Enterprise Software’s Publisher Moment

Enterprise Software’s Publisher Moment

Microsoft recently announced it was folding its consumer and corporate Copilot chatbots into one application. The stated reason was to be more competitive with Claude and ChatGPT. The more interesting story is what the move quietly suggests. Microsoft built a copilot for every product. It watched the sprawl confuse everyone. Now it is collapsing all of it into a single interface that sits above the products. Word, Excel, and Teams become capabilities. The interface becomes the thing you actually use. Hold that picture. It is the whole argument. Software was always a publishing business I spent a career running businesses that looked like opposites and were built on the same engine. InformationWeek. ThomasNet. Pipeline360. Media on one side, software on the other, both running near 80% gross margins on the same model. Software had the best margins in the history of business because it ran a publishing model. Produce once. Sell the identical version thousands of times. High fixed cost to build the first copy, near-zero marginal cost on every copy after. That is not a SaaS invention. That is the printing press. Software just perfected it. No paper, no trucks, instant distribution, infinite copies. Enterprise software was a pure technology business. But those 80% gross margins were publishing margins. So were Oracle's. So were SAP's. The technology was new. The economics were publishing. That distinction did not matter for thirty years. It matters now. A model built on producing once and selling many depends on scarcity holding. Remove the scarcity and the margin goes with it. That is not a theory. We watched it happen to publishing. Software runs the same model. It faces the same disruption. The margin depended on two kinds of scarcity, and AI attacks both The publishing margin rested on two things being true. Production had to be scarce, because building the product was hard. And the product had to be standardized, because you only earned the margin by selling everyone the same version. AI goes after both at once. It collapses production scarcity. When an agent can generate a custom workflow in tokens, the premise that buying the standard product is cheaper than building your own starts to invert. That is the vibe-coding fear. It is already priced into the multiples. Adobe, Salesforce, ServiceNow, and Workday have all watched their valuations decline. Then it breaks standardization, which is quieter and worse. The margin existed because everyone bought the same product. The moment software becomes generated per customer, you have destroyed the very thing that created the margin. That is not lost market share. That is a trade, from a publishing cost structure to a services cost structure. And here is the knife. The replacement model has real marginal cost. Running the model is not free. Every answer burns compute. Tokens meter. Uber reportedly burned its entire 2026 AI budget in four months, with engineers running $500 to $2,000 a month each. So the companies that win the agentic transition may inherit worse unit economics than the incumbents they replace. They are swapping a zero-marginal-cost business for a compute-metered one. Publishing gave software its margins. The agentic model threatens to hand software publishing’s actual fate. Seats were circulation. The agent is the aggregator. I have seen this movie. I sat in the publisher’s chair while it played. Print made its money on a simple model. Produce once, sell many. Priced by circulation. Protected by scarce distribution and an owned audience. Digital killed it three ways. It made distribution free and infinite. It unbundled the product down to the single article. And it inserted an aggregator, Google, between the publisher and the reader. The aggregator captured the value. The publisher became commodity inventory feeding someone else’s interface. Now run the same play on enterprise software. Produce once, sell many. Priced by seat. Protected by switching costs and standardization. Seats are circulation. Seat-based pricing is a subscription to copies, the same model as a magazine’s subscriber file, and it is already cracking. In 2025 alone, the 500 largest SaaS and AI companies changed their pricing more than 1,800 times, the most repricing in a single year since SaaS first moved to subscriptions. The direction is away from seats, toward usage and outcomes. The agent is the aggregator. It is the layer that sits above the software, reads and writes across every system, and captures the value while the applications underneath become inputs. Vibe coding is the collapse of production scarcity. Every force that took print’s margins is loading up against software’s. Everyone can see the pieces. Almost no one sees the pattern. They never sat in the publisher's chair while it happened. The services pivot is already here Watch what the winners are actually doing. Microsoft recently launched the Microsoft Frontier Company. Two and a half billion dollars. Six thousand engineers. Their job is to embed inside customer operations and make the AI actually deliver. Two days earlier, AWS put a billion dollars behind the same idea. OpenAI and Anthropic launched their own versions in May, backed by private equity. This is Forward Deployed Engineering, and it is the opposite of the publishing model. Produce once and sell many needs no humans in the loop per customer. That was the entire point. Embedding six thousand people inside client operations is a services business. It carries marginal cost. It scales with headcount, not with copies. When Microsoft, AWS, OpenAI, and Anthropic all race to build these armies in the same quarter, they are admitting the produce-once model no longer delivers the outcome on its own. Models are becoming commodities, cheaper and more alike by the month. The money is in selling the services that make AI pay off inside a company. That is the publishing margin giving way to the services margin, in real time. Judson Althoff, who runs Microsoft's commercial business, resisted the label. What Microsoft is building, he wrote, "goes beyond" Forward Deployed Engineering. It will be "the largest, most capable, outcome-driven engineering organization in the industry." Read that again. “Outcome-driven”. When you sell outcomes delivered by embedded humans, you are a consulting firm. The denial is the tell. I watched B2B media make this exact move. When the produce-once margin broke, publishers pivoted to services. Agencies, custom content, events, advisory work. Software is now doing the same thing, in the same quarter its multiples are compressing. Microsoft’s stock is down more than 20% this year, the worst among the mega-caps, on precisely this fear. This is not a prediction about where software is heading. It is a press release about where it already went. Where it hits, and where it doesn't Here is the distinction that separates casualties from survivors. The margin compression does not hit the whole category. It hits the application and workflow layer hardest. It does not hit the governed-data layer the same way. Systems of record with real data gravity, permissions, compliance, and audit trails still hold. That is not sentiment. It is structure. An agent needs explicit rules and clean, governed data to act safely, and that is exactly the