When AI Builds Itself: What the Anthropic White Paper Means for B2B Leaders

I'm Tony Uphoff, and this is Uphoff on Media. I've spent 35 years running B2B media, marketing, and technology businesses. Now I'm watching agentic AI rewrite the rules for all of them, and I'm here to help you think through what that actually means for operators, investors, and executives who live in this space. Every episode draws from my Substack, where I publish twice a week. New podcast episodes drop twice a week on Apple and Spotify. Okay, let's get to work. This episode is titled When AI Builds Itself: What the Anthropic Whitepaper Means for B2B Leaders, a guide for CEOs, operators, and investors in media, marketing, and technology. The most useful thing I can offer on the AI disruption isn't a prediction. It's a pattern, one I've lived through before at scale in this sector. I was in the front seat of the roller coaster. At the time, I was running a fast-growing $300 million print-centric B2B media business, and by every internal or external metric, we were succeeding. Then the internet arrived. Not gradually, or at least it didn't feel that way from the operator's seat. One quarter, we're managing circulation strategies, analyzing market share reports and launching events. The next, Employees are stopping you in the hallway. The board is asking pointed questions, and customers want to know, what are you doing about the internet? I had covered technology disruptions for years. The brands I was running served technology professionals. I understood disruption conceptually. What I had never done was lead a business through one in real time, with real consequences, with people's livelihoods and a board's confidence depending on decisions I had to make before I had the answers. No HBR article prepared me for it. No academic framework made it manageable. What got me through it, and what I learned at real cost, was hard experiential learning. The ability to stay calm when the pressure was extraordinary. The ability to move forward before the path was clear. The ability to hold the organization together while simultaneously rebuilding what the organization was actually for. The moment taught me more about leadership than anything before or since. Let me be clear about something before we go further. If you're reading this expecting another dramatic proclamation about the impact of artificial intelligence, stop here and go read something else. There is no shortage of that content out there. AI is saturating every media channel, every business conference, and every board agenda. The volume of coverage is inversely proportional to the operational clarity it produces. Most of it is written by people who are analyzing AI from the outside and with minimal practical insight. Very little of it is written by people who have actually led businesses through a technology disruption of comparable scale. I have, and that informs what I have to say about it. For readers who are digital natives and have built their entire careers in a world where the internet already existed, which is at least half of you and probably more, the internet disruption is history, not lived experience. Some of you were in the early stages of your career when it happened, not yet in leadership positions where the decisions landed on your desk. I'm not telling you this to establish context. I'm telling you this because the playbook from that disruption is the closest thing to a field guide for this one. The internet didn't just change how B2B media companies distributed content. It restructured what content was worth, who would pay for it, and what a media business fundamentally was. The companies that survived and grew were the ones that recognized they were in a different business than the one they thought they were in and acted on that recognition before circumstances forced it. The companies that didn't survive treated the internet as a distribution channel for what they were already doing. They optimized the wrong thing. Until it was too late to optimize the right thing. We are in that moment again, and it may be more disruptive still. Which brings us to the Anthropic white paper and why it matters more than the coverage it has received. The Anthropic Institute titled it When AI Builds Itself, and that title is doing more work than most people realize. This is not a technology document, and it is not another breathless dispatch from the AI hype cycle. It is a document that describes in hard data, the operating environment you are already living in, whether your organization has named it yet or not. If you lead a B2B media, marketing, or technology company or invest in one, the questions it raises are not theoretical. They are arriving in your inbox, your board meetings, and your leadership team meetings right now. What follows is not a summary of the paper. It is not AI hype dressed in business language. It is a field guide written from the operator's seat for the decisions in front of you. What the paper actually says, 3 things that matter. Most AI commentary asks you to imagine the future. The paper shows you the present. Anthropic is documenting in hard operational data, code output per engineer, task completion rates, productivity multiples measured against human performance, what AI is already doing inside one of the most advanced technology organizations in the world. 3 findings have direct consequences for every leader in this sector. First, the development productivity compression is not incremental, it is structural. As of May 2026, more than 80% of the code merged into Anthropic's own codebase was authored by Claude. Before Claude Code launched in early 2025, that number was in the low single digits. The result? In the second quarter of 2026, the typical Anthropic engineer was merging 8 times as much code per day as they were in 2024. In a March 2026 internal poll of 130 Anthropic employees, the median respondent estimated they were producing around 4 times as much output with AI assistance as they would have produced without it. Read that again. 