We’re already starting to see what kind of new jobs AI is creating. AI requires significant technical work and surrounding services to deploy into the economy. This means jobs for AI engineers that build applied AI products sold to (or within) enterprises, FDEs to deploy agents into companies, new services firms for deploying AI, and more. And even the published stats of new jobs undercount all the existing jobs that are transitioning to new areas of AI work in an enterprise. Many prior data, research, and software jobs large enterprises are also being repositioned for working with AI. Every bank, life sciences company, manufacturer, and even law firm is bringing on more technical talent -or repositioning existing roles- to help with agent deployment in their companies. It’s a lot easier to picture what AI can replace vs. what it creates until it starts happening. Now we’re seeing what this looks like.

This metric may tell us more about the business model of consumer AI than the state of AI. AI for consumers almost inevitably will be subsidized and monetized via commerce, ads, or purchases of devices and other services. This metric may level off lower than we think.

AI agent adoption is still very bimodal right now. You have coding and coding adjacent tasks which have taken off, and then everything else. And even within coding, you have a very wide continuum of adoption patterns, with some likely a small percentage of developers deploying background agents working on projects in parallel, and the rest of everyone else still working with agents 1:1. Then, there’s the entire rest of knowledge work where real agentic adoption is still very early. The reason for this is that most of the workflows still need to reengineered to work with agents. This isn’t the same as deploying a chat system for better research efficiency, but instead requires workflows to be rebuilt, data to be wired up in new ways, new practices for accountability and liability of decisions when agents are in the mix, governance and compliance paradigm changes, security upgrades, and more. We’re still so unbelievably early in what this is going to look like outside of a few categories of work right now. This is why you can basically expect 100X more agent adoption from what we’ve seen so far.

A big trend with most enterprises that I talk to is deploying internal FDEs into departments in their organization to help bridge the capabilities of AI into their underlying workflows. This requires a high level of technical expertise, an understanding of AI, and the ability to understand the workflows and processes that the enterprise is trying to automate. There’s no shortcut to getting automation without this combination of skills (either in one person or multiple). What’s exciting is this is an entirely new function in most enterprises, which is going to create a ton of new roles in the economy. “Every generation of technology creates jobs that didn't exist before it, jobs that arise from an immediate need in an emergent industry. The Automation Engineer is a key example of those jobs for this generation. Nobody has ten years of experience doing this yet, because ten years ago the tools that made it possible didn't exist (it didn’t even exist two years ago).” If you have software skills and are diving into AI, this is an area to go deep on.

A big trend with most enterprises that I talk with is deploying internal FDEs into departments in their organization to help bridge the capabilities of AI into the underlying workflows in the organization. This requires a high level of technical expertise, an understanding of AI, and the ability to understand the workflows and processes that the enterprise is trying to automate. There’s no shortcut to getting automation without this combination of skills (either in one person or multiple). What’s exciting is this is an entirely new function in most enterprises, which is going to create a ton of new roles in the economy. “Every generation of technology creates jobs that didn't exist before it, jobs that arise from an immediate need in an emergent industry. The Automation Engineer is a key example of those jobs for this generation. Nobody has ten years of experience doing this yet, because ten years ago the tools that made it possible didn't exist (it didn’t even exist two years ago).” If you have software skills and are diving into AI, this is an area to go deep on.

Monster model right here

There's a huge opportunity right now in being the deployment layer for AI into the economy. The amount of work it takes to change out workflows in enterprises tends to be far greater than anyone realizes or would prefer. Clearly this is what the applied layer of AI is going to look like in the form of software and agents, but also it opens up new services firms opportunities. Legacy systems need to be moved to the cloud, data organization and access needs to be updated, software needs to be connected to agents in new ways, workflows need to be reengineered for agents, HITL needs to be figured out for the process, evals need to be generated and maintained, and the entire system needs to be continually updated as new models get released and new capabilities emerge. And the full list may even be longer. AI is not the same as just deploying software. Software you generally did the implementation of an existing, well understood category of technology, then stepped back and the customer kept running. With AI agents, you're delivering actual work augmentation to the organization, which has a completely different set of complexities associated with it. You're no longer deploying tools that the company is merely enabled by, you're deploying work output in a process. Completely different implementation and enablement process. As a result, this is going to open up lots of new kinds of firms and plays for existing firms to diffuse AI into organizations. We're going to see approaches by industry, by size of company, and by problem inside of companies. Traditional SIs will modernize and adapt (some will clearly not adapt as well), and new entrants will also be founded in this period that take advantage of this window. Great time to be an FDE or FDE firm.

It’s entirely plausible that the AI industry can deal with safety and security through a set of shared standards and practices for the foreseeable future. At some point in capability progress there will inevitably be greater oversight, testing, and layers of liability and regulation, but we are still in a phase where the industry can absolutely align on these efforts collectively. Awesome to see this come together so quickly and for the whole industry to support progress without reducing competition. This certainly points to a much brighter future for AI.

A good direction for the future of AI

At Box, we've been testing Sonnet 5.5 in early access on our complex work eval with the Box Agent, and Sonnet 5.5 is another strong jump from Sonnet 5 on knowledge work with enterprise content. Across the board we saw a 4 pt overall improvement on our hardest tests, but it was roughly 2.4× faster to a finished deliverable on 12% fewer tokens. We also saw major wins in key verticals, like a +18 percentage point improvement in Financial Services, +7 pt in Legal, +8 pt in Life Sciences, and +7 pt in the Public Sector. Here are some specific wins on certain tasks that show the power of Sonnet 5.5: * Financial services (+48 points): a due-diligence review of a trading-tool acquisition, where the deal book's arithmetic doesn't hold up. Sonnet 5.5 caught the miscalculated interest totals and the mispriced options and flagged them, and it also finished this work about 61% faster. * Legal (+14 points): a commercial lease review. Sonnet 5.5 refused to invent a "standard" market benchmark for renewal terms and notice periods, and instead flagged what the lease was genuinely missing. * Life sciences (+22 points): writing up a field trial from the raw measurement data. Sonnet 5.5 got the sample standard deviations right across both treatment groups. Sonnet 5 missed nearly all of them. And Sonnet 5.5 did it about 41% faster. * Public sector (+11 points): a report on a math intervention program. Sonnet 5.5 pulled the right student progress-monitoring figures out of the underlying data, about 71% faster. Customers will be able to build AI agents with Sonnet 5.5 in the Box AI Studio shortly.

