6 Questions Every Enterprise Has to Answer About AI

Today on the AI Daily Brief, six questions shaping enterprise AI. Before that in the headlines, Sam Alman goes to Washington and the conversation has gotten a lot more complicated over the last week. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors KPMG, Blitzie, Retool and Air Table. To get an ad-free version of the show, go to patreon.com/aideallybrief or you can subscribe and Apple Podcasts to learn more about sponsoring the show, send us a note at [email protected]. And to bone up on your AI skills with the last month or so of summer, go check out our latest free self-directed education program, this one being a Choose Your Own Summer Adventure you can find it at summeradventure.ai. While Sam Alman has arrived in Washington to meet with lawmakers and White House officials and when the trip was set at the beginning of last week, the agenda was pretty simple. Alman would brief Washington on the capabilities of OpenAI's new model and discuss a protocol for release, hopefully avoiding a repeat of the Fable and GBT56 rollout. Since then, however, we've had the OpenAI hugging face hack, a public debate about open weights models, and an intention-getting petition for the government to step in and build the capability to slow down the pace of frontier AI. In other words, conversations have become a lot more complicated for Alman in just a couple of weeks. According to reports, Alman met with Senate Commerce Chair Ted Cruz and several Democrat senators on Wednesday, but we got very little information on what was actually discussed. Speaking to reporters, Alman declined to state when or even whether the model being previewed would be released, commenting, "Not sure, that's the part we're here to talk about." Alman also declined to discuss the new capabilities of the model that give cause for concern. Now, of course, the hugging face incident looms large over this visit, but it increasingly appears like the model at the center of that controversy will not see release. In a Tuesday update to their post-mortem blog, OpenAI said that the model was an internal only research prototype never intended for public release. In Washington, Alman told the press that the model has now been permanently deactivated and is inaccessible even for internal research, meaning presumably it's not the model being previewed to lawmakers this week. Now, Sam said that he and Ted Cruz had not discussed specific legislation, but that quote, "We talked about our new model and what it's going to take for America to remain competitive with AI." Alman also said that he didn't support mandatory safety testing, particularly because it could introduce an unnecessary burden on open-weights model developers, but added, "For frontier models at new levels of capabilities, we think it's really important that the federal government has great testing capacity and capabilities." Alman said he plans to meet with a range of other officials to end the week, including White House chief of staff Suzy Wiles, who, for whatever reason, has wound up as one of the key decision-makers on AI policy. That meeting will likely include a discussion of the voluntary AI safety testing framework, which has a deadline of August 1st. Reports state that this framework has been circulated to open AI and Thropic and Google for comment, but Alman declined to comment on the draft. In a hallway interview on Capitol Hill, Alman was asked whether he would talk to the White House about the need to decelerate AI development. Representing the views of his staff from the recent open letter, Alman responded, "I wouldn't use the word deceleration, but we talk about the need to pace it as the models get more capable, which I think isn't everyone's interest." Now, one other story that I'm going to get into in more depth tomorrow is the significant increase in revenue numbers on both the open AI and Anthropic front, but I do not want to bury that in the headlines so that will be a major topic for tomorrow. Come back for that. Suffice it to say, the CFO of open AI Sarah Fryer recently told employees that annualized revenue in July topped all of the previous quarter. One more bit of open AI intrigue. President Greg Brockman says that the company is working on an entire range of devices to give a physical presence to their chatbots. In a new interview with former Wall Street Journal reporter Joanna Stern, Brockman confirmed that open AI's hardware plans are still on track, stating that the company is building a family of devices. He wouldn't confirm the recently rumored smart speaker or any other form factors that have seen speculation this year, nor would he give a timeline beyond commenting you can expect them soon. Still, this is the clearest confirmation we've had so far that a full hardware range is still on the road map, surviving the end of side quests and an IP lawsuit from Apple. Brockman was understandably brief when talking about that lawsuit stating, "We're focused on our own development and technology." One interesting bit of competitive news, which I think sounds good for consumers, particularly those of you who are in the enterprise, without a ton of choice on which models and platforms you're going to use. Microsoft appears to be gearing up to compete more directly with open AI and anthropic, with the development of a co-pilot super app. During Wednesday night's earnings call CEO Sachin Della, confirmed the app is coming later this year with the goal of unifying the co-pilot experience for both consumer and enterprise customers. He