to talk very briefly about the Anthropic labor data study that came out about 2 months ago. I did go over this in more depth. So if you're interested in this, go listen to episode 730. Here's essentially similarly, whereas the GDPVal benchmark looks at actual economic output, This Anthropic labor data study looks a little bit more at the capability gap. So that's why I think these 2 kind of this benchmark, OpenAI benchmark and the Anthropic labor data study in tandem really tell a powerful story of where we're at and why this capability gap even exists. So in this study, Anthropic mapped over 20,000 work tasks, right? So they actually used some public some public jobs data from the federal government. So they mapped over 20,000 work tasks across 800 occupations to millions of real AI conversations with their Claude chatbot. These were anonymized, but essentially they said, okay, what are people using Claude for? And they're matching millions of these chats to 20,000 work tasks across these 800 occupations from this federal data. And what they found in theory was that computer and math roles could solve 94% of— sorry, today's most capable AI models could automate up to 94% of computer and math tasks, right? But they were only seeing 33% usage, right? Because people assume, oh, computer and math, yeah, everyone's using AI for that because everyone knows that's the lowest hanging fruit, right? Anything in coding, software development, Yet even in the area where people assumed, oh yeah, literally everyone, if you're working in anything computer, right? Software engineering, anything dev, anything with math, of course you're using AI. And they said, well, actually not, right? And even office admin, business, financial, and legal roles all revealed the same deep kind of adoption shortfall, right? So for our livestream audience, you can always see the video version of this on our website at youreverydayai.com. So I have the theoretical capability and observed usage by occupational category, kind of map that Anthropic put together. Very fascinating. But all this shows is in certain categories, right, such as management, business and finance, computer and math, you know, architecture and engineering. Legal is another high one, arts and media, right? Like AI's capabilities when mapped to literal, the actual work tasks that the federal government uses to define these roles, I mean, most of them have coverage in the 80 to 90%. Yet the observed— or sorry, that's the theoretical coverage. So in theory, today's most powerful models could do anywhere from 80 to 90% of the actual work, right? Computer and math was the high one where they had an observed coverage of a little more than 30%. But these other areas, are not even 20%, right? They're for the most part, you know, in the 20s, some of them below 20. So you have this huge gap where in most of these general cases, right? Like management, you know, in theory, today's AI models, if you understand the capabilities, they can automate 85+% yet you don't even have a 20% coverage. That is a crazy gap, right? Where essentially you have a magic wand that if you know how to use the magic wand and you know the magic words and you say, poof, work be gone, poof, the work is gone. But people don't know because it's impossible to keep up with. And some other things that they found on this study, which is actually the workers that are most impacted, they said that workers most exposed to AI right now earn 47% more and hold graduate degrees at nearly a 4x rate. So essentially, I think people early on assumed that AI would displace or would potentially have the capabilities to displace workers who were maybe a little more junior or, you know, weren't as high up the kind of quote unquote knowledge work totem pole, so to speak. And what they found was the exact opposite, right? It is those people who have those higher degrees. And then conversely, people with physical and manual occupations registered lowest exposure according to Anthropic's study. So how did we get here so quickly? Right? How do we get, as an example, from GDP Val, where literally 5 and a half months ago, it was even, right? It was a little less. It was about a 47% tie or win rate against humans, right? It was a coin flip. And if the off-the-shelf model that you can pay $20 a month for was as good as a, you know, a human with a decade of experience who specializes in that to now humans aren't going to be able to compete in a couple of months, right? By blind benchmarks, almost every single time within probably 6 months, humans are always going to prefer the AI model. How did we get here so quickly? Right? 2 years ago, the AI models were not good at all, right? They were, they were actually Pretty bad, right? Especially obviously we have today's comparison to draw back on. But one of the biggest reasons I think is recursive self-improvement. Stick with me here if you're not super technical. All right. So recursive self-improvement or RSI, it's a concept within artificial general intelligence where essentially an AI system improves its own source code architecture or training data, leading to a more capable model, which then improves itself further, creating a self-reinforcing loop. All right. So why am I talking about this kind of strange niche concept called recursive self-improvement on something about managing the AI capability gap? Well, because at the end of 2025, the AI companies started to either directly admit or to kind of allude to the fact that they were all now using recursive self-improvement on their models, right? So now their quote unquote big models were good enough that it could start writing its own code. It could start improving itself, right? I think You know, Anthropic with their Claude code, probably one of the most consequential products of the AI lifecycle outside of ChatGPT. You could make the argument that Claude code is one of the most important outside of ChatGPT, right? The lead at Anthropic's Claude code says, yeah, we don't write code anymore. You know, Claude codes Claude code, right? And we've seen the same inferred by different researchers at— sorry, at OpenAI as well. You know, Google has been a A little less direct, but they've still alluded to the fact that a lot of their models today, not just the smaller versions that are distilled from the biggest versions, but the biggest versions are being improved by themselves, right? And these new big scary models, right? Like Anthropic's Mythos or, you know, OpenAI's, whatever it's called, you know, Spud or Glacier, right? If you're listening to this in 6 months, these names, these code names don't matter anymore. But one of the reasons that maybe the general public isn't getting their hands on them is, well, they're too compute intensive, but they are using these models internally to improve the best consumer models and to also ship new products. And that's why Anthropic, as an example, their ship rate in February and March was straight up off the charts, right? And then we found out later, well, one of the reasons was because they were using this Mythos model to put out A lot of this, a lot of these new features that we're all using now. And this is why, right? Because a year ago, it would have taken a team even using AI would have taken teams way longer. But now that you have this kind of recursive self-improvement, or models improving themselves and building new features that use those models, right? This is why it's hard now to keep up. So this started, like I said, in 2025. And ever since, large language model capabilities have far outpaced enterprise training and learning and development. You know, even the companies that wanted to do it right, I think they could, they could do it, you know, in, you know, quarter 2, quarter 3 of last year. But now you can't, right? Unless you have an entire segment of people, you know, maybe 5% of your total workforce That literally has no deliverables and all they do, right? I've said this all along. Your team needs a bunch of me's where all they do all day is they just scope different AI models. They play with AI releases, they sandbox things. They're not building anyone for anything. They're just building solutions for what they think the company needs. And then they're training those people, but companies don't have that, right? And that's why, at least right now, it is Nearly impossible to keep up. And most companies, though, claim AI adoption, but they can't actually prove real results. And it's getting even harder and harder, right? That McKinsey study that I talked about in 2025 said that 88% of organizations are using AI, but only 6% are generating meaningful business profits. And right now, I think leadership confidence is running far ahead of what frontline practitioners report actually seeing on the ground. That's the thing, right? So a lot of times, the people who are in charge of AI at certain companies now because it's getting easier to build. I think maybe their hands are on keyboard or their eyes are on monitoring agents a little bit more, and they are getting removed from what the frontline practitioners are actually experiencing. So not only is it pretty hard to manage that gap just from a technical perspective, but from a change management and a people management perspective, it's getting even And I think that's why some AI capability gaps are closing, but others are still remaining stuck. So as an example, right, AI is helping AI-enabled humans close some gaps, but many still exist and are getting worse. So as an example, coding performance, right, this in a lot of different benchmarks, this surge from single digits to 90% on structured benchmarks in 3 years in terms of what the AI itself was capable to do. And that helps obviously the humans that use this as part of their daily workflow close a big chunk of that gap individually. Right. But you now have this, you know, PhD-level models with, you know, science and standardized math capabilities that are rapidly approaching the ceiling of the current tests. So you aren't even necessarily able to know by the benchmarks what the capability gap is because these benchmarks are becoming saturated. So we, even the AI community doesn't even fully understand these models' capabilities because the benchmarks, right, a lot of them have been stuck in the 90s, to 95 percentile. You know, a lot of them are at 98, 99. So I think the benchmarks themselves are getting saturated. So even the people building the models aren't even fully aware or understand what these models are actually capable of. So let's talk about how you can actually manage this gap and start to close it. All right. Right now, the top performers, right, the top companies are investing more than 20% of their digital budgets to fundamentally Rework their