transcript

Speaker 1:
[00:00] Today, we are discussing how the best companies at using AI are using AI. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Today's episode combines a number of the big themes that we've been talking about this year. 2026 kicked off with everyone coming back from the holiday break, really understanding and grappling with just how big an advance we had gotten between the harnesses like Claude Cote and Codex, and like the models like Opus 45 and the GPT-5253 series that had come out around the end of last year. Ever since then, it has just been a race, and it has been very clear that there is a massive difference and a growing difference between the people who are best at using AI and using it most fully and those who are not. Now clearly, a huge amount of that work is running through code, even when the ultimate output is not building software. Turns out that when you give AI and agents the ability to write code, it in fact unlocks a huge amount of other capabilities that are relevant for knowledge work way outside of just software engineering. When OpenAI announced their new Codex app, they said that a full 50 percent of the usage is not about coding specifically. Now, in many ways, the conversation that we've been having is focused on the individual. What you can as an individual do to get the most out of this new set of tools, how you can build and set up your own open clause, what features of the new Codex or Cloud Code app you should be using, things like that. Yet at the same time, another big theme for this year is that there is clearly a growing gap between the companies that are using AI best and those who are farther behind. A great example of this was on display with PwC study we talked about earlier this week that found that three quarters of AI's economic gains were being captured by just 20% of the companies. And importantly, it wasn't just that a fifth of the companies were the best at using AI, they were using AI in fundamentally different ways. They were not thinking just about productivity and efficiency, doing the same with less. They were instead thinking about AI as a growth technology, as an opportunity technology, in other words. They were two to three times as likely as others to say that they used AI to identify and pursue growth opportunities. And they were 2.6 times as likely to report that AI improves their ability to reinvent their business model. In other words, the leading companies are thinking structurally and differently about how to use AI on a core level. And this wasn't the only big consulting study or memo that we've recently seen that argued this theme. About a week and a half ago, McKinsey dropped its AI Transformation Manifesto. And once again, they pointed out that you can really divide the world of companies into two categories, the AI leaders and the AI laggers. By the way, for any of you who are grammar aficionados, I know that the more technically correct word would be lagerd with a D, but I like the colloquial lager better. So if it's been bothering your brain when I say lager, I apologize, but I'm going to keep doing it. In any case, McKinsey's Transformation Manifesto argued that there were 12 themes that separate these two categories of companies. Many of these will be extremely familiar to you if you are a regular listener of this show. Number one, McKinsey writes, technology alone doesn't create advantage, enduring capabilities do. In other words, you got to build systems around these new tools and the companies that are wired to build those types of systems who have built those types of systems in the past, who are adaptive by design are doing better than those who are just dropping tools on their employees' heads and hoping it works. Second, in another interesting departure from the efficiency or productivity idea of AI, McKinsey argues that the best focal points, the focal points that the leading organizations are focused on with AI are your economic leverage points. They write, any business model has a few key economic leverage points that provide the biggest impact when improved with AI. They give the example of automotive, where supply chain integration is the key leverage point, and one where companies like Toyota have been having breakthroughs with artificial intelligence. Now, relatedly, in number three, McKinsey goes a step farther and doesn't just say that the efficiency productivity approach isn't optimal, but that companies that aren't seeing actual measurable business value from AI are getting it wrong. They said that they studied the impact of 20 companies that they considered AI leaders across different industries and that on average, they found that the business transformations that came from AI and technology delivered a 20 percent EBITDA uplift, reaching breakeven in one to two years and generating three dollars of incremental EBITDA for every one dollar invested. Don't know how many different ways I can say it, this is not about efficiency and productivity, it is about business model transformation, new growth and new opportunity. Perhaps because of that, number four on their transformation manifesto is that you have to build the tech and AI muscle of your senior business leaders. That means not just deputizing IT leaders, not just leaving individual contributors alone to figure it out, but that the leaders who actually own different parts of the business, different lines of business, different units have to be in leadership roles, combining their domain expertise and experience with new AI know-how. And yet, of course, it is not just about leaders. One of the things that you will frequently hear Ethan Malik screaming about is the organizations that think that they can just outsource the transformation process and hire consultants and new experts to do it. But McKinsey argues that every tech and AI transformation is ultimately a people transformation, suggesting that more than 70% of talent for AI should be in-house. McKinsey argues that the companies that are ahead are thinking about their technology platforms in a fundamentally different way than others, not just as tools but as strategic assets that need to be invested in. And as part of that, those strategic assets need to be fed and enriched with what they need