Interesting to think about all the implications in a world where agents begin to make the best or most efficient choice for their users. On one hand, there is some subset of the economy that benefits from the friction customers traditionally have changing something about their habits. In those parts of the market, switching costs will come down and competition is going to increase dramatically until some equilibrium is reached between the disruptive alternatives and the incumbents. On the other hand, there are lots of markets that are hurt by a significant amount of friction, where agents will begin to unlock all kinds of economic activity by making it far easier to buy products or services that were too frictionful before. Healthcare, travel, local services, certain information services, and other entirely new markets probably are net beneficiaries of agents as a result. No matter what, a future meaningfully mediated by agents that work tirelessly for us and our goals can’t possibly function exactly the same as today. Going to be wild.

You can’t automate what you can’t measure. This means that evals are one of the gates to diffusion of AI in the enterprise. We can test our deterministic processes through software, but most enterprises have no useful way of understanding how their non-deterministic processes are working today. Specifically the work that agents are doing for them. Evals are mission critical for enterprises adopting AI because you have no other way of knowing what’s working, what’s broken, what changed, what improved, what you can do more of, etc. if you don’t have a good sense of how agents work in your environment today. All changes, upgrades, and deployments are downstream from good evals. Not only are we going to get vastly more domain specific evals over time for the labs and across the industry, but every enterprise will also need a clear sense of how agents are performing in their environment as well. Huge opportunity.

@alexandr_wang @Box @Muse Imagine being a CIO and opening X right now

Great vision for the future of the creative industry with AI from the one person who can actually predict these things. “As the barriers and the costs come down, more films will get made, not fewer. Studios will get to take more risks. There will be more seats at the table, and very soon entirely new forms of storytelling. In the 1980s, animation was dismissed as a niche corner of the business. Today it is one of the most beloved and profitable forms of storytelling in the world. In live action, filmmakers like Steven Spielberg, James Cameron and Peter Jackson embraced new visual tools not as shortcuts, but as instruments, and expanded cinema in the process. Every time storytelling has met a genuine technological shift, from synchronized sound to color to computer animation, it has redefined the boundaries of the medium and grown larger in the process.” Technology has reshaped the creative industry over and over through the years, and has generally always led to an expansion of opportunity or new approaches as a result. And even when the techniques or mediums change, the underlying need for creative skills and taste don’t go away. Just more people are able to leverage those skills effectively when a new technology arrived, and users of the tools discover novel new ways to tell stories.

Muse and Box, together at last https://twitter.com/levie/status/2102915863989075979/photo/1

What an insane day in AI. The frontier models just became substantially cheaper, with the Opus 5.5 price cuts, and now with GPT-6 Sol and Luna dropping token prices by 50%. The rate at which the cost per task (on a like-for-like basis) drops in AI is unlike any other type of technology in history. And every time the cost of AI drops, the use-cases you can deploy agents against dramatically increase. This is Jevons paradox applied to agents. These improvements will directly lead to broader diffusion of AI in the economy as we can use agents to process all of our data, scan our code for security issues, read through all log data to make decisions, have agent swarms in workflows, and much more. The cost of tokens is directly correlated to these use-cases being opened up at scale.

At Box, we've been testing Opus 5.5 on a variety of complex enterprise knowledge work tasks dealing with unstructured data with the Box Agent. Overall, we saw frontier capability levels, with major performance improvements over Opus 5. 63% fewer tokens used, 42% less verbosity, and 30% faster vs. Opus 5. And the model itself is cheaper, so this is a major win for any agentic computer use, coding, analytics, or data work that enterprises will be doing. Here are some examples of the task wins and performance gains across a variety of industry tests that we performed: • Financial services - due diligence (+39% task accuracy): A year of transaction records, with the job of finding every miscalculation in an acquisition target's pricing tool. Opus 5.5 scored a perfect result on every attempt in half the words Opus 5 used, consuming 82% fewer tokens overall. • Technology - cloud cost analysis (+65% task accuracy): Work out what a company should actually change about its cloud spend. Opus 5.5 picked the right basis for the retention calculation and kept the source data's unit conventions straight all the way through, so the number at the end actually holds up. It took half the time Opus 5 took, with 70% fewer tokens. • Consumer products - client account analysis (+17% task accuracy): Set the onboarding targets for a client account, reading across the signed contract, a satisfaction tracker and a team metrics sheet. The contract never states a senior/junior split, so Opus 5.5 derived it from the 18-person roster and showed the rule it used; several clients had a perfect 10 on individual survey questions, so it averaged each client's responses instead of crowning the single 10. It finished this one in half the time, on 78% fewer tokens. • Clinical diagnostics - data analysis (+15% task accuracy): Malaria rapid-test performance across a dry and a wet season: build the patient records out of two clinical PDFs, compute positive test rates by season and gender, and test whether parasite counts really differ between test-positive and test-negative patients. Opus 5.5 caught that the two groups' standard deviations differed more than 100-fold, re-ran it the right way, and found the dry-season difference didn't hold up after all. This accuracy gain came with a final answer that was half the length of Opus 5's, and also needed 78% fewer tokens end to end. Customers will be able to build AI Agents with Opus 5.5 shortly in the Box AI Studio.