8 times more code per engineer per day, 4 times overall output. at the most sophisticated AI development organization in the world operating at the frontier of what these systems can do. For B2B technology-enabled businesses, companies that use technology to create, develop, and distribute their products and services but are not selling technology directly, this is the most consequential finding in the paper. The engineering backlog that every one of these businesses has been managing as a permanent condition of operating is not a permanent condition. It is a constraint that is beginning to lift. And when that constraint lifts, the question is not how to go faster on the existing roadmap. It is what the roadmap looks like when the binding constraint of the past decade no longer applies. Second, the capability curve is accelerating faster than most executives have internalized. The paper documents that the length of tasks AI can reliably, reliably complete on its own has been doubling roughly every 4 months. In March 2024, Claude could complete software tasks that take a human roughly 4 minutes. By 2026, that horizon extended to 12-hour tasks. The paper projects that tasks taking a skilled person multiple days could come into range before the end of this year, and tasks taking weeks could follow in 2027. This is not a projection from an optimistic analyst. It is a measured trend line from observed performance data documented by the organization building these systems. The pace of change embedded in these numbers is faster than most strategic planning cycles, which means most organizations are planning against a capability baseline that is already out of date. Third, the paper names 3 possible futures, and the one most relevant to B2B operators is already underway. Anthropic outlines 3 scenarios for what comes next. The trend stalls at current capability levels, AI labs see compounding efficiency gains while humans retain direction setting, or— AI achieves full recursive self-improvement. The paper is direct that Scenario 2, compounding efficiency gains with human oversight, is where the evidence points today. And the operational implication of Scenario 2 is named explicitly. A 100-person company can increasingly do the work of a 1,000-person one because each employee sits atop a pyramid of AI agents. That sentence is not a marketing claim. It is a structural shift in competitive dynamics, and it is happening now. The all-too-human reality. Before the frameworks, before the action items, something needs to be named directly here. The decisions described in this post are not technology decisions. They are career decisions, organizational decisions, in some cases, identity decisions. for leaders who have built their professional lives around capabilities that are being restructured around them. The wave of senior leadership departures that we are already seeing— Adobe, Walmart, and others whose CEOs have cited AI transformation as a reason for stepping aside— is being reported as a business story. It is also a deeply human story. Experienced, successful leaders concluding that they are not the right person to finish what they started. That takes a form of self-awareness and courage that is genuinely rare and genuinely hard. I know what it feels like to run a business through a technology disruption where the ground is shifting and the answers are not yet available. The pressure is real. The isolation is real. The temptation to project more confidence than you feel, to buy time with reassuring language while privately figuring it out, is entirely understandable. I've been there. What I learned is that leaders who navigate disruption well are not the ones who pretend— pretended it was simple. They are the ones who stayed honest about the uncertainty while remaining decisive about the next step. Those are compatible. They just take practice. With that said, here's the field guide. For CEOs, the 2 decisions that cannot wait. Decision 1, are you the right person to finish this? This is the question nobody is asking directly in your organization, which means you have to ask it yourself. It is not a referendum on your career or your competence. The leaders stepping aside at Adobe, Walmart, and elsewhere are not failures. Several of them are arguably demonstrating their greatest leadership act, the clarity to recognize that a different phase of a company's life requires a different profile of leader, and the integrity to act on that Recognition rather than hold the seat. The honest version of this question is 3 parts. Do you have the genuine fluency with AI? Not enthusiasm for it, not familiarity with the vocabulary, but actual working knowledge of how these systems operate and where they create and destroy value. Do you have the appetite for the organizational and business model transformation that is required? Not just the efficiency overlay that is insufficient. And do you have the runway, personal energy, board confidence, team trust to lead an 18 to 36 month transformation rather than a 6 month initiative? The Anthropic paper puts this in sharp relief. The organizations that are capturing the gains described in this paper are not doing so by running AI pilots. They are rebuilding how work gets done at a fundamental level. That requires a leader who can hold the organization through sustained uncertainty. Not someone who manages to accordingly plan and waits for the path to clarify. If the answer to all 3 questions is yes, your job is to lead with full commitment and full transparency. If the answer to any of the 3 gives you pause, that pause is worth thinking about before circumstances force that conversation. Decision 2, how do you set expectations that are simultaneously honest and stabilizing? You have 4 distinct constituencies, and they need 4 different conversations. Conflating them is where leaders make the most consequential mistakes. Employees need to hear the honest version before they read the trade media version. Your people are not naive. They can see what is happening around them in the sector. The Anthropic paper documents that AI is already handling