unglamorous machinery SAP, Oracle, and the enterprise data platforms spent decades building. Migrations off those systems famously fail. This is why they are the survivors and thin-workflow SaaS is the casualty. If your product is a nice interface wrapped around a database and some workflow logic, the agent reproduces you. If your product is the governed system of record the agent has to trust, the agent needs you. The cost argument comes with a caveat too. Every time the model answers, it costs money to run. That cost is called inference, and it is falling fast, so the marginal-cost hit softens over time. The catch is that usage is expanding faster than prices are dropping, so net spend still climbs. The unit economics improve while the bill goes up. Demotion, not death The incumbents are not going to go away. That was never the real risk. The real risk is demotion. From the system of intelligence, where the money and the CEO relationship live, down to the system of record, the governed database humming in the background that the agent treats as one input among many. Microsoft is the only one hedged against it, because it is both the incumbent and the distribution proxy for a frontier model. SAP, Oracle, Salesforce, and ServiceNow are long a single bet: that the model layer stays a component they can rent and wrap, rather than the control point that reduces them to plumbing. SAP embedding Claude across its Business AI platform reads two ways. Smart hedging, or an admission that the intelligence is no longer theirs to own. Enterprise software is a technology business that ran on publishing economics. Those economics produced the greatest margins in the history of business. They are now producing the same displacement that came for publishing. The model that made software the best business in the world is the model that puts it at risk. The survivors will not be the companies with the best AI. They will be the ones who own the governed data the intelligence layer cannot reproduce. The views expressed in Uphoff on Media are entirely my own. They don’t represent the opinions of any company I’ve led, any board I’ve sat on, or any investor who’s had the pleasure of debating strategy with me over the years. If something I write here sounds brilliant, I’ll take full credit. If it turns out to be wrong, I was clearly misquoted by myself. “Uphoff on Media” is published by Tony Uphoff, Founder and Managing Partner of Uphoff Advisory, LLC [https://uphoffadvisory.com/]: a strategic advisory practice for founders, CEOs, and investors in B2B information, marketing, and technology. The businesses that drive business. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tonyuphoff.substack.com [https://tonyuphoff.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

The Teammate That Never Logs Off

Anthropic and Salesforce announced Claude Tag on June 23, 2026. Most of the coverage framed it as a Slack upgrade. That positioning is wrong. What shipped two weeks ago is a signal about where enterprise AI is going, and the implications go well past faster meeting summaries. Here’s what you need to understand about what actually changed, which functions will be impacted first, and what the next 24 months look like for anyone running a knowledge work organization. What Claude Tag Actually Is The previous Claude integration in Slack was a chatbot. You asked it something. It answered. Conversation over. Claude Tag is structurally different. It joins Slack channels as a persistent member. It accumulates context from the conversations it observes. It connects to data sources, codebases, and external tools. It works asynchronously, accepting tasks and continuing to work for up to days at a time without being prompted again. If “ambient” mode is enabled, it proactively flags relevant information across channels before you ask. Claude Tag represents a fundamental shift. Every prior enterprise AI product, chatbot or copilot, waited for the human to initiate. Claude Tag doesn't. This is not a tool you use. It is a participant that works alongside your team. Anthropic shared a data point that shows a glimpse of the future. Internally, 65% of the company’s product team code is now generated by its internal version of Claude Tag. The same capability is spreading beyond engineering into product metrics, support tickets, and bug triage. That number is not a demo. It is a production deployment stat from the organization that built the model. The Efficiency Lens Is the Wrong Lens Every enterprise rollout deck for Claude Tag will be built around efficiency. Faster drafts. Better summaries. Fewer handoffs. Those gains are real. But efficiency assumes the workflow stays the same and you move through it faster. Claude Tag changes what the workflow is. The deeper shift is about which tasks require a human at all. Most knowledge work is not judgment work. It is a coordination tax: status updates, context-passing between handoffs, translating what the product team decided into what the engineering ticket says, pulling together account history before a renewal call, following up on threads that went quiet. Claude Tag, with channel memory and connected data sources, eliminates that layer as a human responsibility. Deloitte’s 2026 State of AI in the Enterprise found that only 34% of organizations are truly reimagining their business models around AI. The other 66% are using AI at the surface level while leaving the underlying workflow structure intact. Claude Tag is designed to force the deeper question. Its ambient architecture makes workflow restructuring the path of least resistance, not an optional advanced use case. Where This Hits First Four functions are immediately in Claude Tag’s line of fire. Sales operations and revenue enablement. Pipeline hygiene, CRM update discipline, competitive context-gathering, call prep assembly: these are coordination-heavy, low-judgment tasks that consume time from people who should be selling. AI already saves sales professionals an estimated 12 hours per week on these tasks. Claude Tag, connected to CRM data and deal channels, does not speed up that work. It eliminates it as a human task entirely. Product and engineering handoffs. The translation layer between product decisions and engineering tickets is almost entirely pattern work. Salesforce is already deploying Claude Code across its global engineering organization, using Slack MCP integration to pull spec context from channels and feed decisions directly into development workflows. This is not a pilot. It is production at scale. Customer success and support triage. Deloitte identifies customer support as the highest-impact early function for agentic AI. Issue routing, escalation write-ups, prior case context retrieval, all high-volume, low-variance tasks. The CS role shifts to relationship judgment and resolution authority. The coordination layer disappears. Marketing and content operations. Campaign briefing, competitive tracking, and content calendar management are high-coordination, medium-judgment tasks that map directly to ambient AI’s strengths. The briefing doc that takes a junior marketer two hours to assemble becomes a Claude Tag output that a senior marketer reviews and approves. What Post-Claude Tag Workflows Look Like Take the renewal motion in sales. Today, a CSM pulls usage data, reviews support history, drafts a prep