The Next Wave of Enterprise AI

Today on the AI Daily Brief, the next wave of enterprise AI is upon us. Before that, in the headlines, the very confusing and weird process around the latest AI executive order. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors. KPMG, OutSystems, ZenCoder, and Bolt. To get an ad-free version of the show, go to patreon.com/aiDailyBrief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. Today we begin with the latest in the saga of this Trump AI executive order. This is just one of the absolute strangest policy processes I've seen. So what's going on? How did we get here? And what was actually signed? First of all, by way of context, the reason that this is coming up at all is a couple parts. Firstly, there are some very, very different and contentious groups when it comes to AI, including in Trump's own coalition. Republicans like Governor DeSantis in Florida, as well as very loudly former presidential advisor Steve Bannon, have been squawking quite loudly about AI and more broadly decrying Trump's close alliance with the technology industry for some time now. And yet the specific catalyst for this new round of policy discussion was the cyber capabilities of Anthropic's Mythos model. So the executive order we started hearing about a few weeks ago seemingly had something to do with labs needing to give the government access to their most advanced models before actually releasing them. Indeed, that was the core policy of the draft that was circulated 2 weeks ago that seemed at the time like a done deal. A signing ceremony had been scheduled, a who's who of tech CEOs had been invited to attend. However, hours before the event, President Trump pulled the order, stating, I didn't like certain aspects of it and adding that he thought that it would get in the way of the US lead over China in the AI race. Now, it later surfaced that former AI czar David Sachs had intervened at the 11th hour, placing a call to the president to talk him out of signing the policy, at least for now. The order that was signed this week is substantially the same as the draft order that was scrapped a couple of weeks ago. Both versions of the order made safety testing voluntary, although in the current climate, that's not all that meaningful a distinction. All major AI labs have agreed to submit advanced models for testing, and while some White House personnel were reportedly pushing for compulsory testing, it appears that that position never made it into a draft. Indeed, it seems like the only significant change is that companies are encouraged to make their models available 30 days prior to public release, as opposed to the draft order which had asked for a 90-day period. It was that 90-day period more than anything else that triggered industry backlash for its potential to significantly slow down the release cycle. Neither version of the order provided any mechanism for the government to block a model's release. In fact, one subtle change in the new version is an inclusion of a disclaimer which reads, Nothing in this section shall be construed to authorize the creation of a mandatory government licensing, preclearance, or permitting requirement for the development of new AI models. This sounds like a direct response to the critique that many had had that what the White House was doing with this executive order was a de facto licensing regime. Functionally, however, the policy just allows the government to assess new capabilities before they're available to the public. The NSA has been assigned primary responsibility for model testing with support from various cyber technology and defense agencies. In addition to safety testing, the order establishes a cybersecurity clearinghouse run by the Treasury in consultation with the NSA, the Department of Homeland Security, and the Cybersecurity and Infrastructure Security Agency. There's also provisions instructing civilian and military agencies to harden systems against AI-driven cybersecurity risk. Outside of the 90-to-30-day switch, the other biggest difference with this version of the order is the way that it was presented to the public. Rather than a high-profile signing ceremony, the order was signed in private with zero fanfare. I don't think it's an unreasonable interpretation to see this as the administration treating this sort of AI safety regulation as some version of eating its vegetables rather than the Big Mac or the well-done steak that it would prefer. The order does contain some language reaffirming the administration's commitment to AI acceleration, talking about a commitment to the United States' AI global dominance. But ultimately, this was the sort of Rorschach test policy that gives everyone something to comment on, everyone something to claim victory around, while ultimately doing very little. The New York Times reported the order as, in their words, signaling a shift from the hands-off approach the White House had previously taken toward AI. The White House Office of Science and Technology Policy, however, called this lazy and inaccurate reporting from the New York Times, trying to draw a line in the sand around the difference between oversight and voluntary sharing. The EO, they wrote, creates a process for frontier labs to voluntarily