The monetization potential of personal agents that are transacting on your behalf you is quite significant. If you imagine agents that are perfectly capable of handling an arbitrarily complex task end to end, then eventually a substantial amount of commerce inevitably goes through them. You’ll start by slinging your daily simple and annoying tasks at the agent. Then as people get used to it, they’ll just start to throw more complex tasks at the agent, ultimately leading to even more spend through these systems than what they were doing before. If you can bring down the friction for commerce and services, then you end up spending even more. This thus creates a ton of opportunity for the agent providers (Muse, etc.), but also completely new opportunities to build the layer that the agents want to interact with (commerce, local, b2b services, etc.). Win/win for multiple layers.

AI agents will use software 100X more than people ever did. Even as interfaces begin to fade into the background as you primarily interact with agents, those agents still need many of the core primitives that people have used. In fact, in many ways these core primitives become even more important when agents can take destructive actions in our systems or where the context they’re accessing with make or break the workflow. This will be true of where you house your CRM or ERP or structured or unstructured data platforms. The platforms that can best act as the security layer and guardrails for agents, manage the data for agents and people, and orchestrate the business logic for workflows have a huge opportunity right now. This is true for brand new startups as well as existing platforms that can move fast enough.

Literally impenetrable from agent swarms

Personal agents are the ultimate manifestation of “build something that agents want”. The form factor of a product like Muse is you want to be able to hand off a task to the agent and ensure that it is fully completed end to end. To do this, the agent must be able to successfully operate with your tools or use its own to complete the task. Use your MCP or CLI, easily navigate your site, be able to transact, and more. The new attention you need to compete for is not from the user itself but instead for the agent. This means that the tools that allow agents to order food, handle ecommerce transactions, book flights, work with the local economy, and interact with our data and information best, are the ones that will get used the most. This will ultimately be the biggest opportunity and shakeup in consumer tech since the App Store itself.

@alexandr_wang A literal human thought other literal humans were writing these posts? Incredible.

Jev will be super helpful for agents to make split second decisions in workflows, data classification, judgment calls, and hundreds of other use-cases in the enterprise. Here's a quick demo with Box and Jev to make that real. The demo pulls an incident report from Box, asks whether it's customer-facing and how severe it is, moves the file into escalate, monitor, or review folders, and sets a metadata template instance with the result. This all happens nearly instantly and at almost no cost. You can imagine this in insurance claims, contract management, loan processing, security reviews, customer log analysis, and so on. Definitely a great new class of AI use-case.

Agents already make up the majority of inference. This will quickly trend toward nearly all inference over the next year or two. The vast majority of tokens used in the world will be agents that are executing unbelievable amounts of tasks for us in the background 24/7. Agents will be deployed to read all code changes to secure our software, process all of our data inside of workflows, handle a significant majority of the research that goes into recruiting and customer prospecting, review every event stream and log from every system in the world, go out and execute tasks for us in our personal lives, and hundreds of other use cases. The rate of new agents coming online that will be consuming insane amounts of tokens is not slowing down. Just in the past week I’ve introduced multiple completely new workflows that would not have been possible technically even a month ago. Incredible time to be doing anything in inference and of course building on top of all of this.

Incredibly exciting that there are entire universes of AI innovation that still exist that weren’t even on most of our radars. Being able to process information insanely quickly, at crazy low costs, with high levels of capability is huge for a wide number of enterprise tasks. Data classification tasks, routing decisions inside of a workflow, decision making when handed a particular domain problem, quick judgment calls about safety or security, and more all are the gates in a large number of processes. This model and approach could be quite cool in agentic workflows in the enterprise.

There’s a massive chasm between the power of AI models and the ultimate workflows that enterprises are trying to automate. This gap is the opportunity for the applied AI layer to fill. You need to connect the intelligence to workflows, often reengineer processes, aggregate the right context and data, allow for the right human in the loop experiences, drive change management, do domain specific evals, manage the security and governance of the data and process, and much more. We’re going to see this layer emerge in every vertical and horizontal category. And ironically, even as models improve at incredible rates, this layer still must exist - and may become even more important and useful. Greater capability enables even more complex tasks to be tackled, amplifying the challenges if you don’t do this well. Was super fun chatting with @sonyatweetybird on all the things going into AI diffusion.

We probably need to all update our sense of what’s coming in terms of agentic workloads with the combination of agent swarms, better computer use, the next wave of APIs and MCPs coming online, and new form factors like Muse or Instinct, vertical enterprise agents, background workflow agents, and other products that are emerging right now. We’re going to throw agents at vastly more tasks in our professional and personal lives than we had initially imagined. The kind of information that agents will go out and find for us and do work for us in the background is going to be 100X more volume than what we can imagine previously prompting in a single session. You’re going to have agents go out and surgically recruit talent for you 24/7, look for every signal in your customer’s business for when to pitch them, process every single transcript and conversation for product insights, review every line of code written for security issues and bugs, brute force test your systems for issues, and 100s of other tasks. We’re probably 1% of the way into the shape of what all these agents look like, where they get deployed from, how they get managed, how we budget for them, and so on. But it’s inevitably going to happen at a scale that wouldn’t have been fathomable before.

Protecting enterprise data in a world where agents are using our systems 100X more than people ever did is going to be one of the more complex security and governance challenges of the 21st century. Importantly, security and productivity gains are inexorably linked in the world of AI. If you give an agent too much unfettered information access, it will be difficult to truly control and protect your data; and conversely, if you lock everything down completely, you won’t get any real productivity gains from AI. We need all new ways to protect our systems, environments, and structured and unstructured dada in the enterprise in an intelligent way by modernizing our approach to security and governance. At Box, as one example, we’re building new intelligent ways to protect enterprise data and agentic use of that information. A recent update in Box Shield is to provide granular controls on what content agents can and can’t work with based on document classification level. We’re also working on other features that can automatically detect and alert (or block) when data is being accessed or used in unusual or anomalous ways by agents. And this is just the start. There’s a ton more innovation coming across the entire industry - from the labs like OpenAI or Anthropic; security platforms like Palo Alto Networks, Cisco, CrowdStrike, Okta; or startups like Enclave, Method, Alterion, Runlayer, and many many others - to rethink how we protect information in the world of AI agents. Exciting and wild times ahead.