document, circulates it for input, typically two to four hours of work before the call. Post-Claude Tag, the system monitors the account channel, pulls CRM and product usage data, and surfaces the brief 48 hours before the call without being prompted. The CSM arrives prepared. The prep assembly is gone as a task. The same logic applies to product-to-engineering handoffs: spec-to-ticket translation becomes a review function rather than a drafting function. RBC Wealth Management is already running a version of this in compliance workflows, using Claude in Agentforce to handle advisor meeting prep and portfolio summaries so advisors focus on client relationships rather than administrative assembly. The pattern is consistent across all three: humans move from producing the work to approving it. The Employee Unrest Question This is the part that enterprise leadership is not talking about yet. It should be. When Claude Tag is deployed in a channel, it learns from what people do. Every workflow it observes becomes institutional memory. The employees who are most diligent: who document their work clearly, communicate decisions explicitly, and close loops reliably, are the best contributors to the system that may eventually reduce the need for their role. The people most likely to cooperate fully are the most exposed. This dynamic has a name in manufacturing and consulting contexts: participatory deskilling. It generates serious resistance when workers recognize what is happening. The difference here is speed and audience. This cycle is not happening to assembly-line workers. It is happening to knowledge workers with graduate degrees, professional identities, and the vocabulary to articulate their grievances clearly. The data already shows the anxiety. ManpowerGroup’s 2026 Global Talent Barometer found that regular AI usage among workers jumped 13%, but confidence in using technology fell 18% in the same period. The term the research uses is “job hugging”: workers holding tightly to existing tasks because they understand that mastery of the old task was their job security. As of May 2026, over 113,000 tech workers had been laid off across 179 companies, a pace 33% higher than the same period in 2025. The workers watching those announcements are not going to be neutral observers when asked to onboard Claude Tag into their channels. The Next 24 Months: Six Predictions 1. The coordination PM role disappears faster than anyone expects. Product management will bifurcate. Strategic PMs: those who own the vision, set priorities, and interface with customers, become more valuable. Coordination PMs: those whose primary job is translating decisions across teams and tracking ticket status, get absorbed into agentic workflows within 18 months. This is already visible at Anthropic internally and at Salesforce’s engineering organization. It will spread to every enterprise running Claude Tag or a comparable ambient system. 2. Enterprise trust in AI judgment will be tested publicly and visibly. The first major Claude Tag failure, a consequential decision made on bad ambient context, a sensitive conversation flagged in the wrong channel, a task completed incorrectly over several days before anyone noticed, will become a governance inflection point. Deloitte’s 2026 data shows that confidence in AI governance drops sharply when the question turns from strategy to operational readiness. Ambient AI operating asynchronously across private channels for extended periods is a different governance challenge than a chatbot answering a discrete question. The frameworks do not exist yet. They will be built reactively, after the first visible failure. 3. The knowledge workforce shapes like an hourglass. PwC’s 2026 analysis predicts that as agents take on midlevel work, the knowledge workforce concentrates at the junior and senior levels: junior workers who are AI-native and senior professionals whose strategic judgment is irreplaceable. The middle layer of experienced-but-not-senior knowledge workers faces the most pressure. These are the people whose work is pattern-rich enough to automate but whose institutional knowledge has not yet been formalized into data Claude Tag can learn from. They are the most vulnerable and the least protected by current reskilling narratives. 4. Slack becomes the operating system of record, not just a messaging layer. The average enterprise now uses more than a thousand applications. Employees lose significant productive time to context-switching between them. Every major AI lab has concluded that the right place to intercept that problem is the team-chat surface where work is already being coordinated. Claude Tag is Anthropic’s move into that layer. Microsoft has GitHub Copilot in Teams. OpenAI launched Workspace Agents in April. This is not a product competition. It is a platform competition for where enterprise AI lives. The winner will have more influence over enterprise workflow design than any SaaS vendor of the previous decade. 5. Governance becomes a competitive differentiator, not a compliance checkbox. The EU AI Act’s high-risk system enforcement provisions take effect August 2, 2026, 40 days from Claude Tag’s launch. The enterprises that build rigorous governance around ambient AI deployments: clear channel access policies, explicit human review checkpoints, audit trails for AI-initiated work, will not just manage compliance risk. They will build the trust infrastructure that enables more ambitious AI deployments downstream. Governance is not a constraint on competitive AI use. It is the prerequisite for it. 6. The first wave of restructuring produces the case studies that set the terms for the second. The companies that aggressively cut headcount in 2026 will either prove that AI can truly replace those workers, or discover they cut too deep and need to hire back. Those outcomes arrive in 12 to 18 months and will heavily shape how the second wave of enterprises approaches ambient AI deployment. The organizations that instrument their deployments carefully, measuring what actually gets done better, what gets dropped, and where human judgment proves irreplaceable, will have significant advantages over those that deployed on the efficiency narrative alone. The Central Bet Salesforce’s new President and CPO Rohan Kumar put the strategic thesis plainly: the future of enterprise software is “headless” and “ambient.” Headless means no interface for the human to drive. The human becomes the exception handler, not the operator. That is a meaningful bet. It assumes that trust in AI judgment will increase faster than resistance to AI autonomy. That is the central question of the next 24 months in enterprise software, and Claude Tag is the most visible live test of that thesis in the market right now. Anthropic describes Claude Tag as “the beginning of an evolution.” That positioning is accurate. But evolutions, in business as in biology, produce winners and losers. The organizations and individuals who understand what is actually changing, and design around it deliberately rather than reacting to it after the fact, will be the ones who shape what comes next. The rest will find out what “headless and ambient” means the hard way. The views expressed in Uphoff on Media are entirely my own. They don’t represent the opinions of any company I’ve led, any board I’ve sat on, or any investor who’s had the pleasure of debating strategy with me over the years. If something I write here sounds brilliant, I’ll take full credit. If it turns out to be wrong, I was clearly misquoted by myself. “Uphoff on Media” is published by Tony Uphoff, Founder and Managing Partner of Uphoff Advisory, LLC [https://uphoffadvisory.com/]: a strategic advisory practice for founders, CEOs, and investors in B2B media, marketing, and technology. The businesses that drive business. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tonyuphoff.substack.com [https://tonyuphoff.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

Loop Engineering: The Bridge to Workflow Transformation

A few posts ago, I made the case that Vibe Coding [https://tonyuphoff.substack.com/p/vibe-coding]eliminates the queue: the wait for an engineer to build the thing you need. Last post, Solo Scale [https://tonyuphoff.substack.com/p/solo-scale-the-new-business-model] showed what happens when you remove the ceiling on how much one expert can supervise. Neither one, by itself, changes how your business actually works. Loop Engineering is what does. It’s the bridge between “I can build this” and “this runs the business.” And it’s the part most business leaders haven’t discovered yet, even as it’s already quietly reshaping how businesses operationalize Agentic AI. That quieter shift is starting, and the people closest to the technology have already made it. In June, Boris Cherny, who leads Claude Code at Anthropic, said he no longer prompts the model at all. He writes loops that do the prompting for him, and his job now is to write the loops. Engineers at Google and OpenAI have been saying versions of the same thing. The shorthand term is “loop engineering.” Here is why it matters to you, even if you never write a line of code: the move from prompting to loops is not a coding story. It is an operating story. And it is the piece that turns everything agentic AI has promised into something a business can actually run on. From Running the Tool to Designing a System Start with the plainest version. Prompting asks a model for an answer. You type a request, read what comes back, refine it, ask again. You are running the tool the entire time, one turn after another, and the quality of the work depends on your skill at the exchange and your patience for it. You are the bottleneck. A loop is different. A loop is a short set of instructions that says what you are trying to accomplish, what the system may use to get there, what counts as evidence it succeeded, what to do when it fails, what to remember between runs, and when to stop. Then it runs: checking its own output against your standard and continuing until the result is good enough to use. Prompting gets you one good answer. A loop builds the thing that produces good answers without you in the room. What a loop really is Strip the code away and a loop is something every good operator already carries in their head: a standard for what good looks like, and the judgment to know when work meets it. What you have never had is a way to enforce that standard at scale, on many fronts at once, without being present for every step. That is what a loop gives you. You define the goal, the tools, the evidence, the failure response, the memory, and the stopping point once. The system carries your standard forward every time it runs. The scarce thing just moved For two years the value lived in the interaction: knowing how to prompt well, working the tool skillfully turn by turn. That skill is being commoditized. The new scarcity sits upstream of it: knowing what good looks like, and being able to specify it precisely enough that a system can hold to it without you. The premium is moving from doing the work to knowing what good looks like, and being able to say so precisely. That reallocation changes what each of you should do next. Here is where to start, depending on where you sit. If you run the work: deploy Pick one process you run over and over that leans on your judgment: a weekly competitive brief, lead qualification, a first-pass content review, campaign QA. Something with a clear standard for good that you currently apply by hand. Write that standard down: the goal, what the system may draw on, what a good result looks like, what to do when it isn’t. Then let a loop run it while you check the output instead of producing it. Start with one. Get it right before you add a second. The skill you are building is not technical. It is learning to state a standard clearly enough that a system can hold it. If you run the company: reallocate If value is moving from execution to judgment, your org chart is about to feel it. Ask yourself three questions. What are you staffing, promoting, and paying for today that is execution capacity about to get cheap? Who on your team can author standards, not just follow them, because that is the scarce talent now. And the pointed one: which part of the business would you most want running on your standard instead of your headcount, and what is keeping you from starting there? The executives I watch most closely are not asking whether this is real. They are asking where to apply it first, and how fast they can move. If you fund the work: evaluate When execution is cheap and judgment is scarce, the shape of a good business changes. The moat is no longer how much a company can build, or how fast. It is the quality of the standard encoded in its systems: proprietary judgment about what good looks like in a specific domain, enforced by loops a competitor cannot easily copy. In diligence, the question moves from how big is the team to whose judgment is in the loop, and will it hold. And learn to tell a real loop from loops-as-lipstick: genuine encoded expertise versus a thin prompt in a nice interface. If you’re starting something: build This is the part that should keep you up at night, in the good way. If one expert, armed with the right standards and a set of loops, can produce what used to take a department, then a single person with deep domain knowledge can start and run a business that used to require a team and a raise. Not by writing code. By encoding what they know about their field into systems that run. The barrier to starting has never been lower for the person whose real asset is expertise. So the question is simple: what would you build if execution were no longer the thing in your way? One hard truth first A loop running unattended is also a loop making mistakes unattended. The skill of loop engineering is not the loop. That part is simple. It is knowing where the system will be confidently wrong, feeding it inputs you can trust, and keeping a human on the standard. The people learning this the hard way keep arriving at the same lesson: the model rarely fails because it is dumb. It fails because it was handed bad inputs or a fuzzy standard, and it faithfully delivered exactly what it was told to. Verification stays human. That is not a limitation to engineer away. It is the job. Three posts, one argument Now step back. Vibe Coding removed the constraint on who can build. Anyone who can describe software can produce it. The queue disappeared. Solo Scale named what becomes possible at the far end: an expert-led, agent-powered business running at software margins, one person, or a very small team, producing what once took many. Loop engineering is the bridge between them. It is the mechanism that turns “anyone can build” into “one expert can run a whole business.” Without loops, Vibe Coding gives you faster one-off builds, impressive, but still one turn at a time, still bottlenecked by your bandwidth. With loops, those builds become systems that run: your judgment, encoded, working on many fronts at once. That is the machinery beneath Solo Scale. It is what makes the margins real. And the same mechanism, pointed at an organization instead of a solo operator, opens a much larger door. When you stop automating single tasks and start encoding judgment into loops that run across a whole process, you are no longer improving a workflow. You are transforming it, rethinking how work moves through the business from end to end. That is the next frontier, and loop engineering is the way in. So: three posts, one argument. Vibe Coding opened the door. Solo Scale showed the room on the other side. Loop engineering is how you walk through, and it is the same threshold every business will cross on the way to changing how work actually gets done. The queue is gone. The ceiling is going. What is left is the standard, and whether you can name it clearly enough for a system to carry it. That is the work now. The views expressed in Uphoff on Media are entirely my own. They don’t represent the opinions of any company I’ve led, any board I’ve sat on, or any investor who’s had the pleasure of debating strategy with me over the years. If something I write here sounds brilliant, I’ll take full credit. If it turns out to be wrong, I was clearly misquoted by myself. “Uphoff on Media” is published by Tony Uphoff, Founder and Managing Partner of Uphoff Advisory, LLC [https://uphoffadvisory.com/]: a strategic advisory practice for founders, CEOs, and investors in B2B information, marketing, and technology. The businesses that drive business. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tonyuphoff.substack.com [https://tonyuphoff.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

Solo Scale: The New Business Model AI Just Made Possible

There’s a conversation happening about what AI means for software creation. Elena Verna at Lovable calls it Mom-and-Pop SaaS. The thesis is smart: as the cost of building software collapses, domain experts — not just developers — become builders. She’s right. And it’s only half the story. The bigger shift isn’t happening in software products. It’s happening in services. Agentic AI and vibe coding aren’t just lowering the cost of building apps. They’re dismantling the core constraint of every services business ever built: the ratio of expert time to revenue. That constraint: one expert, one engagement, finite hours, is what kept professional services firms trapped at 10–20% EBITDA margins while SaaS companies ran 60–80%. That constraint is ending. The Old Model Was Always a Labor Problem Services businesses sell expertise. But expertise doesn’t scale. You hire more experts, manage more overhead, and your margins compress. The best consulting firms in the world, running mature operations, optimized for utilization, still average under 10% EBITDA. The 2025 SPI Professional Services Maturity Benchmark, covering 509 firms managing $63 billion in revenue, put the industry average at 9.9%. Historic low. The fundamental problem is structural. Services revenue scales linearly with headcount. Software revenue doesn’t. That gap is what the entire VC-backed SaaS industry was built to exploit. Now the gap is available to everyone. What Agentic AI Actually Changes Gartner projects that 40% of enterprise applications will embed task-specific AI agents by end of 2026, up from under 5% in 2025. That’s an enterprise story. The more important story is what’s happening at the edges. Agentic AI doesn’t just make individuals more productive. It allows one person to deploy systems that do execution work that previously required a team. Research. Analysis. Outreach. Reporting. Monitoring. Drafting. These aren’t tasks an AI does instead of you. They’re tasks an AI does alongside you, continuously, at scale, without adding headcount. The economics flip. A solo practitioner running AI agents across ten client engagements simultaneously isn’t running a consulting business anymore. They’re running something new. Vibe coding accelerates this further. The market for AI coding tools hit $4.7 billion in 2026 and is projected to reach $12.3 billion by 2027. Sixty-three percent of vibe coding users identify as non-developers. The cost of building a functional product has dropped from roughly $200,000 to about $5,000. Build timelines compressed from six months to six weeks. The combination matters. Agentic AI handles execution at scale. Vibe coding eliminates the technical barrier to building the systems and tools that deliver it. A domain expert with deep practitioner knowledge and a willingness to learn these tools now has access to a business architecture that didn’t exist two years ago. Introducing Solo Scale I want to name what’s emerging, because it’s distinct from anything we’ve described before. Solo Scale is a new category of AI-enabled business: expert-led, agent-powered, running at software margins. The entity at the center of a Solo Scale business is what I call an Expert Scale operator. A practitioner who combines deep domain knowledge with agentic AI systems to deliver services at a leverage ratio that traditional firms can’t match. This isn’t freelancing. Freelancing trades time for money at a fixed rate. Solo Scale trades expertise for outcomes, with AI handling the execution volume. The margin structure is fundamentally different. This isn’t SaaS. There’s no product to sell at scale without the practitioner. The expertise is the product. But the delivery infrastructure is AI. Three things define a Solo Scale business: 1. Domain depth. The operator has years, often decades, of practitioner knowledge in a specific field. Not generalist knowledge. Vertical expertise that took time to accumulate and can’t be easily replicated by a model alone. 2. Agentic infrastructure. AI agents handle the execution layer: research, analysis, drafting, monitoring, outreach, reporting. The operator sets strategy, reviews output, exercises judgment. The agents do the volume. 3. Software-like margins. Because the execution scales without proportional headcount, the margin structure approaches software economics. Not 10–20%. Closer to 40–70%. Territory that professional services firms have never accessed before. Elena Verna’s data from Lovable points in this direction: 80% of builders intend to monetize. 