more than 80% of code production at one of the world's leading technology organizations. People doing knowledge work, writing, analysis, research, production are drawing their own conclusions about what that trajectory means for their roles. Vague reassurance does not calm that anxiety. It amplifies it because it signals that leadership either doesn't know or won't say. What employees need is an accurate picture of what is changing, a genuine account for what it means for their roles, and a believable plan for how the organization is going to navigate it. That plan does not need to be complete. It needs to be real. Boards need a different frame entirely. Most boards are running on mental models formed in the last disruption cycle. They are simultaneously overexcited about AI as a narrative and underequipped to evaluate it as a business reality. The paper's finding that Anthropic engineers are merging 8 times more code per day than 2 years ago is exactly the kind of concrete, operational data point your board needs, not as a benchmark to match, but as evidence of the pace of change they are governing through. Your job at the board is not to match their enthusiasm or manage their anxiety. It is to be the most credible voice in the room about what AI actually means for your competitive position, your cost structure, and your product roadmap over the next 18 months. Specificity is your currency. Vague AI optimism is not. Investors, particularly in PE-backed companies with defined fund timelines, need the transformation timeline mapped against the investment thesis clearly and honestly. Efficiency gains from AI adoption can show up in 6 to 12 months. Revenue model transformation, new intelligence products, subscription conversion, data licensing is an 18 to 36 month story. If those timelines do not align with fund structure, That is a conversation to have now rather than at the next quarterly review. Customers want proof, not vision. The AI positioning that is landing in the market right now ranges from genuinely substantive to transparently hollow, and your customers can usually tell the difference. What they need from you is specific capability milestones, honest timelines, and a clear account of how your AI investments translate into value for them. The vendors who earned trust in this period will be the ones who underpromised and overdelivered. For operators, 3 decisions about how the work gets done. Decision 1, the development bottleneck is lifting. Are you positioned to act? The Anthropic paper documents something that should fundamentally change how every B2B technology-enabled business thinks about its product development process. When AI systems can write production-quality code 8 times the previous rate. And when that code is passing the review standards of the world's best engineers, the operating assumption that has governed every technology roadmap for the past decade, that engineering capacity is the binding constraint, is no longer reliable. This is not a small adjustment. For most B2B technology-enabled businesses, the operating rhythm has been more demand than we could build for, so we prioritize ruthlessly. And leave significant product opportunity on the table. The AI-driven productivity shift does not eliminate prioritization. It changes the calculus for what is worth building at all. Workflow improvements that were previously uneconomical become viable. Vertical-specific modules for individual enterprise customers become feasible. The long-deferred infrastructure work that has been on every product roadmap for 3 years finally gets addressed. The paper notes that Anthropic engineers are now using AI to fix problems that have accumulated for years, including one case where Claude shipped over 800 fixes that reduced a class of API errors by a factor of 1,000, work an engineer estimated would have taken a human 4 years to complete. The question for operators is not how do we use AI to go faster on an existing roadmap? It is what does the roadmap look like when the constraint we have been managing for 10 years is no longer the binding constraint? Those are different questions, and they produce different answers. Decision 2, is your implementation model a competitive advantage or a liability? The Anthropic paper describes a world where AI agents can work autonomously for hours on complex, open-ended tasks. Tasks with no clear specification where the approach itself has to be figured out along the way. Claude's success rate on the most open-ended tasks reached 76% in May 2023. Up 50 percentage points in just 6 months. For B2B technology vendors, this has direct implication for professional services revenue. If AI agents can execute configuration, data migration, and workflow setup tasks that currently require billable professional service hours, then implementation revenue, a significant line item in many B2B technology contracts, compresses or disappears. Some operators will experience this as margin erosion and fight it. The sharper operators will get ahead of it by treating faster time to value as the selling point it actually is. Shorter implementation cycles mean shorter time to demonstrated ROI, which shortens sales cycles, improves renewal rates, and reduces customer acquisition costs. The math on giving up implementation revenue in exchange for those outcomes is often compelling. The operators who do this analysis now rather than after the revenue line is already under pressure will be in a better position. Decision 3: What not to do and what to do now. The mistakes being made at scale right now in this sector are worth naming plainly because they are tempting mistakes made by smart people under real pressure. Premature restructuring announced before the capability exists. Declaring an AI transformation in a board presentation or a customer communication without the underlying product reality to support it. This is not a positioning mistake. It is a trust mistake, and it is a very difficult one to recover from once customers' employees have seen the gap between the announcement and the reality. Cutting the