“Pacing” can be somewhat of a trigger word because it sounds like an arbitrary slow down of capability or a way to hobble competitors through undue regulation. However, the specific improvement goals laid out in Dario’s are absolute necessities in AI development. We expect equivalent in areas like aerospace, life sciences, health care, and other industries, and AI development likely shouldn’t be that different. AI is going to be the technology underpinning our financial trading systems, our medical devices, our biotech breakthroughs, defense systems, government workflows, and many more mission critical domains. So it completely stands to reason that we want these systems to be safe and “aligned”. How we get there -without meaningfully slowing down innovation progress or reducing competition- is one of the most complex questions in the 21st century, but the need is clearly real.

“Pacing” can be somewhat of a trigger word because it sounds like an arbitrary slow down of capability or a way to hobble competitors through undue regulation. However, the specific improvement goals laid out in Dario’s are absolute necessities in AI development. We expect equivalent in areas like aerospace, life sciences, health care, and other industries, and AI development likely shouldn’t be that different. AI is going to be the technology underpinning our financial trading systems, our medical devices, our biotech breakthroughs, defense systems, government workflows, and many more mission critical domains. So it completely stands to reason that we want these systems to be safe and “aligned”. How we get there -without meaningfully slowing down innovation progress or reducing competition- is one of the most complex questions in the 21st century, but the need is clearly real.

Good post. Don’t agree with all of it, but it represents many of the actual practical realities for the general path forward in frontier AI whether we like it or not. At the level of capability we’re seeing from models there will inevitably be some form of coordinated self-regulation of the industry; that’s generally a good thing. The big question will be if everyone can agree to what that is. However, the way things are headed politically there’s a chance that the labs won’t even get a say in it. And then the bigger question yet is do all the countries participate or not. Any “slow down” hinges on broad participation, which from a game theory standpoint isn’t likely until the risks are more severe and obvious. Expect this to be pretty messy for a while.

Good post. Don’t agree with all of it, but it represents many of the actual practical realities for the general path forward in frontier AI whether we like it or not. At the level of capability we’re seeing from models there will inevitably be some form of coordinated self-regulation of the industry; that’s generally a good thing. The big question will be if everyone can agree to what that is. However, the way things are headed politically there’s a chance that the labs won’t even get a say in it. And then the bigger question yet is do all the countries participate or not. Any “slow down” hinges on broad participation, which from a game theory standpoint isn’t likely until the risks are more severe and obvious. Expect this to be pretty messy for a while.

Great time to pivot to software

This one is very cool. Now you can mount Box to agent sandboxes to make it far easier for an agent to read and write files on the agent's computer. As AI agents execute critical workflows in the enterprise, they're going to need the same primitives that people have had.

Some more tales from the road. Met with a couple dozen technology leaders this week across banking, media, information services, insurance, and consulting to discuss agents in the enterprise. Some of the biggest trends right now: * Cyber! Everyone nervous about the growing rate of vulnerabilities coming at them from AI, and the implications of the OpenAI Hugging Face incident. The conversation is not as existential as it is in Silicon Valley, but still highly concerned and pragmatic about what to do about it operationally in their environments. Lots of new discoveries due to AI, and still hard to keep up with all the changes they have to execute now. * Model battles persist. Most companies are deploying multiple frontier models within their enterprise. Too hard to standardize on anything and seeing different preferences across their teams and use cases. But the dollars are still concentrated on just a few vendors. Open weights still in infancy at scale in most of these organizations, often due to lack of domestic “frontier” OSS options. Plenty of appetite for more options here, but so far few places to go. * Agent security and identity. Somewhat tied to Hugging Face, there’s much more awareness to the new challenges around agent security and identity management in a world when agents are trying to get into every system they can. In a perfect world enterprises could setup identities for all their agents and control what they’re doing, but of course sometimes the agent needs to act exactly as the user as well. * Process reengineering. Most companies realizing that the big upside of agents is when they can change the actual workflow itself to get the full gains from AI. Far more ROI when companies can adjust their workflows to support agents changing how the work happens instead of just layering on agents into the existing flow. But the big question is who can actually tackle driving these changes, where does that live, etc. Best lessons were still around embedded FDEs in the functions. * Ruthless adjusting of architectures. Most companies had examples of changing systems out multiple times just in the past year or two with different vendors. I probably haven’t heard “we tried X and it didn’t work so have gone with Y” more than in today’s environment. The lesson here is that because innovation is happening so fast, no one hangs around until a vendor gets something right, they just move on to the next one. * Evals! Still very early for most companies to have a good grasp of evals of their workflows. A few customers out of a couple dozen called this out - huge opportunity right now for enterprises to have a good sense of how their work actually happens and how well AI is doing against it. * Legacy systems still a hurdle. As always, legacy systems still remain a mainstay issue that holds back enterprises from rapid adoption of AI in enterprises. Data is fragmented across legacy environments that weren’t built for an agentic world. Companies spending a lot of time just cleaning up these old platforms. Many more topics, but these tend to be some of the more top of mind items at the moment in the enterprise.

Excited to partner more deeply with OpenAI so you can work with your enterprise content from Box securely in ChatGPT. Software continues to go headless, and we’re massive believers that the future of work will be AI agents that can process data and execute workflows anywhere.

The world is going to be using coding agents for far more than anyone would have thought before. We’ll be building tools in all new categories, building software for companies that couldn’t afford it before, using agents to protect systems from cyber risk, automating research in life sciences, processing insane amounts of data in complex workflows, upgrading legacy systems and infrastructures, and hundreds of other use cases. In short, by lowering the cost of code, we are going to do so much more with it. This will result in software becoming more useful and the leverage of engineers going up dramatically, which coincidentally will mean we need more of them, not fewer.