35% are already generating revenue. But the real money won’t be in software products built by domain experts. It will be in services businesses run by domain experts, using software as the delivery infrastructure, not as the product. The Five Categories That Emerge First Not every services domain is equally positioned for Solo Scale. The highest-value early categories share three characteristics: high complexity (which protects against commoditization), high existing fee levels (which means the margin expansion is large in absolute dollars), and repeatable workflow structures (which AI agents can execute reliably). Here are the five categories where I expect Solo Scale businesses to emerge first: including two that may surprise you. 1. Specialized B2B Advisory Management consulting, strategic advisory, and fractional executive services are the clearest near-term Solo Scale opportunity. The practitioner provides judgment, relationships, and pattern recognition accumulated over a career. AI agents handle the research, analysis, competitive monitoring, and report generation that junior associates and analysts currently do. A senior advisor running five engagements simultaneously, with AI handling the execution layer across all five, is running a business with fundamentally different economics than a traditional consulting firm. No associate overhead. No utilization management. No bench time. The margin structure changes completely. This is not hypothetical. It’s happening now, across advisory practices of every kind. 2. Vertical Content and Intelligence B2B information and intelligence businesses have always been constrained by the cost of producing high-quality, specialized content at volume. The editorial team is the cost structure. AI changes that equation. Consider a practitioner with genuine vertical expertise: supply chain, procurement, healthcare operations, industrial technology. That practitioner can now produce research reports, market analysis, buyer intelligence, and curated content at a volume and quality that previously required a full editorial operation. The practitioner provides the framing, the sourcing judgment, and the editorial voice. AI handles the production volume. The Vertical Intelligence Company I’ve written about in this series fits squarely here. Solo Scale is the business model underneath it. 3. AI-Augmented Professional Services Law, accounting, financial advisory, and compliance-adjacent services are early-stage Solo Scale territory. Regulatory complexity and liability requirements mean pure automation can’t displace the practitioner. But AI dramatically compresses the time required for the research, drafting, and analysis that underlies professional work. A specialist attorney who handles a narrow, high-value area of commercial law, and deploys AI agents to handle the research, precedent review, and initial drafting, operates at a leverage ratio that a traditional associate-dependent firm cannot match. Same expertise. Dramatically lower cost structure. 4. Supply Chain and Logistics Intelligence (The Surprising One) Here’s a category most people aren’t talking about yet. Supply chain consulting has historically required large teams: data analysts, logistics modelers, procurement specialists, demand forecasters. The work is extraordinarily complex. That complexity is also the moat. A veteran supply chain practitioner: someone who spent twenty years running logistics operations, sourcing organizations, or distribution networks, can now deploy AI agents that monitor supplier risk, model tariff scenarios, track demand signals across markets, and generate procurement recommendations. BCG estimates agentic systems already account for 17% of total AI value in supply chain and are projected to reach 29% by 2028. Supply chain leaders report that 78% anticipate disruptions to intensify over the next two years, but only 25% feel prepared. That gap is a market. And the practitioner who combines deep operational experience with agentic infrastructure to serve mid-market manufacturers, distributors, and retailers, companies too small to staff an internal supply chain intelligence function, is sitting on a genuine Solo Scale opportunity. The client can be enterprise-scale. The operator doesn’t have to be. 5. Trades and Skilled Work Operations (The Very Surprising One) This one will raise eyebrows. Solo Scale isn’t just a knowledge economy phenomenon. It extends into the physical economy. Particularly into the operations layer that sits above skilled trades. Consider the experienced HVAC contractor, the seasoned electrical contractor, the veteran construction project manager. The actual hands-on work requires licensed, skilled tradespeople on-site. But the operations layer: estimating, scheduling, compliance documentation, supplier negotiations, customer communications, change order management, warranty tracking, is all information work. And information work is exactly what agentic AI transforms. A solo operator with deep trades experience can now run the back office, customer acquisition, and operations management for a services business that employs 10–20 tradespeople in the field, without a team of administrators, estimators, and project coordinators. The leverage point isn’t the work itself. It’s the operations intelligence layer that organizes and monetizes the work. The trades are facing a documented expertise gap as veteran operators retire. The Solo Scale model lets that expertise be preserved, leveraged, and deployed at scale. This is the physical economy version of the Vertical Intelligence Company. The Prediction: What Solo Scale Does to the Economy The internet was supposed to create the creator economy. It did, but only partially. The promise was that anyone with expertise and a laptop could build a scalable business. The reality was messier. The numbers tell the story. The creator economy was valued at roughly $250 billion in 2024. More than 300 million people worldwide identify as creators. But only 4% earn more than $100,000 annually. More than 50% earn less than $15,000 per year. The top 10% of creators received 62% of ad payments in 2025. The long tail of the creator economy doesn’t produce economic independence. It produces economic fragility. The bottleneck was always the same: the internet gave everyone distribution, but it didn’t give everyone leverage. Attention was the scarce resource. And attention, unlike expertise, doesn’t compound. It chases novelty. The creator who went viral in 2022 has to go viral again in 2023. The game never ends. Solo Scale changes the bottleneck. Expertise compounds. Client relationships deepen. Reputation concentrates in narrow domains. A Solo Scale practitioner isn’t competing for attention. They’re deploying knowledge that took decades to accumulate, through infrastructure that didn’t exist until now. Here’s the prediction: we are about to see the largest wave of high-value small business formation in American history. The data is already signaling it. Business applications in the US hit 5.62 million in 2025: up 8.2% from 2024, and nearly double the pre-pandemic annual average. In the first four months of 2026, applications are running 17.4% ahead of the same period last year. Solo-founded startups surged from 23.7% of all new ventures in 2019 to 36.3% by mid-2025: a shift that tracks almost precisely with mainstream AI tool adoption. And 47% of respondents in a 2026 Entrepreneur survey said AI availability makes them more likely to start a business. But this wave will be different from prior entrepreneurial surges in one critical way. It won’t be driven by people selling to consumers. It will be driven by domain experts, veterans of industries, functions, and markets, selling to enterprises. Small businesses at software margins, serving Fortune 500 clients. That combination has never existed at scale before. The economic implications compound. When a solo practitioner can serve enterprise clients at software margins, several things happen simultaneously. More experienced operators leave large organizations to compete against them. The addressable market for small businesses expands upmarket. Enterprise buyers gain access to specialized expertise they couldn’t previously afford. And the income distribution of entrepreneurship shifts: away from the winner-take-all dynamics of the creator economy, toward the more defensible economics of expertise. 