said, "Copilot is rapidly evolving from chat to co-work to autopilots. This quarter we are bringing these co-pilot experiences together, including code in one super app. This is a major step forward and I look forward to sharing more soon." Microsoft is beginning to see open AI and anthropic as direct rivals, thanks to the capabilities of their new M.A.I. models. Della told analysts that the combination of cost and data privacy concerns gives Microsoft an opportunity to sell customers on their own cheaper models. When asked about the rolling debate about open versus closed, Della suggested the framing is too simplified. He said, "The goal is to have the firm be in control of their own destiny. We are very, very clear about the architectural design of the platform, which is you get to keep your harness separate from the model. That means any model at any given time is swappable." Now, I'm sure some of you will think that that's kind of a corporate answer, but I actually think that his assessment of how most enterprises feel is correct. In that, I don't think that most enterprises actually care ultimately about whether a model is open or closed. They care what they can do with it, what control they have, and what sacrifices around control they're making to someone else to have access to the systems they're using. Anyway, overall, Microsoft is increasingly positioning themselves not as a reseller of open A.I. or Anthropic products, but rather as a model agnostic platform offering a full range of options. Said Nudella, "Every customer wants the right model for each task based on latency, quality, cost, and compliance. We offer the broadest model catalog in the cloud with over 11,000 models, including the leads from OpenA.I. and Thropic, Mistral, X.A.I., as well as our own M.A.I. family." So along this world of all these big discussions and jockeying for position in A.I., Mark Zuckerberg has made the case for A.I. acceleration in a new op-ed in the Wall Street Journal. In an essay titled "The A.I. Future is for Everyone," Zuckerberg argued that the defining question of the A.I. age won't be whether superintelligence will exist, but who will have access to it. In other words, whether we end up in a world where superintelligence is closely held by a handful of institutions or broadly distributed to normal people. Zuckerberg wrote, "It is surprising that the discourse from many of those who are developing artificial intelligence is so filled with doom. I don't understand why anyone who believes that A.I. will eliminate most jobs and much of humanity's relevance would rush to build that future." This is for what it's worth exactly the point that I was trying to make yesterday when I was discussing what I think the "normy" response to the pacing the frontier letter would be that the only acceptable answer to why are you building A.I. is not, well, if we don't, someone else will, but instead, because we think A.I. will be awesome and dramatically better than all the risks that it comes with. Zuckerberg continued, "The notion that A.I. is so dangerous that the only safe path is an extreme concentration of power seems dangerous. Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened hasn't led to safe or positive outcomes." Zuckerberg's view is that much like previous technologies, like the internet, the best result will come from diffusing the technology freely across society. He wrote, "Rather than centralizing this power, we believe that delivering personal superintelligence to everyone is the way to answer this question. This has the potential to begin a new era of personal empowerment, in which individuals have greater freedom to pursue their interests and reach their full potential." The op-ed came as part of a press tour that linked up with Meta's new A.I. optimism campaign. In a separate interview with the journal Zuckerberg called for the U.S. government to accelerate A.I. development rather than restrict it. He argued that the benefits of broadly distributing A.I. outweigh the risks by quite a margin, adding, "I get that it's always hard to debate about the future because it hasn't happened yet, but I do think we have a lot of data points at this point and that should point us to be much more optimistic than I believe a current discourse reflects." Specifically, he warned against thinking a 30 or 60-day government review window is harmless commenting, "The field is moving so quickly that actually is quite a meaningful amount of time." Now notably, Meta is the only frontier A.I. lab that hasn't agreed to the government's voluntary testing framework. Zuckerberg also said that the U.S. government shouldn't ban Chinese A.I. in a separate interview with the financial times. Not only does he think a ban won't be effective, but he believes it would open the risk of regulatory capture and could stymie the release of open models more generally. Now Meta's A.I. CEO Alexander Wang recently said that the company will begin launching open source models again, suggesting that this isn't just hollow sentiment. Still, overall the core message is simply that more A.I. optimism is needed. Speaking with the New York Times, Zuckerberg said, "So much of the discourse from a lot of the other labs that are developing this is overwhelmingly filled with doom. There needs to be a voice or several voices that are bringing realism to this debate." Now, I think unfortunately Mark Zuckerberg's power to be the leader face of A.I. optimism is limited by history and people's fairly negative view of the overall impact of social media on society. Still, to start to have loud, sustained discourse that other people can pick up and run with is immensely important, and you better believe I will be here amplifying that message. For now, however, that's going to do it for today's headlines. Next up, the main episode. One of the most important A.I. questions right now isn't who's using A.I. It's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace A.I. interactions and found something surprising. The highest impact users aren't better prompt engineers, they treat A.I. 