to live up to their potential, which in the case of AI is of course data. McKinsey argues that in most organizations, especially those that aren't using AI as well, data is the constraining factor. And it makes sense because doing data well means investing time in it, building data products, initiatives for data enrichment, and not just as a one-time transformation process but as a new ongoing operational discipline. Speaking of which, in another part of the manifesto, they talk about something which will come up again later in the show and which will provide a little bit more structure for, which is the idea of agentic engineering becoming the next capability to master. McKinsey writes, leading companies are moving quickly to master agentic engineering. They are ingesting unstructured data, extending their AI platforms with agentic capabilities, automating guardrails and controls, and rapidly experimenting to codify what works into a repeatable agentic playbook. What's interesting and what you'll see in a minute is that I think that this is one area where even the leading enterprises are a little bit behind the absolute frontier companies in general because constraining this to think about agentic engineering as simply a domain of software development is increasingly going to be its own constraint. Luckily, McKinsey's 12th bullet point is a potential solve for that, relearn like your business depends on it. They point out that the half-life of skills is shortening, and when combined with number six, speed is the defining organizational advantage. You can see just what a ride enterprises are in for in the years to come. Now, one valuable piece of the discourse comes from George Sivulka, who recently wrote an opinion piece for A16Z called Institutional AI versus Individual AI. His core thesis is that while AI has made every individual 10x more productive, no company has become 10x more valuable as a result. In the piece, he argues that there are seven pillars of institutional intelligence that show just how different individual AI versus institutional AI really is. And the biggest takeaway from George's piece is that institutional AI is not just a matter of aggregating the impact of individual AI. It is instead a different set of processes that make individual AI's value point in the same direction and that solves problems that individual AI creates. The first example he gives is around coordination. He writes, Imagine you doubled your organization's headcount tomorrow with clones of only your best employees. Each of these employees have minor differences, predilections, quirks and perspectives. Especially true if they're your best employees. If they're not sufficiently managed, if they're not sufficiently communicating, if their swim lanes, OKRs, roles and responsibilities are not well defined, you've created chaos. The organization, while measured on an individual basis, may be more productive, but thousands of agents or humans rowing in opposite directions creates a standstill at best and destroys organizational harmony at worst. George continues, This isn't hypothetical. It's happening right now in every organization that's adopted AI without a coordination layer. Every employee has their own ChatGPT habits, their own prompting styles, their own outputs that don't talk to anyone else's outputs. An org chart might exist, but the actual flow of AI-generated works has something else entirely. Coordination then becomes an absolute imperative for humans and agents alike. In another example, he talks about how institutional AI has to find signal through the noise and the massive expansion of content that's being generated. He writes about how institutional AI has to create a certain category of professional objectivity that can interact with individual AI's overalignment, harkening to the PwC study that we were just discussing. If individual AI cares about saving time, institutional AI cares about scaling revenue, and so again, you're not just adding up a bunch of time saved, you have to actually harness it in specific ways. Now maybe the most interesting thing to me is that when you get beyond not just the companies that McKinsey is profiling that are leaders, but companies that are actually completely reinventing themselves in AI native ways, you're starting to see some patterns emerge for how the companies are using AI to create this institutional AI organizing function in ways that are actually producing significant results. Let's look at the case study of Ramp. Ramp co-founder Eric Gleeman recently tweeted, 99% of Ramp uses AI daily, but we noticed most people were stuck. Not because the models weren't good enough, but because the setup was too painful and unintuitive for most. Terminal configs, MCP servers, everyone figuring it out alone. So we built Glass. Every employee gets a fully configured AI workspace on day one. Integrations connected via SSO, a marketplace of 350-plus reusable skills built by colleagues, persistent memory, scheduled automations. When one person on a team figures out a better workflow, everyone on that team gets it and gets more productive. The companies that make every employee effective with AI will compound advantages their competitors can't match. Most are waiting for vendors to solve this. We decided to own it. Earlier this week, I did a show about Harness Engineering, and this is effectively Harness Engineering at an organizational level. Seb Godagin, who runs internal AI at Ramp, wrote a post about this that's been seen now a million times. And the key thesis is about harnesses. The models he writes are good enough, the harness isn't. The setup is the same context Eric gave, so we'll dive into the essay. In the next section, everyone can be an AI power user. Seb writes, The models are already exceptional, but most people use them like driving a Ferrari with the handbrake on. Not because they aren't smart or lack ambition. They've just never seen what a well-configured environment looks like or what it can do. Solved this problem, we aligned around three core principles for Glass. Glass is the name of this internal system. One, don't limit anyone's upside. The default approach, Seb writes, for non-technical users is to simplify. Put the product on rails, offer fewer options, and make