78% of solo businesses currently make under $50,000 annually. The Solo Scale model, executed well, moves that number dramatically. Twenty percent of solopreneurs already earn between $100,000 and $300,000 annually without any employees. AI is early. The ceiling is rising fast. The long-promised democratization of the economy is happening. It’s just not coming through social media. It’s coming through agentic infrastructure. A Personal Note Lest you think I’m only researching and analyzing this trend, I’m living it. Uphoff Advisory, LLC is a Solo Scale business. In 90 days, as a solo operator, I’ve built a thriving multi-client advisory practice serving organizations that range from fast-growing entrepreneurial ventures to established B2B information brands. The client engagements are substantive. The margin structure looks nothing like a traditional consulting practice. And the AI infrastructure running underneath: research, analysis, content production, web site design and build, business development, is doing work that would have required a team of three to five people a few years ago. I’m not describing a theory. I’m describing what happened when I built exactly this. Solo Scale indeed. Why This Is Different From What Came Before The creator economy taught us that one person could become a media company. Substack. YouTube. Podcasting. The economic unit of production changed. But the economics of creator businesses were still constrained: audience attention was the bottleneck, and monetization was tied to distribution scale. Solo Scale is different. The bottleneck isn’t audience. It’s expertise. And expertise, unlike audience, is defensible. It compounds. It gets more valuable with specificity, not less. What Shopify did for merchants, eliminating the infrastructure barrier to selling, agentic AI and vibe coding are doing for expert practitioners. The infrastructure barrier to running a high-leverage services business is collapsing. The firms that built defensible practices on the old model, that hired in order to scale, that conflated headcount with capability, are not positioned for this shift. The practitioners who built deep domain knowledge and are willing to redesign the delivery model are. What to Watch The signal isn’t in the tools. Tools are proliferating everywhere, and most of the noise about AI is tool noise. The signal is in the business models that start showing up with unusual margin structures. Solo practitioners handling client loads that would have required teams. Boutique advisory firms with economics that don’t match their headcount. Intelligence products delivered by two-person operations that used to require twenty. A trades contractor running a 15-person field operation from a home office. That’s Solo Scale. And it’s only beginning. The views expressed in Uphoff on Media are entirely my own. They don’t represent the opinions of any company I’ve led, any board I’ve sat on, or any investor who’s had the pleasure of debating strategy with me over the years. If something I write here sounds brilliant, I’ll take full credit. If it turns out to be wrong, I was clearly misquoted by myself. “Uphoff on Media” is published by Tony Uphoff, Founder and Managing Partner of Uphoff Advisory, LLC [https://uphoffadvisory.com/]: a strategic advisory practice for founders, CEOs, and investors in B2B information, marketing, and technology. The businesses that drive business. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tonyuphoff.substack.com [https://tonyuphoff.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

AI Isn’t a Tool. It’s Infrastructure

A couple of weeks ago I gave a talk to the Southern California chapter of the Counselors of Real Estate. Seasoned operators. People who have financed, built, and managed commercial real estate through multiple cycles. Smart, experienced, skeptical. When I shared a slide showing that five technology companies plan to spend nearly $700 billion on AI infrastructure in 2026 alone, the room went quiet. Thanks for reading Uphoff on Media! Subscribe for free to receive new posts and support my work. Not skeptical quiet. Stunned quiet. During the Q&A session, someone said what everyone was thinking: “Those numbers can’t be real.” They are real. And understanding why they’re real, and what they mean, matters for every business leader working in B2B markets right now. We Have Seen This Before. Just Not at This Scale. The human mind has trouble processing numbers beyond a certain size. $700 billion is difficult to put into context. It doesn’t connect to lived experience the way a house price or a company budget does. So let’s anchor it. The U.S. Interstate Highway System remains the benchmark for transformative American infrastructure. Eisenhower signed it into law in 1956. It took more than 35 years to complete. Its final cost, in 1991 dollars, was $128.9 billion. That investment is credited with generating $1.80 in economic output for every dollar spent. The foundation of American commerce for half a century. The AI infrastructure build will exceed that total in a single year. The Hoover Dam, one of the engineering wonders of the 20th century, cost $49 million in 1931 dollars, roughly $811 million today. It powered the American Southwest and enabled the growth of cities like Las Vegas and Phoenix. Its construction required an act of Congress, years of planning, and thousands of workers. Amazon alone is projecting $200 billion in capital expenditure this year. That’s roughly 250 Hoover Dams. In one year. By one company. The Manhattan Project, which produced the atomic bomb and reshaped the post-war world order, cost approximately $24 billion in today’s dollars. The five largest AI infrastructure investors: Microsoft, Alphabet, Amazon, Meta, and Oracle, will spend that amount every few weeks. This Is Not a Tech Story. It’s an Infrastructure Story. The numbers become comprehensible when you stop thinking about them as technology spending and start thinking about them as infrastructure investment. Every transformative infrastructure build in history produced numbers that seemed disconnected from reality at the time. Because they were. They were civilization-scale investments, not household-scale ones. The people financing the transcontinental railroad or rural electrification weren’t spending money in the same category as a business investment. They were building the foundation an entire future economy would run on top of. That is exactly what is happening now. Goldman Sachs projects total hyperscaler capital expenditure from 2025 through 2027 will reach $1.15 trillion. More than double the $477 billion spent from 2022 through 2024. Global AI spending is projected to reach nearly $1.5 trillion this year and exceed $2 trillion in 2026. This is not a bubble narrative or a FOMO story. It is a recognition that compute, in the machine age, is what roads and rails and electrical grids were in prior eras. It is the foundation everything else gets built on. And building that foundation at civilizational scale costs civilizational money. The Nearest Comp: The Fiber