talent you will need in 18 months. The current wave of AI-driven workforce decisions contains real efficiency gains and genuine AI washing in roughly equal measure. The leaders who will regret decisions made in this period are the ones who eliminated domain expertise in the name of efficiency. Editorial judgment, deep customer relationships, vertical market knowledge— these are exactly the inputs that make AI-generated intelligence valuable rather than generic. Cut them now and you will spend the next 3 years trying to rebuild what you gave away. The Anthropic paper is instructive on this point. The human role that remains irreplaceable, even at the most AI-intensive organization in the world, is research, taste, and judgment. Choosing which problems matter, knowing which results to trust, recognizing when an approach is a dead end. That same judgment applied to your vertical market is your most defensible asset. Protect it accordingly. And ignoring the mid-level leadership layer. CEOs get the attention, investors get the attention, the VP of product, the chief revenue officer, the managing editor, these are the people who will execute the transformation or block it. They are navigating this period with almost no guidance designed for them. The operators who invest in their mid-level leadership now, who bring them into the strategic conversation rather than cascading decisions down to them, will have a significant execution advantage. What to do now? Audit your product roadmap through the lens of a changed development constraint. Identify your implementation revenue exposure and model the unit economics of faster time to value as an alternative. Inventory your proprietary data assets genuinely, not aspirationally, and protect the domain expertise in your organization as a strategic asset it is because It is what will make your AI-generated intelligence valuable rather than generic. For investors, the filter that matters now, the 3-characteristic test. The PE-owned B2B media and marketing landscape contains a wide distribution of asset quality, and AI is going to widen that distribution significantly. The Anthropic Papers Scenario 2 compounding efficiency gains with a 100-person company doing the work of a 1,000-person one is not a uniform benefit across the sector. It is a capability multiplier, and what it multiplies matters enormously. Some companies in this sector are being structurally disintermediated. Media brands whose primary value was information access, marketing services businesses whose primary value was human execution of repeatable tasks, Research businesses whose primary value is data aggregation rather than insight synthesis. For these companies, AI does not create a rescue opportunity, it accelerates the pressure. The companies that can use this moment to break out share 3 characteristics. Number one, a defensible proprietary data asset, not licensed data, not aggregated public data, behavioral data, transaction data, intent data that comes from a genuine and exclusive relationship with an audience or customer base. Number 2, a trusted audience or customer relationship, the kind of relationship where the brand is the reason people show up, not just the container for the information. And number 3, a leadership team willing to pursue genuine business model transformation, not an efficiency overlay on a declining model, a fundamental rethinking of what the business is and what it sells. The efficiency overlay is the more common response because it is easier to underwrite, easier to model, and easier to defend in a board meeting. It is also insufficient. The B2B media and information companies that navigate this period successfully will be the ones that use AI not to do the same thing cheaper, but to do something fundamentally more valuable, moving from content delivery to intelligence execution. From audience access to workflow integration, from quarterly reports to continuously updated decision support. The timeline reality. The Anthropic paper's capability trajectory has a direct implication for investment timelines that most sponsors are not explicitly addressing. If AI systems can complete tasks that take human days by the end of 2025 and the tasks that take human weeks in 2027, then the window for business model transformation is not a 3 to 5 year horizon. It is compressed. Companies that begin genuine transformation now have a meaningful advantage over those that begin in 18 months. Not because the technology would be unavailable, but because the organizational learning curve, the product development cycle, and the audience trust that makes intelligence products valuable all take time to build regardless of how capable the underlying AI systems become. The timeline reality for PE sponsors is specific and needs to be addressed. Efficiency gains show up in 6 to 12 months. Revenue model transformation is an 18 to 36 month story. Most fund structures were not built around that gap. The result is pressure to declare transformation victories on an efficiency story that acquirers will correctly identify as incomplete. That misalignment is a risk factor in its own right. Sponsors and portfolio company leadership need to have an explicit conversation about whether the fund structure can actually support genuine transformation. Most are not having it. What AI-native leadership looks like in diligence. The standard diligence question about AI readiness, how are you using AI in your business, produces answers that are almost universally unreliable because every management team has a version of that answer that sounds good and and very few have any substance behind it. The better questions: What proprietary data does this business own that gets more valuable as AI systems get better at synthesizing it? What is the gap between the intelligence this business could produce from its data and the intelligence it is currently producing? Has the leadership team demonstrated the ability to move through ambiguity and make structural decisions rather than incremental ones? And specifically, who on this team has led through a prior