The way to the reconcile capability level of AI vs. GDP impact is that the diffusion of AI will take much longer than people think. And it will also show up in ways that are hard to measure in GDP immediately. You could bring the world’s greatest superintelligence to many workflows, and still be bound by the laws of corporate physics: getting data prepared and put into a pipeline, process reengineering and change management, aligning on how the new workflow should function, and so on. Even after you solve all that, you’re still bound by the speed of the real world: waiting for a customer to respond to a proposal, getting a permit for a project, a drug discovery pipeline taking years to eventually reach the consumer, and so on. Not to mention lots of positive daily AI use-cases are entirely net neutral to GDP, at least in the near term. AI diffusion is going to be the theme of the next decade. The upside is that there’s a tremendous amount of opportunity in building the bridges between superintelligence and real-world workflows.

@alexandr_wang 😮

Personal assistant agents are going to be a very exciting AI category. It’s the first time you can have high token volume agentic use-cases that make sense for consumers. Lots of different approaches emerging right now, and it’s going to be hyper competitive because these agents will mediate a lot of consumer spend over time. But this certainly plays directly to Meta’s strengths: lots of compute required, can monetize with ads and commerce, software focused experiences so can distribute it at scale, and so on.

Good reminder here that with rapidly accelerating AI progress, you should likely be building with a vision that contemplates at least a few orders of magnitude of capability improvement or token volume available to you. The best opportunities right now are those where you can value to customers today with things that are barely only possible, but with a full mission that feels nearly impossible with today’s tech. There are lots of products the require orders of magnitude more processing of data (and thus are barely economically viable ideas) or where model capabilities still are in their infancy relative to what’s needed to work. Best to be building for those scenarios.

The AI jobs prediction so far is playing out the opposite of what many thought. The reason this is happening is because AI is bringing automation to areas that don’t have finite demand characteristics. AI allows companies to do far more than they could before, but agents still require human operators and oversight to generate value from them. And not only has AI not replaced many of the types of jobs that theoretically would have been most affected by AI - engineers or lawyers - but AI is also creating or growing entire categories of jobs across the economy. Cybersecurity, FDEs, agent operators, engineers in non-software domains, and tons of job categories are going to explode. It’s not clear if or when this trend will reverse because we keep discovering all new areas of demand that AI is going to open up.

The internet is almost entirely unprepared for a future where everyone’s personal agents are running around executing tasks for them. There are going to be endless new challenges and opportunities at the infrastructure layer, for user experiences, new business models, and more.

If agents produce the vast majority of software in the future, and they’re most trained on open source software, they will inevitably do their best work with those tools. If you cycle this enough times, it means that open source effectively becomes the dominant software in the future as it’s what everything important gets built with. This was already the trend in many critical domains, but agents will accelerate this far faster than humans ever could have.

Good post on how to think about process redesign in an enterprise with AI Agents. Unfortunately for most workflows, there’s no “easy” button. The best way to get real efficiency gains is you have to actually reengineer the process to take advantage of what agents are actually good at and can do differently and faster than people. One of the challenges of course is that the most important workflows actually space many functions at once, so someone needs to be able to go in and retool how work happens. “Nobody in that chain is empowered to walk into finance, for example, and say that their 14 step process should actually only be 5 steps. So nobody does it. Instead, you 'apply AI' on what you currently have, and you end up with faster sh*t.” This is the road ahead for AI diffusion in enterprises. This also presents the clear opportunity for companies - both at the applied AI layer (regardless of the article’s title) and teams or companies that can go and implement transformation into the enterprises. And all roads lead to FDEs one way or another.

@OpenAIDevs Great Model!

Palo Alto Networks for message boards is going to be a $100B company

GPT-6 Astra is out. We've been testing the model in early preview on our enterprise complex work eval at Box. It is now the best model we've ever tested on our expanded and hardest test set. Overall, GPT-6 Astra offers a breakthrough level of capability in coding, analytics, logic, and domain specific knowledge for dealing with complex enterprise knowledge work. In our eval, we use the GPT-6 Astra model in the Box Agent and give it a set of difficult tasks that represent real-world work in various industries. These tasks often will take people hours to execute properly and require deep domain expertise. GPT-6 Astra scored 77% overall vs. 74% with GPT-5.6 Sol, representing frontier performance. But the story is in some of the bigger gains on individual task types that are incredibly complex across. Here are a few examples across the tests that show how much of an impact this will be in different industries: * Media and entertainment (48% → 100%, +52 points). Ranking film genres and countries by profitability across a year of production data from four teams, where the trick is to apply the following year's tax-incentive corrections without double-counting them. GPT-6 Astra was perfect on every attempt. Sol had the rankings right but the underlying ratios were wrong. * Technology (69% → 97%, +28 points). Choosing which region to fund first, from a stack of performance and infrastructure documents that never state the metric the brief asks for. GPT-6 Astra flagged the gap, labelled its own figure a proxy, and caught a growth claim that didn't match the numbers beneath it. GPT-5.6 Sol reported the proxy as the real thing. * Legal (69% → 93%, +24 points). Reviewing an NDA against a company's own contracting policy, where reaching the right verdict isn't enough; the answer has to cite the provision behind it. Both models declined to approve the draft, but only GPT-6 Astra separated whether the liability cap's structure was permissible from whether its amount was defensible, and pointed to the policy language that settles it. GPT-5.6 Sol argued the same conclusion without citing the provision, which is a critical error in the legal industry. * Healthcare (53% → 77%, +23 points). Auditing a batch of radiology reports for errors and rating how serious each one is. GPT-6 Astra caught a terminology error in a knee-imaging report that GPT-5.6 Sol missed on both attempts, and was more reliable at grading severity rather than just flagging that something was wrong. * Energy (82% → 97%, +15 points). Building a consumption report on two facilities from a year of meter logs, where only one of the two sites is actually missing data. GPT-6 Astra found it and drew a line most reviewers wouldn't: the other site's odd solar readings are a measurement problem to investigate, not a gap in the record. GPT-5.6 Sol reported both sites as incomplete. Astra clearly is going to offer a meaningful jump in powering and orchestrating enterprise workflows. We'll be making GPT-6 Astra as an option in the Box AI Studio for customers to build agents with shortly as it continues to roll out.