Buildout of the Late 1990s There is one prior episode that maps to this moment more closely than any government project. Not the dot-com stock bubble. The actual physical buildout underneath it. Between 1996 and 2001, telecom companies spent more than $500 billion laying fiber optic cable across the United States, with global totals estimated as high as $1 trillion to $2 trillion once equipment, acquisitions, and debt-financed expansion are included. Telecom companies issued more than $500 billion in new bonds in that window alone, betting that internet traffic would grow fast enough to fill the capacity they were building. It didn’t. Not on that timeline. By 2001, an estimated 95 percent of the fiber laid during the boom was dark, built, buried, and unused. The companies that built it mostly went bankrupt. Capital markets had financed a buildout years ahead of the demand curve. Here is the part that matters most for this conversation: the infrastructure was not the mistake. The financing structure and the timeline were the mistake. And the dark fiber didn’t just eventually get used, it changed what was economically possible. Bandwidth costs collapsed so far, so fast, that companies could build business models that made no sense at 1999 prices. Netflix could not have pivoted from DVDs-by-mail to streaming in 2007 if it had been paying 1999 bandwidth rates. Amazon Web Services launched in 2006 by renting out infrastructure capacity that existed only because so much had been overbuilt. YouTube’s 2005 bet that ordinary people would upload and stream video at scale only worked because bandwidth was nearly free. None of this was the plan in 1999. It was the byproduct of a trillion-dollar overbuild that bankrupted the companies that financed it and then handed the next generation of entrepreneurs capacity at a fraction of its true cost. It would be an overstatement to say none of this happens without the fiber bust. Someone, eventually, would have built the bandwidth the modern internet needed. What the overbuild actually bought was time and price. Cloud computing, SaaS, and streaming video arrived years earlier, and at a dramatically lower cost basis, because a previous generation of investors had already paid for capacity the market wasn't ready to use. The infrastructure outlived the investors who built it. The value accrued downstream, to companies and business models that didn’t exist yet when the fiber went into the ground. That is the precedent that matters here. Not as proof the AI buildout is a bubble. Not as proof it isn’t. As a reminder that physical buildout and financial outcome are two different questions, and history answers them on different timelines. The fiber was real. The capacity was real. The mistake was assuming the company that laid the cable would be the company that profited from it. The Valuation Numbers Follow the Same Logic The IPO and valuation numbers circulating right now are equally disorienting. OpenAI closed a funding round in March 2026 at a post-money valuation of $852 billion, with a public listing targeting above $1 trillion. SpaceX has confidentially filed for an IPO at a reported target valuation of $1.75 trillion to above $2 trillion. For context: the largest IPO in history remains Saudi Aramco’s 2019 listing at a valuation of roughly $1.7 trillion. Aramco sits on the largest proven oil reserves on the planet. It was built over decades with the backing of a sovereign government. AI companies are approaching those numbers in years, not decades, with private capital. That either means the market is pricing the AI infrastructure build as one of comparable civilizational importance to the petroleum economy. Or it means the market is wrong Both are possible. History will settle it. But the framing matters: these are not tech stock valuations. They are bets on which companies will own the infrastructure layer of the next economic era. What B2B Markets Need to Understand I’ve made two arguments here on Uphoff on Media that this build puts to the test. The first: The Vertical Intelligence Company [https://tonyuphoff.substack.com/p/the-vertical-intelligence-company], not the traditional media company, is the sustainable model, one that lets business information companies scale at software margins. The second: brand is the one thing AI cannot manufacture, the case I made in Your B2B Marketing Career in the Age of Agentic AI. [https://tonyuphoff.substack.com/p/your-b2b-marketing-career-in-the] The infrastructure build makes both arguments more urgent. For B2B Marketing, this build is what makes agentic AI execution at scale possible in the first place. Content generation, campaign orchestration, and personalization down to the individual account are becoming nearly free. That should terrify anyone whose value proposition was “we can produce more content than you can.” It should not terrify anyone whose value proposition was trust, relationships, and brand. Execution is being industrialized. Judgment and brand are what’s left standing. Here is the pattern history keeps repeating: the companies that built the great infrastructure of each era rarely captured the most value from it. The railroad barons were real. So was the destruction of railroad economics once the infrastructure became commoditized. The fiber companies of the late 1990s built the network and mostly didn’t survive to use it. The electricity utilities built the grid. The consumer products companies, the manufacturers, the retailers — they built the businesses that ran on the grid. AI infrastructure is being built at a scale and speed that will make it a commodity faster than most people expect. The economics will shift to the companies that know how to use it — not the ones that own the pipes. For investors: it means the infrastructure bet is largely made. The application layer is where the next generation of value will be created. The Numbers Are Real. The Question Is What You Do With Them. The CRE professionals in that room were right to be stunned. These numbers are stunning. What they shouldn’t be is paralyzing. Eisenhower’s highway system seemed incomprehensibly expensive in 1956. It became incomprehensibly valuable. The question for every business leader in that moment was not “are these numbers real?” It was “what do I build on these roads?” That is the right question now. The infrastructure is going in. The roads are being paved. The machine age is not arriving. It has arrived. The only remaining question is what you build on it. The views expressed in Uphoff on Media are entirely my own. They don’t represent the opinions of any company I’ve led, any board I’ve sat on, or any investor who’s had the pleasure of debating strategy with me over the years. If something I write here sounds brilliant, I’ll take full credit. If it turns out to be wrong, I was clearly misquoted by myself. “Uphoff on Media” is published by Tony Uphoff, Founder and Managing Partner of Uphoff Advisory, LLC [https://uphoffadvisory.com/]: a strategic advisory practice for founders, CEOs, and investors in B2B media, marketing, and technology. The businesses that drive business. Thanks for reading Uphoff on Media! Subscribe for free to receive new posts and support my work. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tonyuphoff.substack.com [https://tonyuphoff.substack.com?utm_medium=podcast&utm_campaign=CTA_1]