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 A.I. access to real capability, KPMG's research on sophisticated A.I. collaboration is worth your time. Learn more. or at kpmg.com/us/sophisticated. Blitz's deep code based understanding unlocks the thing every roadmap owner cares about, shipping new features. Here's the truth about building inside a massive enterprise code base. Writing code was never the bottleneck. Context is. Which system does this touch? Which contracts can't break? Which standards apply? Blitz already knows because it reversed engineered your entire code base into a dynamic knowledge graph before feature work began. With that complete picture, Blitzy builds features end to end. Architecture, APIs, UI, and tests all validated against your existing systems. One Blitzy customer built an AI native application from scratch with 100% autonomous completion, saving over 2700 engineering hours. Features that respect your code base instead of fighting it. Stop letting your backlog grow faster than your team. Accelerate your roadmap at Blitzy.com. That's BLI-TZY.com. This episode is supported by retool. AI made building software easier than ever. So more people are building it than ever, usually without a thought for security. Right now, people in your company are vibe coding and every ungoverned app that touches your data is a risk you own. Retool takes that risk off your shoulders. Build apps however you want. Natively with retool or with cloud code, codecs are any coding agent and ship it in a secure government environment. Security lives in the platform not any chap, so however it was built, it's governed the moment it ships. It's why teams at Amazon, Stripe, and Brex build on retool. New enterprise customers who sign up by September 30th get up to $10,000 in AI credits per year. Learn more at retool.com/aideally. This episode of the AI Daily Brief is brought to you by HyperAgent where you run fleets of agents your team can manage together. New users get $1,000 in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. HyperAgent deploys always on agents in the cloud doing real work across the tools your team already uses. Marketing's agent turns competitor moves into landing pages. Apples as agent enriches leads, drafts emails, and updates the CRM. Ops agent chases the paperwork and tracks the budget. Every agent has access to shared context and follows your rules about scope and approvals. It's time you add agents that feel like teammates. Hire yours at HyperAgent built by the team at Air Table. Claim your $1,000 in inference at hyperagent.com/aideally brief. Welcome back to the AI Daily Brief. This week I had the chance to be out in Utah with KPMG for their annual tech and innovation symposium. Now, this is the second year that I've been at the event. In each case, I've had a chance to do a similar type of presentation. This time around, both my focus and the focus of the conversation after was about trying to sum up, broadly speaking, the big questions that are currently shaping how enterprises have to think about AI. It's extremely notable to me is how much the conversation has changed since last year at this time. Now, I'm going to go through a version of the presentation I gave, but before that, I actually want to zoom back to last year. Last year I did where AI is 15 slides and 15 minutes, which was actually, as you can see, 23 slides. And looking back, it's almost quaint. What we found interesting or fascinating and what we were discussing at that event. The first thing was acceleration, and we talked about how AI wasn't just moving faster, but it was actually getting faster and the speed it was being adopted. I discussed the more than 100% growth in the total monthly tokens that Google was processing between May and July, where they reached nearly a quadrillion tokens. Now, as many of you know, a quadrillion tokens at this point is about what a single open claw left unattended will do in a month, but it was a big deal back then to see this massive inflection point. And indeed, some of the themes from that presentation were effectively set up to where we are now. The compute shortage has done nothing but get worse as we've moved to a new era. And of course, even back then, the big conversation was agents. Now, what was interesting is that, at least in the way that we use agents today, Agent AI was still firmly in the domain of the future. This was the clawed force on it, '03 type time horizon. And we were just wrapping our heads around the meter time horizon task graph that showed that AI capability was doubling every few months. Now speaking of themes that would continue to be important, even back then, it was clear that Agentic coding was the breakout Agentic use case. And again, to give a sense of just how long ago this was, we were all gobsmacked because we had hit a billion dollar revenue run rate in just about a single year. To put a fine point on how much this has changed, right before this I read a post from Door Cash that suggested that Anthropic could get to $100 or $150 billion revenue run