Another huge moment for open weights AI. The infra platforms are working. The models are getting better. The ecosystems are being deeply invested in. The business models are working.

This week is absolutely nuts for AI releases, and now a major new model update from Muse. We appear to be reaching escape velocity in capability progress. If this thing gets released as open weights it completely changes the dynamic of US open weights competitiveness.

AI for cyber is about to go vertical. The models increasingly becoming insanely good at finding and exploiting vulnerabilities. Frontier models are ahead, but we’re already seeing that open weights is not far behind. Most enterprises are already inundated with cyber discoveries, so now that’s only going to multiply. Triaging and automating the fixes with more AI -along with human oversight- is essentially the only way forward. In case you were wondering what jobs AI was going to create, it’s definitely going to great time to be in security.

At Box, we’ve been testing Fable 5.1 in early release against our complex enterprise work eval. Fable 5.1 delivers a huge 7 percentage point jump over Fable 5 for unstructured data tasks in the enterprise. On this updated test, we use the Box Agent with Fable 5.1 to work through a wide range of real world enterprise scenarios with documents in financial services, life sciences, the public sector, and more industries. All of these tasks require a high degree of analytical, math, logic, and domain knowledge to perform successfully. Here are a few examples of wins: * Financial Services (+17% improvement): on a tax-adjusted profit projection, Fable 5.1 correctly applies capital allowances before computing tax liability; this is a subtle ordering that Fable 5 misses, producing wrong figures all the way through retained earnings. * Technology (+37% improvement): on a cost-optimization analysis where a key metric's normalization is ambiguous, Fable 5.1 recognized the ambiguity, computed both forms, and presented the correct one, where Fable 5 committed to the wrong normalization. * Public Sector (+16% improvement): on an educational data analysis task, Fable 5.1 works through the full weighted-mean methodology and produces correct rankings, while Fable 5 miscategorizes one item and cascades errors through the whole sheet. These are just a few of the wins we saw. Overall major jump in capability for long running agentic workflows in the enterprise. Fable 5.1 will be available shortly in the Box AI Studio for building custom AI agents with enterprise content.

Now that the base open weights AI models are getting far better, and post training infra is becoming more mature and commercialized, there are going to be all new plays for companies that have large amounts of data to have their own models. Licensing data for external model training was previously the only play if you had a large corpus of information, but you can now reasonably go and train your own models as well without incurring the cost and complexity of competing with the labs on research. The general purpose frontier models will still have a leg up in broad areas due to the ability to handle the widest set of tasks, but you can absolutely see a future where we have far more models than today across every vertical and domain.

As AI security events pick up, it’s going to be critical that we have the most sophisticated AI agents for being able to detect and prevent security issues. Frontier models are clearly still ahead in cyber, but open models are catching up quickly. Wild times.

Sometimes people don’t have an intuitive sense of what Jevons paradox looks like for token consumption. Those of us working with enterprises get to see this first hand every day. Basically enterprises have an unending stream of tasks that they’d love to be able to bring automation to. But for each individual task, it’s either ROI positive or not based on the cost of bringing automation to it (including the setup time, change management, and so on). As tokens get cheaper (at a certain capability threshold), enterprises can afford bringing them to more of the work that they do. This could be processing every contract, reading every log, watching new streams of data for insights, having background agents execute workflows, and so on. Any time we can lower the cost of tokens, we will see a disproportionate increase in consumption. Even a 50% drop in token prices could result in a 5X increase in tokens for these kinds of workloads. This is why it’s critical to keep bringing down the cost of AI, and why that’s good for all market participants.

The average strongly held belief these days in AI has a half life of 6 months at best. Here are just a few the industry has cycled through and probably has no consensus on at the moment: * OSS is too far behind to catch up * The labs can’t be profitable at scale * All software will be replaced by agents * You can’t build moats on top of models * Cheaper models will mean less compute * You don’t need evals * RAG is dead * AI will decimate engineering jobs * Prompting won’t matter in the future * We’ve hit a training wall * Frontier models are too dangerous to release …and dozens and dozens more. The key here is to remain flexible in your thinking because we’re going to be in a constant state of change for a while.

This week’s tech earnings calls were a critical reminder of how valuable the relationship is between software and AI. Software provides the guardrails for how data is managed, how business logic in workflows are maintained, how you manage and govern access to information, how you protect data, and more. You want these systems to work the same way every time, without fail. Agents, on the other hand, work inside these systems and with this data to execute on tasks just like people would have. But at a scale far greater than people ever did, which is especially why the deterministic controls matter more than ever. And in fact many of the best ways to deploy agents will be directly within the software systems themselves, like from Salesforce, Box, Harvey, ServiceNow, and so on. These platforms can optimize agents for their workflows and domains, with evals and context that are unique to their vertical, as well as connect to other agents headlessly. The net result is both software and AI adoption go up, hand in hand. Software and Agents are going to work together, and ultimately grow the size of the IT TAM dramatically as a result over time.