it dummy-proof. We couldn't disagree more. At RAMP, power users thrive on multi-window workflows, deep integrations, scheduled automations, persistent memory, and reusable skills. The goal isn't to remove complexity, but to make it invisible while preserving full capability. Now, I want to pause here, because I don't think that I can stress how much I am also seeing this, and I think it breaks a lot of our conventional wisdom in important ways. Every time I drop some program, like Claw Camp, I am reminded that AI is this very different paradigm, where people are not looking for the simple, easy, or dumbed-down way to do things, but are instead looking for the full capability set that can unlock totally transformational approaches. Now, I think part of what makes this different than previous categories of software is that even when you are sitting there by yourself, in your office or at your desk trying to figure things out, you are not actually alone in that pursuit. Because of AI, we have the world's greatest tutors, coaches, and build partners sitting there just waiting to help us work through whatever problems we encounter. So the difference is that where in the past you couldn't necessarily support every user inside your organization to take on the hardest challenges because you simply didn't have the human resources to help them work through their problems at any sort of reasonable pace, the fact that they can now partner with the AI itself to work through those problems totally changes the equation. What Ramp is showing you is that this can turn into a design principle that changes the shape of what a great AI using organization looks like. I think this is going to be one of the most important mental resets. It is not that in your large enterprise, you should throw everyone into the deep end of the pool all at once. But you also should not assume that there are some people who are going to be just chat users and some people who are only going to be co-work users and some superstars who get to use quad code. That might be the pattern of skill acquisition they follow. But you should not limit anyone's endpoint to only one type of those interfaces because of some pre-existing technical limitation they have. Now, continuing back into Seb's piece, and again, we're in the section, Everyone can be an AI power user. He writes, two, one person's breakthrough should become everyone's baseline. The biggest failure mode wasn't that people couldn't figure things out. It was that everyone had to figure things out alone. A workflow discovered by one person didn't help anyone else. Third, the product is the enablement. Becoming an effective AI user is a skill. People improve through repetition and experimentation, but the product can accelerate that curve by suggesting the right skill at the right time and showing what good looks like in the moment. No amount of workshops can match a targeted nudge while you're already doing the work. Okay, so if that is the mindset shift, what is some of the structure of this Glass tool that maybe you as a listener can imitate in your organization? The next section of Seb's post is everything connects on day one. He writes, Glass comes auto-configured on install, people sign in once, and all Ramps tools become available to them with a one-click setup. This includes homegrown products like Ramp Research, Ramp Inspect, and our newly released Ramp CLI. This is the unsexy foundation that makes everything else possible. When a sales rep asks Glass to pull context from a gong call, enrich it with Salesforce data and draft a follow-up, it just works because everything is already connected. And so now, stepping back to NLW here and outside the essay, we've got an example of organization-level context engineering. They have designed their harness to take advantage of the full organizational context by integrating all that context as the default state for any employee who's interacting with that harness. And importantly, the harness evolves. This is that idea that when one person figures out some skill, by which in this case we are referring to agent skills, mark down files that teach your agent how to perform a specific task, those skills can find their way into an internal marketplace that Ramp calls Dojo. An example they give is a CX engineer building a Zendesk investigation workflow that pulls ticket history, checks account health, and suggests resolution path, which now, through this Dojo marketplace, gives the entire support team the ability to have access to that skill as soon as the originator of the skill makes it available. Now with 350 skills already shared, that's a lot to wade through. And so of course, they used AI once again, building an in-AI guide, which they keeping with the theme of Dojo called the Sensei, that looks at the tools that people have already connected, what role they're in, and what they've been working on to recommend skills that are most likely to be useful to that particular person. A new account manager, Seb Reitz, doesn't need to browse a catalog of 350 skills. The Sensei surfaces the five that matter most on day one. What about memory? We talked about context and context engineering in terms of the organization as a whole, but what about on an individual level? Seb writes, when users first open Glass, we build a full memory system based on the connections they have authenticated. This gives every chat session context on the people they work with and their active projects along with references to relevant Slack channels, Notion documents, linear tickets and more. As a result, the agent spends less time searching, entering each conversation with the context the user expects. Now, this is one of those areas where, especially if you're thinking about imitating the system, I think there's the most hidden technical complexity. Seb continues, under the hood, we also run a synthesis and cleanup pipeline every 24 hours, mining users' previous sessions and connecting tools like Slack, Notion and Calendar for updates. This means Glass can adapt to their world without them having to re-explain things every session. This is one of those things that's easier to write about than to build even with agentic coding. Managing memory and context is kind of the whole game right now, and