processes. All right, let me repeat that: 20% of their digital budget. So whatever your digital budget is, I'm guessing most companies are probably at 1%, maybe maybe 5%. Very few, only the top performers are investing that 20% of their digital budget to rework their day-to-day knowledge processes. Right? Because you need to learn to separate workflows into risk. Tiers with verification and human approval matched to each level. And organizations that wait for AI to become ready, they're going to find that their competitors have already captured the advantage because in the same way, right, just to draw a little parallel here, what Anthropic— and I keep saying they've won 2026 so far. Well, it's because they were first, right? They were first to at least internally close that capability gap and put it to work. Right? And that's why they've been able to outpace their competitors. I don't know how long it'll last. We'll see, because I think the other labs have closed that gap, and I think we'll start to see that soon. But think of that within your own organization, right? The first to close the gap is going to be able to accelerate at a pace that we haven't seen before, right? That's why you see, unfortunately, a lot of these big companies, you know, Block as an example, cutting 40% of their workforce, and they seem fairly confident that they're going to be able to actually grow revenue because of the way they've completely reworked their organization. Actually, Jack Dorsey had a very fascinating essay that we shared about in our newsletter about how they're essentially flipping the work pyramid on its head. So here's what I want to have you focus on. One metric. Okay. I want to make this very digestible for you, uh, and hopefully actionable as we close out today's show. I want you to track the percentage of AI-assisted workflow steps right now that are accepted without any rework or incident. And here's why. Because you probably, most organizations, they kind of get their AI plan or their AI training for the year. Right? Or some companies, unfortunately, it's kind of like one time. Maybe the companies that are really investing heavily, you might get it once a quarter. But when was the last time that you used the most capable thinking models? I'm talking about GPT-5.4 Pro. I'm talking about Opus 4.6 with extended reasoning. I'm talking about Gemini 3.1 Pro with a higher thinking budget. Right? When was the last time that you used those? And track the percentage of your workflows that get accepted without rework or incident. And then I want you to break that metric down by risk tier, right? So the low risk, you know, drafting versus the high stakes legal or financial outputs. So organizations that measure and understand that operational reliability for the high percentage of AI-assisted workflows that can be accepted that are lower risk, you will be surprised. That depending on what your team does, depending on what your company does, depending on what sector you're in, you'd be surprised without too much investment aside from just reverse engineering your current day-to-day processes. You'd be surprised to say that about 30 to maybe 60% of a lot of the work that many of us do, right? I'm not talking about, you know, people in specialized role. I'm saying if your organization has 1,000 employees, you know, if you look at that work collectively, You'd be surprised. I would say 30 to 60%, if you rescope everything, would fall in that can essentially be automated without rework or incidents and is in a lower risk or a medium risk tier. And if you're not measuring that on an ongoing basis, you can't scale. But that's step one to managing the AI capability gap. And you have to be doing this Right? At least monthly. Again, a year ago, you could get away with quarterly. The way models, right? If you're using a model from last quarter, good luck, right? That's showing up to a, to an F1 race in a bicycle. Good luck. You're going to get smoked. You don't stand a chance. So you can no longer take these year-long pilots, these quarter-long plans. You have to be agile to actually understand Number one, the capability gap, but to begin to manage it. All right. I hope this was helpful in our Start Here series. Like I said, make sure to repost today's episode on LinkedIn. Here's why. We put together the AI capability gap report card to help you know where your organization stands when it comes to number one, understanding this gap, and number two, how you can actually tackle it. So in this report card, it's a great guide. For you and your team to go through it together to understand the latest capabilities of all the models and how they break down for different types of work. Alright. So if you didn't know, yes, if you're listening on the podcast, this is actually livestreamed on LinkedIn. So in the podcast show notes, we always put a link to today's LinkedIn show. So just go click repost, and I will send that capability gap report card Your way. All right. Thank you for tuning in. If you haven't already, please go to starthearseries.com. That's going to give you free access to the Inner Circle community, and you can go listen to all of the Start Here series in order in the playlist that we have, and you can go read and listen to all of the shows there. So thank you for tuning in. We hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers, and we'll see you next time.