rate this year. Now I won't go through all of these different slides, but what stands out to me reflecting back then was that the questions of AI and agents were really still for some if questions. One slide that I didn't have in this particular chart, but I know I had an longer presentation that was being given around the same time was this chart from McKinsey that showed the growth in the number of organizations that had implemented at least one or two or three different AI use cases. Yes, the big deal in mid 2025 was still that something like 40% of enterprises were up to two or three use cases. A year on the conversation has changed immensely. And the historian in me thinks it's worth reflecting how we got here. The big capability jump as we now know came towards the end of the year, the November December time period where we got Opus 4.5 and GPT 5.2. For whatever set of reasons, those were the model updates where agents in agent workflows actually came online in a major way. Now what was fascinating is that it actually took a couple of months for people to really grok that something had shifted. Everyone went home for the holiday, had a little bit of time to decompress, and when they fired up their instance of cloud code or whatever tool they were using, they found that the stuff that they could do was significantly different than what it had been before. I still remember vividly the absolute title wave of tweets in that week between Christmas and New Year's of entrepreneur after entrepreneur and developer after developer coming back gobsmacked about what they could now build that they simply couldn't be for. Now what's interesting is that this almost immediately translated into organizational practice as well. Part of this was because software organizations had been adapting to greater and greater capabilities throughout the year and even before. At the turn of 2026, we were long past software engineering organizations viewing AI coding as just an autocomplete solution and they became some of the first groups in the enterprise to actually shift from viewing their job as writing code to managing the agents that wrote the code for them. That said, maybe because the enterprise folks had been paying attention for over two years at that point, it wasn't like there was some major lag from the AI early adopters to enterprises thinking about what this new agent at capacity was going to mean for their work. Looking back into 2026, it was absolutely not just software engineering organizations that were racing to put into practice these new ways of working. You saw Vanguard builders and early adopters across domains from marketing to legal to finance, starting to figure out how to bring this new capabilities into their work as well. Alongside the model jump, folks also recognized that part of the new capability set was actually about the harness that you situated the models in. Now, cloud code had been growing in adoption throughout 2025, but became a real focal point in the new year, which was perhaps augmented by open AI going all in on their codex product as well. Still, I think in many ways, where this whole idea of harnesses, and frankly, a much deepened understanding of what we actually mean when we say agents and what it means to build and manage an agent, came when open-cloth became popular. Hundreds of thousands of people, perhaps millions if you include the people who were standing in line in China to get access to an open-cloth, really got their hands dirty, figuring out the guts of how these agents work. And while you don't necessarily see everyone running their Mac mini setups anymore, the explosive learning of that early period of open-cloth, I think will be seen as a key inflection point moment for the history of agent AI. Now of course, all of this wasn't just happening to individual builders, and the evidence that something fundamental shifted started showing up, particularly on the revenue side of the ledger for the big labs. For the first few months of this year, it seemed like every time we turned around, anthropic at particular had released some new jaw-dropping number about how much their revenue run rate had grown, eventually eclipsing open AI, although it's not like they've been particularly slow in their revenue growth either. Now the interesting thing is that the enterprise's experience is the inverse side of that revenue chart as a cost chart. And on the one hand, this was always inevitable. For years we've been talking about the idea that AI in the enterprise is not just another category of software spend, but represented something fundamentally different. Something more akin perhaps to labor. The explosion of intelligence consumption reflected in that growing revenue and the growing cost for enterprises were simply a manifestation of that fact coming to bear. Now as an aside, the recognition that we were not talking about seats but instead talking about tokens did a whole lot to collapse the AI bubble narrative zone Wall Street from Q4 of last year as well. Pretty soon we were getting stories of enterprises absolutely torching their annual budgets in just a few short months. Uber was the most notable of this, and although these stories were presented as surprising, if you actually think about it, it's really not that surprising at all. How are we going to expect organizations to effectively budget for the agent token era of AI when no one knew that that was right around the corner when those budgets were being made? Subsequently, and regular listeners of this show will know that these are the themes