Yesterday, at Box we reported our Q2 results with revenue of $321.1 million for the quarter, up 9%, or 11% in constant currency (our highest constant-currency growth rate in 14 quarters), and we raised our full-year revenue target to $1.290 billion. Overall, this growth is being driven by the demand we're seeing from enterprises aiming to get the most out of their enterprise content and transform in the era of AI. As I shared on our earnings call, during the quarter I spoke with many enterprise technology leaders who highlighted their primary goals and challenges in implementing AI. One of the most common topics is how enterprises can get the right context to AI agents, as well as tap into the full value of their unstructured data, in a secure and governed way. The world’s most advanced superintelligence is only as useful as the underlying enterprise knowledge it has access to. Agents need access to the critical information that makes up the bulk of corporate knowledge --like contracts, research materials, financial documents, marketing assets, product roadmap information, and so on-- most of which lives within an enterprise’s unstructured data. We've long been able to query, calculate, and automate work that deals with structured data; but the vast majority of data in an enterprise is unstructured, often in the form of our enterprise content. Instead of sitting on millions or hundreds of millions of files they know very little about, companies can now use AI agents to ask questions about this data, mine it all for intelligence, and automate nearly any workflow that involves this enterprise content. The technology leaders I'm talking to are also recognizing that as AI model capabilities advance across AI labs --like OpenAI, Anthropic, SpaceXAI, Meta, Google, NVIDIA, and a variety of open-weight providers-- they will need applied AI platforms that can bring the full power of these models to their processes. And with AI token usage continuing to grow exponentially, the ability to draw the right cost-performance mix becomes essential. Rather than migrating data and workflows into separate systems to unlock AI's benefits, enterprises need the ability to swap models or agents on their workflows and information at any time. Finally, as we saw with the recent OpenAI Hugging Face incident, enterprises will increasingly need platforms that can securely protect their corporate data and ensure that neither humans nor agents can get access to information they shouldn’t have access to. As external AI agents interact with enterprise data, enterprises will need platforms that offer robust guardrails, comprehensive audit logs, and real-time security alerts to ensure content remains protected and governed at all times. Given all of these dynamics, there's going to be a tremendous amount of opportunity for various plays to bring intelligence to the critical workflows inside of enterprises at the applied AI layer. At Box. we're focused on building the leading intelligent content management platform to empower enterprises to transform how they work with their unstructured data at scale. And we're looking forward to partnering with all of the other companies bringing intelligence into organizations as well.

Good post on what the applied AI strategy looks like at scale. It’s clear that there’s a wide gap between the AI models and the underlying workflows of an enterprise, which leaves a ton of opportunity for applied AI companies. “The world doesn’t just want raw models and agents; it wants problems resolved and outcomes achieved. The premium will sit with the companies that can diffuse this intelligence through every aspect of civilization, converting raw tokens into real world outcomes, transforming industries, and creating economies in the process.” This requires understanding the context, driving the change management, having a harness that can route to various models, connecting to the critical business systems in that vertical, solving the UX challenges of connecting users to agents in the right way in a workflow, understanding the evals in the space, and so on. That’s a ton of value that goes beyond just the model intelligence itself. And there’s a window of opportunity right now to build the defining companies that can bring intelligence to the critical domains in an enterprise.

Systems of record have never been more important than in a world where you have AI agents that will do 100X more work on these platforms than people ever did. Agents will be querying the data in these systems, processing tasks, executing workflows, collaborating with human and agent users, and more. Thus, the governance, reliability, security, access controls, and business logic remain critical more than ever — the OpenAI Hugging Face incident is just a small peek into what the future will look like with agents running around trying to execute on their goals. But this only works for the systems of record that respond with the right product experiences, APIs, and business models to support agents on and off their platform executing in these systems. Huge moment of opportunity and disruption right now in all of software.

It’s not fully appreciated that ZDR is responsible for a substantial amount of the growth of AI, because it dramatically simplified the compliance process most companies deal with in handling data with subprocessors. The alternative path would have been a far slower adoption rate. For the applied AI tools that are primarily responsible for driving AI growth, they generally have agreements in place with their customers to only offer models with ZDR to ensure that customers don’t have to inspect each model they turn on. Anything else is an exception process that takes far longer. And then inside enterprises directly, most companies have internal governance requirements that say they can only use ZDR models because they have no way of explicitly separating PII and other sensitive or confidential data they put in the context window across their enterprise. Now, all of this could change if the entire industry went this direction, but with one off models hold outs it would take years for these policies to change. Without ZDR, AI diffusion grinds to a halt.

I’m good to retire now https://twitter.com/levie/status/2091580656598102047/photo/1

AI diffusion is far more rate limited by having good evals than most realize. The kind of evals that you see for every model release are incredibly helpful, but only tell you the shape of general AI progress and the relative capability level of models. The far bigger space over time are evals on all the major workflows that enterprises do, down to the specifics of an individual company. This will be a huge space over time because you can’t automate what you can’t assess the progress on. Enterprises will not be able to go just on vibes.

The rate of progress in AI right now is unlike any other period in tech history. Models are getting cheaper on a like for like task basis, more generally capable, faster, and they’re going deeper in almost every domain. As intelligence becomes too cheap to meter, then the big opportunity is driving diffusion of AI into the economy. This is a great time to be applied AI company because there’s an incredible amount of innovation and competition acting as a tailwind to your own progress. Huge moment for startups right now that can take advantage of this.

Great post on what post training looks like for applied AI use-cases to bring down costs and improve accuracy on certain tasks. This will increasingly be an approach that companies that can get closer to the underlying workflow in an enterprise will take. The key is that once you understand a domain well enough and have enough volume on a set of similar tasks, it can start to make sense to purpose design models just for that work. “In post-training, we incentivized efficient tool use and reasoning through reward shaping, preferring trajectories that would reduce tokens consumed at inference-time given equivalent performance. This allowed us to co-optimize for both cost and quality, gaining significant performance while keeping cost stable.” Now, this won’t make sense in every domain, as general purpose frontier closed or open models will be good enough out of the box -or necessary- for the work at hand. But once you have deep enough vertical expertise, and either the costs are too high to do at scale *or* you have a unique enough task type not being trained on otherwise, this will make a ton of sense. Very compelling value proposition for being an applied AI company, and awesome to see multiple paths to winning in the market right now.