share cutting-edge cyber models in order to secure critical infrastructure and strengthen the government's own cyber defenses. We are not conducting oversight of all new models, as that level of government overreach would have chilling effects on free speech and innovation. David Sachs himself chimed in to explain that the policy is intended to only cover models that, in his words, represent, quote, meaningful step change in cyber capabilities, i.e., mythos, not incremental changes to existing models like Opus 48. He also took the chance to comment on the slippery slope argument, writing, I understand the concerns of many that this could morph into an FDA for AI. Of course, bureaucratic mission creep is always a danger, and this should be closely monitored, but the EO expressly forbids the creation of a new licensing, preclearance, or permitting regime. From the labs, all of the commentary was pretty similar. Clearly talking points were circulated, suggesting they all use the language of this being a, quote, important step. Former White House advisor Dean Ball was concerned about the implications of this first step, calling it a, quote, fairly major win for the safety contingent within the administration and a significant loss for the Sachs accelerationist wing. Now Dean, despite Sachs' assurances to the contrary, thinks this heads in exactly one way. He writes, This is clearly teeing up the infrastructure for a model licensing regime, and the fact that the administration is classifying the details of how this voluntary system will work is egregious. The public and the employees of the labs have a right to know how this works. Most lab staff don't have clearances, but if the literal regulatory thresholds that trigger pre-deployment review are classified, researchers themselves won't know whether what they are training is regulated by this EO. All for a benefit that is barely articulable? What exactly is the intelligence community going to do in 30 days to make the models safer? It's not a huge mistake, but a small to medium-sized one. But I am fairly confident this is a mistake nonetheless. And again, while Sachs says this isn't a step to more, the more safety-minded certainly see this as a crack in the Overton window that they can pry open. Right-wing pundit Steve Bannon said, "For the first time, it's on a piece of paper, a structure, and a process. That process is still pretty ill-defined. It doesn't meet our requirements. But as I tell people, we're going to eat the elephant one bite at a time. I strongly believe we're heading towards mandatory within the next couple of months. We intend to ramp up the pressure campaign.". And showing how weird AI makes different political bedfellows, Bernie Sanders seems to want the same thing as Steve Bannon, saying, after calling efforts to regulate AI foolish, Trump finally acknowledged AI poses a real threat. That's the good news. The bad news? His executive order is voluntary and does almost nothing to protect Americans. Congress must act. Ultimately, the reason that everyone is speculating around what the implications are is that the order itself doesn't do all that much. The AI labs already have agreements in place to share new models with the government ahead of release. David Remler from the Center for a New American Security said that the order, quote, effectively formalizes what has already been happening between the US government and the leading AI companies. And so speculation about what comes next is the natural next place to go. Certainly there will be a lot to watch here, but for now we actually do still have a few more headlines, one of them from the very project which got this whole ball rolling, which is Project Glasswing, the specific way of course that Anthropic is releasing their Mythos model. Anthropic has just expanded access to Mythos, adding 150 new partners. With this new announcement, Anthropic is rolling Mythos out to firms across 15 countries, with the expanded group of partners including new sectors that weren't covered by the initial project, including energy, water, communications, healthcare, and computer hardware. Writes Anthropic, what each partner has in common is that a successful attack on their codebase could be catastrophic. For most partners, we estimate that a major attack could affect more than 100 million people, with important ramifications for both global and national security. Now, the announcement also included some further discussion of a public release. You might remember that during the rollout of Opus 4.8 last week, Anthropic said that they expect to have a Mythos-level model ready in the coming weeks. In Tuesday's Glasswing update, they wrote, We're working as quickly as we can to safely release Mythos-level capabilities in general access. To do so, we'll need highly robust safeguards that prevent the model's cyber capabilities from being misused. Safeguards that we, and to our knowledge, all other AI developers have yet to develop. Because cybersecurity has both helpful and destructive uses, making safeguards that are both strong and precise enough is a major challenge. Which to me honestly kind of feels like walking back the language that they had used in the Opus 4 announcement. All the way back then last Thursday, they said it was coming in the next couple of weeks, but now they're saying they need infrastructure that doesn't exist. Honestly, the messaging is about as confusing as the government's executive order. Meanwhile, The Information checked in with some of the teams working with Mythos, finding that although the model is powerful, it is also eye-wateringly expensive. Most of the testers are finding themselves running through millions of dollars worth of tokens very quickly. And what's more, for now, Anthropic is subsidizing use, so firms aren't even paying the full cost. At the same time, it also appears valuable enough to justify that cost, with many of the firms that The Information talked to saying that they're basically aligning their budget so that they can build their strategy around Mythos when it becomes more broadly available. And lastly today on that same theme of the token shortage, SK Hynix now plans to double their manufacturing capacity for memory chips to help address the global shortage. This year's rapid growth in token use has led to shortages throughout the AI supply chain, with one of the more prominent shortages being in memory chips. The cost of high-bandwidth memory for AI servers has more than doubled so far this year, and up until now, memory manufacturers have been reluctant to build new plants to boost supply. In the past, cyclicality in chip pricing has punished long-term investment, with new plants often missing the window of peak demand. SK Hynix's new plan suggests that they now view AI-driven demand as a structural change, and they're deploying capital to take advantage of it. Now, to be clear, they're talking about doubling capacity by the end of the decade, so this is unlikely to do much for the chip crunch in the short term. Indeed, Chairman Chey Tae-won told reporters that the shortage could last until 2030. And yet still, the deeper investment is the right policy, with Chairman Chey arguing, "The whole AI industry needs to be more sustainable. We have to continue to grow, but sudden jumps in price can become a problem and actually hurt sustainability." So lots of big movement today, but that is going to do it for the headlines. Next up, the main episode. One of the most important AI questions right now isn't who's using AI, it's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers, They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com/us/sophisticated. That's kpmg.com/us/sophisticated. One thing I keep seeing in enterprise AI companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends. The teams actually shipping, they've picked a lane and they move fast. That's one of the reasons I like today's sponsor, Robots and Pencils. They've gone all in on AWS. They're an advanced tier and AWS pattern partner, and they ship production AI coworkers in 45 days. That's led to them doing some of the more interesting work I've seen on AI coworkers. And by that, I'm not talking about chatbots. I'm talking about actual agentic systems that sit inside a business architecture and do real work. That kind of focus matters if you're an enterprise leader trying to get something real into production or an AWS rep trying to move a customer from interested to deployed. Request an AI briefing at robotsandpencils.com. One conversation with Robots and Pencils and you'll know. You know AssemblyAI for having the most accurate streaming speech-to-text out there, but they just went a step further and launched a full voice agent API. The idea is simple. One connection and they handle everything. The listening, the thinking, the speaking. You just stream audio in and get your agent's voice response back. We're talking about things like outbound sales calls that actually qualify leads, customer support that handles complex requests without a script, scheduling agents that sound like a human assistant, and you can build one in 5 minutes with one API. And importantly, their streaming model is the best at catching all the stuff that breaks on other voice agents, things like phone numbers, emails, names, and medical terms. And for those of you who are still in experimentation mode, there are no contracts and unlimited concurrency, so you can actually test it out without any friction. Head to assemblyai.com/brief and try the live voice agent demo right there on the site. No signup needed. This episode of the AI Daily Brief is brought to you by OutSystems, a leading agentic systems platform built for the enterprise. Organizations all over the world are building, orchestrating, and governing agentic systems on the OutSystems platform, and with good reason. OutSystems' open and unified platform allows teams to architect, deliver, and scale governed agentic systems with agility. Teams of any size and technical depth can use OutSystems to build, deploy, and manage AI apps and agents quickly and cost-effectively without compromising reliability and security. With OutSystems, you can rapidly launch ideas from concept to completion. It's the leading agentic systems platform that is unified, agile, and enterprise-proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI. OutSystems. Build your agentic future. Welcome back to the AI Daily Brief. Yesterday we had two dueling events, both focused on enterprise AI. One was from OpenAI and one was from Microsoft, and they each provided in their own way some indications of where enterprise AI is currently and where it's headed. Now, the context for this, of course, is the broader shift we've been discussing on this show of moving from the subsidy era of AI to the scarcity era of AI. Or the token shortage era of AI. The basic idea is that as we move from assisted to agentic workloads, the sheer quantity of the AI tokens we use goes up, and we're now running into the limits of what the available compute and physical infrastructure can produce, meaning that business models are realigning, costs are going up, and everyone's scrambling to figure out how to adapt to all of that. In the meantime, though, part of what makes this challenging is that it's not at all clear to most organizations how to best use this new set of tools. In other words, the question of enterprise AI adoption is not just a question of costs, but also one of tool and use case fluency. And increasingly, enterprise users are living inside the power tools like Cloud Code and Cowork and OpenAI's Codex. Interest in Codex has been surging for a while, with Google searches for Codex actually spiking past Cloud Code for the first time in May. The Information wrote about how the, quote, vibe shift on Codex has been palpable. And OpenAI's event yesterday centered on a set of new updates for Codex that are all about it moving out of the strict realm of the developer into the broader world of knowledge work. Now, alongside the event, OpenAI released a report called The Next Era of Knowledge Work. And the TL;DR on the thing was not only that Codex was growing, hitting 5 million weekly active users, but that the biggest source of its growth was not developers, but non-technical knowledge workers who are now adopting Codex at a 3 times faster pace than developers are. And one of the things that's interesting about the report is that it's not just a bunch of reported stats, but actually shows quite a bit of the design philosophy and the first principles understanding that's going into how OpenAI is thinking about Codex. One of the central themes is what OpenAI calls a strange abundance. Modern workers, they write, can produce documents, messages, dashboards, models, and presentations faster than ever, yet they spend a remarkable share of their time looking for context, reconciling conflicting versions, waiting for responses, and moving information across systems. They point to a McKinsey study that found that the average knowledge worker right now spends more than a quarter of their work week managing email and almost a fifth of it looking for internal information or trying to find people who can help within their company around some specific task. Overall, they say 3 frictions define the daily cost of knowledge work. The first is the cost of finding relevant inputs across, as they put it, sprawling, untransparent systems. Second is information coordination costs, and third are approvals and verifications. In fact, they argue that these frictions are what accounts for the delays between a new technology being introduced and it actually showing up in the productivity statistics. Knowledge work, they write, is still waiting for its factory redesign. Previous generations of workplace software lowered the cost of producing intermediate artifacts but did not reduce the attention required to consume them. Email made correspondence cheap, then multiplied correspondence. Docs made drafting cheap, then multiplied drafts and review cycles. The result is an excess of documents and tools and even scarcer time and attention. And you might be seeing where they're going with this. Codex, they write, is that factory redesign. So what are they seeing in how people are actually using Codex? First of all, everyone is producing artifacts. 72% of knowledge workers using Codex are producing some sort of artifact, be it a PDF or a spreadsheet or something else, on a weekly basis. Outside of coding and software engineering related tasks, They're also doing research, 41%, data analysis, 27%, as well as implementing what they call business function workflows at 15%. Importantly though, people are doing a lot of these at the same time. The most consequential shift in behavior, they write, is towards parallel tasks. Roughly 50% of users now have more than one Codex task running simultaneously at some point during the day, up from less than one-third in mid-April. The shift, they write, from sequential to parallel use is what lets a single knowledge worker operate at the scale of a small team. One turn to inspect a dataset, another to draft a script, another to assemble a report, another to check an application. The user becomes the orchestrator of workstreams rather than executing a single task at a time. So what goodies did we actually get? The three highlights are annotations, plugins, and sites. Annotations are effectively a more precise way to interact with context. Within Codex, when you're looking at a specific document or artifact, you can highlight rather than having to explain with words the specific part of the document that you want to discuss or query about or change. You can use the annotations feature to select just that part of the document for the model to reason over within Codex. Simon Smith from Click Health wrote, you can already use annotations to give feedback on websites, but now it looks like that interaction model is expanding across outputs. I love working with Codex by selecting things in the preview pane, adding them into chat context, and then talking to Codex about them. This makes that way of working more powerful. Next up was an expansion of Codex plugins. Now previously plugins were a way to connect specific software into the Codex ecosystem, but with this latest update, Codex is adding a set of role-specific plugins for common functions including sales, data analytics, creative production, product design, public equity investing, and investment banking. Now, given the IPO horse race dynamics and competitive storyline between Anthropic and OpenAI, this is the update that a lot of mainstream media focused on as it resembled to them Anthropic's strategy of releasing a set of tools for specific functions and industries as well. In their announcement, OpenAI writes, each role-specific plugin bundles the relevant apps, skills, instructions, and workflows. Across these 6 new function plugins, For instance, they include access to 62 apps and 110 skills, basically about 10 apps and 20 skills per role. You can almost think about the role-specific plugins as organized bundles of features that were already available, but presented in a way that requires much less setup. Another way to think about it is that this goes a long way to productizing best practices. You can think about it kind of like this. If you took the best user across each of these 6 functions from a wide variety of companies, and you looked at the app integrations and skills they most often drew from, and then turn those average set of skills and app plugins into bundle, that would effectively be what Codex is releasing here. And interestingly, this becomes sort of product-led education where the salesperson, for example, who now has access to the plugins and skills that are used by the salespeople who are best at getting the most value out of Codex, can start to imitate those best practices by virtue of what's being presented in this functional plugin. Simon Smith again notes, plugins seem to follow what Anthropic is doing with plugins focused on different business domains, but what's interesting is that OpenAI plugins seem like they can do more than provide instructions and connectors. They can add interactivity inside the Codex preview pane, like buttons buttons and guided actions that make powerful workflows more clickable. Still, the update that I'm most excited about, and one which at some point I might do an entire operator episode on, is the new Sites feature. My guess is that a lot of you have had the experience at this point of realizing as you're going about your normal work that something that you might previously have done as some sort of static document might now be better suited to presenting as some sort of small website. For example, instead of some PDF presentation, maybe you just send them a URL, It's easier to share because they don't have to download anything, plus you can update it as makes sense. Codex Sites productizes that type of behavior. It allows you to turn any sort of artifact that you've built inside of Codex into a full website or web app that's shareable with your team. They give the example of a revenue forecast planner that represents a sort of much more interactive way to look at budget planning than a traditional spreadsheet or presentation might have been, an event operations dashboard, and a product launch hub with both the event operations dashboard and the product launch hub representing a way to keep track of operational progress with a highly customizable set of inputs. Rounding out his analysis of these three, Simon Smith again writes, "Sites are kind of like Claude artifacts but on steroids. This puts vibe coding even more directly in the hands of everyone in an organization. You can build stuff, share it, deploy it, and importantly do it in a more secure way, which has been a real issue with some internal vibe-coded tools." Now I think that Simon's analysis is right, but I think this is where we have a terminology problem. Part of why vibe coding never really fit for this type of use is that these sort of sites are effectively disposable software and web apps. They're meant for a specific purpose for a specific set of time. And the only thing that they have in common with software engineering is that they use code to deliver an output. But this isn't non-coders all of a sudden becoming product designers and engineers and trying to get in on the product building game. This is people using code and websites to improve how they share things and collaborate with colleagues. My argument would basically be that in the same way that building a slide deck or writing a document or interacting with a spreadsheet is a core knowledge work primitive, Building websites and disposable web apps is also going to be a core knowledge work primitive going forward. Codex Sites is a hyper-simple version of that experience that's going to make that primitive much more accessible to a large number of people. Like I said, I actually think that Sites might be deserving of an entire Operators episode to dig into different types of things that people might be able to do with it. But if you had just one thing to play around with in the short term, that's where I'd be looking. It's very clear that OpenAI and the Codex team see the Codex app as the new interface for knowledge work and are going to continue pushing to figure out all the implications of what you can do to do in this new type of environment. But as I said at the beginning, the question of the next phase of enterprise AI is both one of interface, which we've been discussing with Codex, but it's also one of efficiency and cost management. Uber, a company that has somehow found itself in the headlines as Exhibit A in the changing tides of agentic AI, has now put a $1,500 monthly cap across token spending for all employees. Now, I have a lot more to say about what I think does and doesn't work about that strategy, but we'll save that for another episode. The point for us today is that cost management is at another vector of the next wave of enterprise AI is to be cost management. And interestingly, that seemed to be at the core of the announcements from Microsoft Build. Nominally, the big announcement was 7 new Microsoft AI models: Image 2.5, Image 2.5 Flash, Transcribe 1.5, Thinking 1, Voice 2, Voice 2 Flash, and Code 1 Flash, a family of models that were optimized around different sets of use cases. And certainly, just like any other time that we get model releases, there was a bunch of discussion of the benchmarks. The headliner was MAI Thinking 1, a 1 trillion parameter model using a mixture of experts architecture for inference optimization that Microsoft tried to place as a model somewhere in the SONNET-46 to OPUS-46 type of range. Now to some, the discussion was just about Microsoft making progress in the model training game at all. Wrote Sean Wang, you have to give Microsoft props for training all these in-house models from scratch and getting all of them to near state-of-the-art. Mustafa Suleyman built a full-fledged NeoLab inside Microsoft in 2 years that Microsoft now fully controls from chip to model to harness. Absurdly impressive. Prime Intellect's Ellie Bakausch writes that Thinking-1 uses zero synthetic data or distillation from previous models. This means reasoning, agentic behavior, tool use are all learned fully during post-training with no cold start. Bold choice that makes it harder and requires more iterations to reach state-of-the-art, but you get full control over your model series and it proves they are serious about being a frontier lab. Ethan Mollick lamented the fact that no one really has gotten their hands on these things, so we're just left to squint at the benchmarks, which themselves are confusing. He writes, it's difficult to know how good MAI Thinking-1 is from the scores alone. Own, like weirdly low GPQA and TerminalBench 2.0. But Microsoft makes it really hard to try its models upon release, so I dunno. Others, like leaker I rule the world, poo-pooed the releases saying, "In case it's unclear, the Microsoft model isn't competitive, particularly not for anything agentic." And indeed, Thinking One's scores on the agentic coding tests like TerminalBench 2.0 and SuiBench Pro were meaningfully lower than competitors even one generation ago from Anthropic and OpenAI. But go one step deeper and it's quite clear that Microsoft is playing a different game. I believe that they have very clearly identified cost optimization as an issue and believe that their approach can be part of the answer. In his announcement post, Mustafa Suleyman wrote, all of this is the foundation for Microsoft Frontier Tuning. It lets you customize our models to create custom company-specific agents that only you control. Early adopters are already seeing a difference. When we tuned our models for McKinsey's tasks, MAI delivered the highest win rate, outperforming GPT-5.5 on quality while being 10x lower on cost. On stage, Microsoft CEO Satya Nadella called this a pretty significant shift. He said, we believe the time has come for every company to just move from consuming a frontier model to fully participating at the frontier in the frontier ecosystem. In other words, I don't think that we should be looking at this series of models completely in raw terms as something where one of us as listeners is going to decide to fire up MAI Thinking 1 instead of GPT-5.5 or Opus 4.8. Instead, they are very self-consciously being positioned as part of an overall strategy to not only get state-of-the-art performance, but to do so at a lower cost. And given that Microsoft already has the strongest distribution inside the enterprise of any company, their play here is worth taking seriously. If you want to simplify it, when it comes to enterprise AI, the second half of 2026 is going to be about wrestling into a workable, cost-effective approach all of the opportunities that the first half of 2026 unlocked. In different ways, both OpenAI and Microsoft showed off big plays yesterday to those ends, and I certainly don't anticipate that's the last we'll be hearing about. And it's very clear that the race for the next wave of enterprise AI adoption is fully on. For now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.