that have dominated for the past several months. We have seen adaptation to this new agent paradigm run in all sorts of different directions. In some corners we're seeing token caps where companies are going with limits per user per month. We're saying companies have to experiment with and try to figure out measurement and monitoring and observability systems. As cost spiral it puts a whole new emphasis, something that was already coming up in the harness conversation, around the fact that we were no longer just talking about AI as a choice of which models, but as an architectures and systems design question. The router of course the product is yours is one response to this, but when it comes to enterprise buyers and planners and strategists, I don't think anyone and certainly my conversations this week at the KPMG event have confirmed this, is looking to open router or any other solution as some silver bullet that's going to solve all these problems. And there are new problems. Specifically, the capability gap is growing on both an individual and an organizational level. The capability gap of course is the space between what AI can do and the value that we're getting out of it. Now the good news is that it's grown largely because the upper bound of what AI can do is rocketing upwards at an incredible rate, and yet still there are real consequences to that gap widening. One of my bully pulpit issues is that I believe that the upskilling bill is coming due in a huge way. When AI learning was just about whether you could prompt well, maybe you could get away with not investing a ton in training your workforce. Now on the other hand, we are talking about a fundamentally new work primitive. The way that people work is changing in a core way in many disciplines and functions, from "I do my work to I manage agents that do my work for me." The need that that creates for training is radically heightened from the previous era of AI, and indeed one of the things that a lot of folks are talking about here at this event is how to deal with apportioning these incredibly powerful tools that are inherently technical tools to folks that aren't engineers and aren't technical by background. There are a lot of stories floating around this event of people accidentally unleashing agents on critical systems, not because even necessarily they were doing anything wrong, but because they weren't the right guardrails or access provisioning, and these incredibly capable models with their new tenacity just didn't stay in their boxes. Now this is not an upskilling question alone. Again the watchword at the moment is systems and architectures, but without that training organizations are almost doomed to face this sort of issue in increasing fashion, or on the other hand restrict the opportunity for people who could really be doing incredibly valuable work with these tools to do so because they're not trusted to do so. And this gets us to the questions that were explored not only in the panel discussion that followed this presentation, but honestly in these side conversations all over the event as well. The first question is how are enterprises redesigning for the agentic era? And the key word here is redesigning. The biggest caution that folks like Steve Chase from KPMG on the panel had was the warning of the problems with and ill effects of trying to simply bolt on an AI strategy to existing processes and systems. Now that has always been problematic and at least under maximizing for the potential of AI, even when we were firmly in the assisted AI and efficiency AI era, but in this time of new agentic capability that gets even worse. Relatedly, the second question is about the nature of that redesign and why organizations need to be thinking in terms of architectures, systems, not just models. If previously an organization's response to some new challenge drop by technology was to figure out which vendor was best suited to solving that problem, that is simply insufficient for the moment that we find ourselves in now. Thinking about architectures means thinking about complex model systems that allow different levels of intelligence for different types of tasks. It means thinking about yes, the routing systems, whether they are products off the shelf or be spoke or something else, that allow that routing to happen. It's also about that harness design, about which functions and people have access to what types of context and data and systems integration and what the guard rails that surround it need to be. And as we get into the third question, how are you provisioning costs across different groups? The big thing that underlies that is another systems design need, which is systems for monitoring and measuring AI usage. You have not seen the word token used more at an event since the height of the crypto era man. And obviously the tokens we're talking about at this event are very different, but there is a very broad recognition here that without better visibility into the cost of AI and its relationship with outputs, it gets very hard to figure out which individuals, which groups, which functions, which projects should be getting access to which types of models and in what magnitude. Given that bully pulpit I mentioned before, I have certainly been gratified to see how big a concern enablement in education really is among these organizations. If I had to characterize the average discourse I've seen around that, there is a lot of throwing up of the hands and saying, screw it, we're just going to have to do this ourselves and experimentation with bespoke customized solutions for this that work for the organization and the population that it has. In other words, there's a recognition that this is not going to be a bunch of video courses of the pattern of corporate trainings, your, but instead is going to involve real messy work of getting people to use these tools in new ways to do new things and then figure out how to transmit knowledge between parts of the organization that are figuring it out. Well, versus parts that are not figuring it out so well. Indeed, one of the big patterns that I am seeing over and over and over again is various forms of collaboration between both AI redesigned software engineering organizations and business units, but also AI early adopters and AI champions and other types of business units. I think the sophistication in the conversation is that no one is talking about the marketing folks replacing the engineers, but they are now talking about the 10 or 20% of the types of skills and even more than that mindsets that engineers or product managers have that can become a part of the essential toolkit for those people and other functions, be it marketing or sales or back office or what have you and how to best do that new sort of transmission. Now I would say that a lot of the discourse at this event has been focused on internal transformation. 2026 is very clearly the year that for this representative sample of enterprises, AI is not a technology problem, but a transformation problem has really come home to roost as the reality. And yet, there is also the entire dimension of agent transformation that has to do with what happens externally as well. In other words, how are agent opportunities reshaping business cases? Some of the examples of that that people are discussing here include shifts to the business model, people experimenting with outcomes based pricing instead of input based pricing like hourly billing. There is some discussion of new types of products and new types of services that become available in this new context. And there's also a lot of re-evaluation of what the core state of the old product actually means. What is, for example, an audit if agents can be doing a lot of that work and if they can be doing it not just on a one off basis, but on a persistent basis. It feels to me as though that while that type of conversation is happening, most organizations are viewing themselves as patient zero. Let's call it for whatever their external AI strategy is and are focusing on showing up how they work first before necessarily making radical changes to what they sell externally, although certainly for certain types of organizations that change is being forced upon them. Now of course, when it comes to business model disruption, it's made all the more difficult by the fact that no one gets to just shut things down for six months to figure this all out. They got to do it in real time even as they're servicing legacy customers on legacy products with legacy methods of delivery. And on top of all of this, the last question that we explored and that was floating around here is if and as we are successful in designing new systems, how can we build dynamism into that that has almost planned obsolescence and an appreciation of a femurality built into it. The harnesses around them are going to change. And patterns are going to change. Customer expectations are going to change. Market expectations are going to change. Policy is going to change. And so whatever new that gets built has to assume and designed for the fact that it's few months down the line from whenever it is ready will likely require it to change all over again. If all of this sounds head spinning, it is, but I think that there is something immensely positive. Last year, even at this event, which is about as AI-pilled as an enterprise event can be, there were still as I said, so many if questions, how do I convince others in my organization that this is real and that we should be doing it? How do I show ROI to prove that what we're doing is worth the time and money that we're spending on it? Now it's not that ROI questions and things of the like are gone, but by and large, the questions that people are asking now, are it feels like to me the foundational questions for redesigning for a new era that we are going to be answering for the next call at half decade? Companies asking about designing and allocating token budgets. Are now exploring this new category of spend that is just going to be common essential part of their organization when companies are talking about building observability systems around the new intelligence they're using. While the models and harnesses may change, it is very likely that whatever gets updated is still going to need that sort of observability. I guess the point is that the paradigm shift has happened. For years, basically since the chat GPT moment, enterprises have been anticipating the shift from assisted AI to agent AI. The opportunity for AI not just to help us do work, but to actually do the work itself. Now that that is here, all of the questions are about how we solve all the new problems that that new way of working brings and how we best seize the opportunities that it opens up. Almost none of the questions have answers right now, but it should feel good, I think, that the questions being asked are the right ones. Anyways, thanks to KPMG for having me out. It was a great event and I look forward to coming back next year, where honestly I can't even imagine how different it's going to be by then. 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!