technology disruption cycle at the operator level, not the advisory level? That last question matters. The Anthropic paper is clear that the human role that persists even as AI takes over the execution of complex tasks is judgment about which problems are worth solving. That judgment in a business context comes from having made hard calls under real pressure with real consequences. It cannot be learned from a course or acquired by reading white papers. The difference between leaders who have run businesses through disruption and leaders who have studied businesses through disruption is significant, and it is not always visible in a management presentation. The vertical business intelligence window. For small and mid-sized B2B media and information companies, the Anthropic paper describes something specific and time-bound. The infrastructure cost of building a sophisticated intelligence product has collapsed. The paper documents that Anthropic engineers are now using AI to complete work that would have taken humans years, not through artificial intelligence as a concept, but through AI agents executing complex multi-step workflows autonomously over hours or days. The same capability that is transforming software development at Anthropic is available. To a focused B2B media company building a vertical intelligent product today. 5 years ago, building a genuine intelligence product on top of a vertical dataset asset required a data science team, an engineering team, and capital that most B2B media companies just didn't have. Today, a focused team of 2 or 3 people can build and deploy an agentic pipeline that ingests, synthesizes, and services vertical market intelligence at a level of sophistication that it would have required an enterprise analytics operation in 2020. The capability documented in the Anthropic paper is not confined to Anthropic. It is available now at a cost structure that has never existed before. This matters because the moat in B2B media has always been the proprietary relationship with the audience, the behavioral signals, the transaction data, the intent signals that come from a community that trusts the brand. What was missing was the productization capability. Agentic AI closes that gap. The companies that recognize this and move are positioned to make the long-predicted but never quite arrived shift from advertising revenue to audience intelligence revenue. The companies that treat AI as a content production efficiency tool are going to miss the more consequential Opportunity. The window is open. The Anthropic paper makes clear it will not remain open indefinitely. The large players are moving, the capability is diffusing, and the organizations that build institutional AI experience now are accumulating a compounding advantage over those that wait. The advantage small and focused B2B media companies have right now is speed, domain specificity, and audience trust. Those advantages are real and durable, but they require action to be realized. What to do now. If you're a CEO, have the honest internal conversation before your board forces it. Assess your own AI fluency with the same rigor you would apply to assessing any other capability gap in your organization. Begin the 4 constituency communication sequence now, starting with your employees. They need to hear The honest version from you before they read the trade media version and make the organizational decision about AI leadership explicit. Who in your organization owns this transformation with real authority and real accountability, not as an addition to an existing role, but as a primary responsibility? If you're an operator, audit your product roadmap through the lens of a change development constraint. The 8x productivity figure from the Anthropic paper is a leading indicator of where this goes, not a ceiling. Identify your implementation revenue exposure and model the unit economics of faster time to value as a strategic alternative. Inventory your proprietary data assets genuinely, not aspirationally, and build toward the intelligence product your vertical market will pay you for. Protect the domain expertise in your organization as the strategic asset that it is. If you're an investor, apply the 3-characteristic filter to every portfolio company: proprietary data asset, trusted audience relationship, leadership capability of genuine transformation. Have the explicit conversation with each portfolio company without— about transformation timeline versus fund timeline, and be honest about the misalignment where it exists. Build an AI-native leadership evaluation framework before your next diligence process because the standard questions are producing unreliable answers. The capability trajectory in the Anthropic paper compresses your planning horizon. Act accordingly. The path forward. I came out the other side of the internet disruption having learned things about leadership that I could not have learned any other way. The cost of that education was real in organizational stress, in strategic mistakes made and corrected, in the gap between the confidence I projected and the certainty I actually had. What I know now is that the leaders who navigate disruption well share a common characteristic. They do not wait for clarity before they act. They move toward the uncertainty rather than away from it. They stay honest with the people depending on them, and they build the organizational capacity to learn and correct in real time rather than organizational theater of having all the answers. The Anthropic white paper is telling you in hard data that the ground is shifting. The question is not whether this disruption is real. The paper answers that. The question is what kind of leader you are going to be in response to it. That is the most important decision of all, and it is entirely yours to make. Well, that's it for this episode of Uphoff on Media. If this was useful, please rate the podcast, and if so inclined, write a quick review on Apple Podcasts. It helps other executives and operators find the show, and I read every one. Thanks. Back at it next week.