One of the more interesting dynamics with AI agents is trying to find the right form factor for how they should show up inside a workflow. As agents can do longer and longer tasks and need less intervention, they can show up more in the tools that we already use when working with other people or processes. This wasn’t possible when you had to constantly oversee the work they were doing, but becomes much more plausible when you send off a task and just see the work that is done at the end (or in chunks). Of course this will mean most software has to evolve to support this style of interaction, introducing new components and patterns to make this agent interaction work well. So it’s very cool to see what Slack is doing on this front. We can expect to see much more of this across platforms, and it will define which ones end up making sense in the future and those that don’t.

There tends to be a debate between being an expert or generalist in the era of AI. So far, the experts appear to have the upper hand, and that’s not slowing down. AI makes it 10X easier to get started with any kind of task. Coding, legal work, research, financial analysis, and so on. But having the right judgment for how to direct the agent onto the right work, how to veer it to course correct it, being able to review or test the output, and having the right sense of what “good” looks like all requires a high degree of skill depending on the field. That skill can be developed far faster due to AI for anyone interested, which is amazing for uplifting anyone in their career. But there’s no replacement to needing judgment and many of the core skills that go into most domains of work. And if anything, AI will be a technology that exacerbates differences in skill levels because the experts have far more leverage than ever before. Net net: don’t give up on being an expert at something.

Good details on the Stripe + OpenRouter deal here. For AI to diffuse more broadly, developers and enterprises will want ways of being able to mix and match intelligence from a variety of provers seamlessly and better manage costs. This will be an important part of that.

Stripe + OpenRouter coming together is huge. For AI to diffuse more broadly, developers and enterprises will want ways of being able to mix and match intelligence from a variety of provers seamlessly and better manage costs. This will be an important part of that.

As we’re seeing in case study after case study, it turns out that the amount of value that can be created between the AI model and the ultimate end-user workflow is far larger than many people assumed or realized. Model capability is obviously doing a lot of heavy lifting in agentic products, but there’s still a lot more work to diffuse AI into the enterprise. 1. Getting agents to work well (and alongside people) in mission critical workflows tends to need to be represented differently depending on the business process. Sometimes it’s a chat experience. Other times it’s a background agent running in a deterministic workflow. And dozens of other variants. This is a mix of needing a harness that’s tuned to specific domains of work, but also making it show up in the right product experience. 2. Different workflows connect into entirely different enterprise systems and need access to very different data. Working with that data -whether it’s life sciences, financial, legal, etc.- requires contextual approaches, understanding of the data, having the right user experience for data interaction, and more. 3. The need for domain-specific change management remains critical in most verticals. The way you talk and implement technology at a bank is very different from a law firm. Having the right talent with a singular mission ends up being extremely useful for something as complex as process reengineering. 4. The ability to work with a variety of models means you can tune the workflows to different cost and performance levels. And you can eventually post train models for specific tasks to tailor the outcomes and eke out gains that aren’t coming otherwise in frontier models. 5. Evals! AI is basically not useful if it can’t be evaluated. Domain-specific evals that let you dramatically improve the performance of your harness for specific workflows just has a crazy long tail given how many tasks there are in the economy. Nearly impossible for one system to be tuned for all of them. 6. Lots of verticals and domains require pricing models that reflect relevant abstractions on top of tokens alone. The ability to price in ways that work for your industry’s consumption model ends up mattering in a variety of spaces. This just touches on some of the things that go into the applied AI layer. But it all adds up to being a huge surface area for being able to sustainably innovate and differentiate.

When you hear that data is the new oil, this is ultimately what that looks like. AI has such a thirst for data that we’re entering an era where it’s valuable almost in any form. In a world of AI, information actually belongs as an asset on the balance sheet. This data sale is just one example of the implications of what that will look like. Ultimately, the way that companies manage and mine their organization’s intelligence will be one of the deciding factors for competitiveness and value creation in the future.

AI spend is nowhere close to hitting any walls. Obviously this data is weighted toward engineering-centric companies, with the top 1% spending $7,500/mo and top 10% spending $660/mo per employee on AI. But the general trend of continued exponential growth is occurring throughout all firm types. And what the top 10% is doing today may easily be what the top 50% is doing in 3 years from now, at least in terms of token volume if not in dollars. This is why there’s still so much opportunity right now in the diffusion of AI in the enterprise. Especially as token costs come down, we will throw larger portions of work at agents. They’ll be scanning all our code for security issues, testing all of our software, writing coding for much larger projects, processing nearly all data, and much more. So this trend has no end in sight.

Pretty remarkable what’s happening with open weights AI right now. We’re seeing models achieve SOTA results on specific tasks, and getting close to frontier on some areas of coding and other domains. The more that open weights is able to maintain only a marginal gap from the…

The fact that open weights models are being discussed credibly at this level of capability should be a huge update for many. The implications of open models getting to frontier performance ensures that you can always have sovereign AI, have the ability to post train for your…

This whole Fable export control situation is actually net positive to regulation discourse. It’s an early peek into what AI regulation would end up looking like at scale when enacted at the model layer instead of the specific application of the AI. The government would have sole…

Token costs are becoming one of the hottest topics for any enterprise I talk with right now. It’s very bullish for AI in general because it means these systems are being used at a scale that wasn’t contemplated before. It also gives way to another form of differentiation that…

Again, maybe counterintuitive, but in the majority of conversations I have with CIOs, CTOs, and CEOs in large enterprises, they are either growing due to AI (in new job functions like FDEs, engineering, etc.) or at a minimum reinvesting efficiency savings back into the business…

Another week on the road meeting with a couple dozen IT and AI leaders from large enterprises across banking, media, retail, healthcare, consulting, tech, and sports, to discuss agents in the enterprise. Some quick takeaways: * Clear that we’re moving from chat era of AI to…

Box just launched its plugin within Codex, which means you can take any content within Box and automate workflows around it using the power of a coding agent. Here's a quick example of processing earnings call documents to extract structured data at scale, which you could then… https://twitter.com/levie/status/2037383187245273477/video/1