there's a reason why it's a problem that so many people are still working on. Now, one other interesting piece about this system that Ramp has built is that it takes a bunch of the features that people have looked to from OpenClaw and brings them directly into the company's agentic operating system. For example, Seb writes that employees can schedule automations to run daily, weekly or on a custom cron that can then take actions like posting directly to Slack. This, by the way, is also some of the new features that we saw coming into Codex in that episode from a couple of days ago. Importantly, employees don't actually have to be on their devices for this to happen. There's a bunch more, but I thought that this conversation about why they took the time to build this might be relevant for those of you organizational listeners who are trying to decide whether some version of this type of effort is worth it for you. Seb writes, The obvious question is why not just buy this? And argues that there are three reasons that they built it in-house. One he writes, Internal productivity is a moat. Using AI well is now a core business need. The companies that make every employee effective with AI will move faster, serve customers better and compound advantages their competitors cannot match. That makes internal AI infrastructure part of your moat, and you do not need to hand out to a vendor. Two, speed. When you own the tool, you see exactly where people get stuck. You can ship fixes the same day someone reports a problem. We have a Slack channel where users report issues, and our team triages them into tickets automatically, with most resolved in hours. You cannot do that while waiting on a vendor's roadmap. Three, it directly informs our external product. Ramp is an AI first company building products for finance teams, and many of the problems we solve for internal users translate directly to customers. How do you build memory that actually helps? How do you enable people to build, distribute and maintain effective skills? How do you surface functionality through usage? Solving these problems internally gives us conviction about what works before we ship it. Glass gives us reps on the hardest AI product problems without those reps happening at customer's expense. Now, I think if you try to take this argument together, the subtext of what Seb is arguing is something that I agree with wholeheartedly, which is that AI use is now a core primitive of organizational operations. It is not something that can be outsourced or externalized because it is functionally core to how we work going forward. Taking the time to build this then is actually doing double duty. It is building the new muscles that need to be built while also building the system that can help those muscles flex to actually do things that the business is tasked to do. The last section of Seb's piece is called What We Learned, and it's basically an argument for learning by doing. He writes, the people who got the most value weren't the ones who attended our training sessions. They were the ones who installed a skill on day one and immediately got a result. The product taught them faster than we ever could. This realization reframed how we think about the entire project. Every feature in Glass is secretly a lesson. Skills show you what great AI output looks like before you know how to ask it for yourself. Memory shows you that context is the difference between a generic answer and a useful one. Self-healing integrations show you that errors aren't your fault, the system has your back. None of this was designed as education, but it turns out that when you hand someone a tool that just works, they learn by doing and they learn fast. Seb concludes, This is what excites me most about what we're building. Not the product itself, but what happens to an organization when the floor rises for everyone at once. When a new hires for a session in Glass already knows their team, their projects and their tools. When someone who's never opened a terminal is running scheduled automations that would have required an engineer six months ago. The compounding is real and we're only at the beginning of it. We don't believe in lowering the ceiling. We believe in raising the floor. And this is what I was saying about McKinsey's point about agentic engineering becoming the next capability to master. It's not that I think that that's wrong. I think that it turns out that for the real leading organizations, agentic engineering is kind of the work of everyone now. Now luckily, not everyone is going to have to build their entire own systems like this. When a little over a month ago, Cloud Code launched Scheduled Tasks, entrepreneur Ryan Carson wrote, It's exciting seeing all the big model labs launch automations. It's exciting to see everyone moving towards code factories. I think we'll probably have complete solutions from all the big players by the end of the year that allow you to create an end-to-end code factory for your startup without having to jump around to different tools or duct taping things together. And as if to validate Ryan's point, Guillermo Roche, the CEO of Vercel, just announced that the company was open sourcing a reference platform for cloud coding agents to help others build the type of systems that Stripe, Ramp, Spotify and Block are all building. So to sum up, how the best companies that using AI are using AI. First, it is about growth and opportunity and new business and growing the business, not just about efficiency and productivity. Second, they are recognizing that AI success involves creating systems, systems that can coordinate the outputs of individual AI users, systems that can provide context, and increasingly as we see from the example with Ramp, systems that actually build harnesses around AI use that aren't content with dividing the employee base into good users, medium users and bad users of AI, but which fundamentally believe that everyone can be a great AI superuser. It's a mindset shift plus some specific activities, and I'm very excited to see more companies diving in and sharing what they find along the way. For now though, that is going to do it for this long read slash big think edition of the AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace!