Second Place Is a Slow Death

“adopting AI.” They are embedding it into how they operate, and the moat they are building is one you will not see until it is too late to close the gap.

The scene in the title image reflecting the above illistration is not about whether the chart is clever. It is clever. The scene is about what it tempts senior operators to believe. It tempts you to believe you have time.

It tempts you to believe this is a consumer adoption story that will take years, that your plant, your supply chain, your quality system, your customer commitments, and your margins can wait for the world to catch up. It tempts you to treat AI like a tool category. That is how second place becomes last place. The piece you are reading is not a defense of AI. It is not a sales pitch for agents. It is an accusation of a widely held assumption that keeps repeating inside enterprises that are otherwise competent and serious. If the world is still mostly gray, why are some firms already pulling away. What is the mechanism that turns an early lead into a gap you cannot buy back later. What would have to be true for this outcome to keep repeating. That chart is not about AI Start with what is observable, then admit what is inference. The observation is that interactive AI use is still rare at the scale of humanity. Even the green strip, 1.3 billion, is a minority. The paid sliver is microscopic. The coding corner is almost invisible. The inference many executives draw is that “we are early,” therefore competitive pressure is still distant. They tell themselves the enterprise can let this play out in the consumer world while they do pilots, training, and cautious governance work. They tell themselves it is prudent to wait for standards, for better models, for a clearer legal environment, for the first public failures to teach the market what not to do. That inference feels responsible because it uses the language of responsibility. It also mistakes the arena. Enterprise competition is not a referendum of humanity. It is a knife fight inside a peer set. If you share customers with a competitor, you do not need eight billion people to adopt anything for your margins to get repriced. You need one rival to convert software into authority inside the seams where your value leaks. This is why the chart is not about AI. The chart is about the mismatch between how people think competitive advantage forms and how it actually forms. Competitive advantage does not require mass adoption. It requires differential capability in the places that convert time into money. One more observation tightens the point. Stanford’s AI Index reports that 78 percent of survey respondents said their organizations used AI in 2024, up from 55 percent in 2023, and that reported generative AI use in at least one business function more than doubled from 33 percent

to 71 percent. Those are big numbers, and they should end the fantasy that AI is still a fringe activity inside firms. Now the counterweight. OECD data, using a different lens and different coverage, reports that in 2025, 20.2 percent of firms across OECD countries with available data reported using AI, up from 14.2 percent in 2024 and 8.7 percent in 2023. That is fast growth, but it also underlines how uneven adoption and measurement remain, and how much “use” depends on what you count. Both can be true. AI can be widespread as a tool and rare as a decisive operating property. The chart shows rarity at the human scale. Stanford shows acceleration at the enterprise survey scale. The OECD shows uneven penetration across the broader economy. This is where most leaders stop. They talk about adoption. Adoption is consumption. It is a license, a pilot, a usage statistic, a vendor roadmap, a training program. Agency is something else. Agents don’t create value. Authority does. Permission is where profit leaks Most enterprises do not lack intelligence. They lack permission throughput. They do not have a shortage of ideas. They have a shortage of decisions that survive contact with the organization and arrive as authorized action where work occurs. They do not have a shortage of dashboards. They have a shortage of legitimate control. When legitimacy is weak, the enterprise pays a tax in time. The tax is not called a tax. It is called governance, alignment, diligence, and risk control. Each word sounds responsible. Each step can be defended in isolation. The bill arrives when defensible steps become permanent architecture and time becomes the most expensive input you do not measure. That bill is visible in the places operators hate to talk about because it is politically expensive. The meeting that exists because the data cannot be trusted. The approval chain that exists because decision rights are vague. The extra report that exists because nobody believes the last report. The rework loop that exists because accountability is unclear. The status theater that exists because people fear being blamed more than they fear being wrong. AI lands on top of that environment like gasoline on damp wood. It does not automatically ignite value. It does something more revealing. It exposes that the enterprise is not optimized for decision and action. It is optimized for safety, career protection, and blame distribution.

That is why so many firms report “AI use” without reporting material gains. The tools are present, but authority has not moved. The enterprise still routes decisions through the same human intermediation layers, only now with faster drafts and prettier summaries. The core mistake is to treat “agent” as a synonym for “automation.” An agent, if the word is going to mean anything, must be able to shape an outcome. If it cannot, it is a tool. If it can, then it is operating under authority, and authority must be explicit, bounded, and auditable. That is not a semantic argument. Definitions create architecture. Architecture decides whether value compounds or collapses. An enterprise that keeps authority implicit cannot safely let software act. It will either refuse to act, trapping AI at the edge, or it will act in ways that produce fear and backlash, triggering bans, committees, and centralized choke points. Either way, permission stays slow. The enterprises that pull away are the ones that treat authority as an engineering surface. They do not ask, “How do we deploy AI.” They ask, “Where are we willing to let authority move, and what would make that safe.” That is permission architecture. It is not a policy PDF that sits in a folder. It is not an ethics statement on a website. It is machine legible boundaries on who can do what, with what evidence, with what constraints, with what rollback, and with what audit trail. This is the seam where value leaks. Not because people are lazy. Because the enterprise is a coordination machine, and coordination consumes most of the day. When permission stays vague, coordination grows like mold. Every new system adds another place where truth can be contested. Every contested truth creates another meeting. Every meeting lengthens decision latency. Every delay turns operating variance into financial variance. A recommendation without authority is a memo. Agents do not create value The popular story says AI will win because it is smart. That story flatters people who like intelligence as a status marker. It also misses how enterprises really run. Enterprises win when they can convert signal into action faster than competitors, without losing control. That conversion is not a model capability. It is an organizational capability.

The first part is signal. Sensors, transactions, customer feedback, defects, delays, inventory, downtime, service calls, forecast error, scrap. Enterprises already have more signal than they can metabolize. Many are drowning in it. The second part is decision. Who decides, how fast, under what constraints, with what evidence, with what accountability. The third part is permission. The decision must become authorized action that can occur without convening a tribunal. The fourth part is learning. The action must feed back into better future decisions, or the enterprise stays stuck in superstition and ritual. That chain is where agency lives. The firms that treat AI as an assistant often stop at the first link. They ask AI to summarize. They ask it to draft. They ask it to search and explain. That can save time. It rarely changes the enterprise’s competitive physics. The firms that treat AI as governed authority work on the second and third links. They encode decision rights and constraints. They create a ledger that can answer, after the fact and in the moment, what the system recommended, who authorized it, what evidence was used, what action occurred, and what happened next. This is what Travis means by embedding AI into how you operate. It is not a user interface project. It is not a chat rollout. It is the conversion of authority from tribal memory into auditable structure. Now bring in the ugly truth that makes “no second place” real. The moment one firm in a category does this at scale, the basis of competition changes. Decision cycles compress. Permission moves down. Coordination overhead drops. Learning rates rise. That advantage accumulates. The economic language for this is increasing returns and path dependence. W. Brian Arthur’s work on increasing returns describes how early leads can lock in through self reinforcing feedback loops, not because the winner is morally better, but because the system rewards early advantage with compounding benefits. Enterprise agency behaves like that because authority and learning are complements. A firm that can safely delegate authority can close loops faster. Closing loops faster produces better policies, better instrumentation, and more trust. More trust allows further delegation. Delegation reduces coordination cost. Reduced coordination cost frees capacity. Freed capacity funds more redesign. The loop repeats. Second place does not lose because it bought the wrong model. Second place loses because it tried to buy the loop instead of building it.

There is a second economic piece that makes the trap worse. Productivity gains from general purpose technologies often follow a J curve. Firms invest in complements like process redesign, training, data plumbing, and organizational change before measured productivity rises. Brynjolfsson, Rock, and Syverson describe this pattern and why early gains can look small, then rise as intangible complements mature. This is where executive impatience becomes a strategic error. The laggard watches the leader absorb friction in the early complement building phase and tells itself it is wise to wait. Then the curve turns, and the leader begins harvesting returns that are no longer available to a fast follower because the leader has already rewired the operating system. No second place does not mean you can never catch up. It means catching up is no longer a procurement exercise. It becomes a multi year reconstruction of how your enterprise grants permission, records decisions, and closes loops. Human agency is the asset at risk Here is the point most “enterprise agent” narratives evade because it forces moral seriousness. Human agency is not a cosmetic value. It is the reason an enterprise exists as a human institution and not as an automated factory of obedience. Human agency is the capacity to act with intent toward an outcome under constraint, to choose among options, to take responsibility for consequences, and to learn. Albert Bandura describes core features of human agency through intentionality and forethought, self regulation, and self reflectiveness. This is not motivational language. It is a definition of what makes humans more than executors of procedure. Translate that into the enterprise. Human agency is what allows a plant manager to decide when a line should stop, even when the schedule says keep running. Human agency is what allows a quality leader to call a containment, knowing the quarter will take a hit, because the customer matters more than the optics. Human agency is what allows a CFO to say no to a project that flatters growth stories but bleeds cash. Human agency is what allows a COO to trade a short term metric for long term control. The question AI forces is not “can software do tasks.” Software can do tasks. The question is whether the enterprise uses software to expand human agency or to collapse it. A system expands human agency when it returns time, focus, and decision quality to humans, removing coordination drag and giving people better options with clear constraints. It collapses human agency when it turns humans into liability absorbers who rubber stamp outputs they do not understand, or when it strips discretion while keeping accountability. The first produces resilience. The second produces brittle compliance. This is why your definition matters. If you call a tool an agent, you will eventually give it agency like authority without admitting you did. If you give it authority without explicit constraints and

auditability, you will trigger the oldest reaction in institutional life. Fear. Then you will build more procedure. Then you will slow down. Then you will declare that AI failed. The only stable path is to treat authority as explicit, bounded, and revocable, and to preserve human ownership of intent. The enterprise can delegate actions, but it must not delegate accountability for values. There is also a justice angle that senior leaders should name without theatrics. The capability approach, as summarized in the Stanford Encyclopedia of Philosophy, treats freedom as substantive opportunity, not merely formal permission. That distinction matters in firms. If AI becomes a layer that only a minority can understand and steer, then the enterprise creates a two tier labor system. A small group holds intent and authority. A large group executes without comprehension. That is not a culture problem. It is an agency problem. The enterprise that wins will be the one that increases human agency for the many, not the one that builds a priesthood of prompt power. It will use AI to compress permission latency while keeping intent and accountability legible. If it cannot shape an outcome, it is not an agent. The moat you will not see Travis says the moat will be invisible until it is too late. That line lands because it matches what operators already know about competitive gaps. The gaps that matter do not announce themselves. They show up as a customer that stops returning calls. A quote you used to win that you now lose. A lead time you can no longer defend. A service backlog that becomes permanent. A quality escape that should have been prevented. The moat is not the model. It is the movement of authority. Authority becomes a moat when it is encoded into workflows and backed by auditability. That is what allows permission to move down without panic. That is what allows decisions to be executed without meetings. That is what allows learning to be captured without politics. This is the part many executives will resist because it exposes a painful truth about why the enterprise is slow. The slow part is not the people. The slow part is the architecture. The enterprise has built a system where permission is negotiated socially because it cannot be granted structurally. AI does not solve that by being smarter. It solves it by forcing the enterprise to make decision rights explicit, to make constraints executable, and to make audit trails automatic. If you do that work, AI becomes a control surface. If you do not, AI becomes another dashboard.

The chart in Travis’s post is the consumer view. The enterprise view is different. The enterprise question is not who has chatted with a bot. The enterprise question is where authority has moved from meetings into software. This is where the adoption story becomes a repricing story. When one firm can close decisions faster without breaking trust, it starts buying the future with speed. It takes more learning turns per month. It reduces the cost of coordination. It increases the reliability of execution. It begins to offer the market a kind of certainty competitors cannot match. Second place can still look fine for a while. Second place can still have strong people, good products, and respectable financials. Then the market notices the compounding. That is the moment when second place becomes a slow death, because the only way back is to rebuild your permission architecture under pressure. Counterevidence that deserves respect A fair objection appears immediately under Travis’s post. One commenter notes that LLMs are not the only AI. People use AI through Google Maps, Uber, Netflix, and other systems without ever having a “first AI conversation.” The chart, in that view, undercounts AI exposure and overstates the gray zone. That objection is correct as far as it goes. Many people receive AI mediated outputs without using a chatbot. It is also irrelevant to the enterprise agency argument. The chart is not measuring whether an algorithm touched someone’s life. It is measuring the spread of interactive AI tools as a conscious activity and a paid behavior. That matters because conscious usage is a proxy for skill accumulation. Enterprises do not compete on whether their customers watched a recommended movie. They compete on whether their own people can use software to compress decision and permission latency. Even if you assume the chart understates AI exposure, the core signal stands. Paid usage is tiny. Coding scaffold usage is tiny. The distribution of advanced interaction is narrow. That narrowness is the reason the window exists. It is also the reason the window will close. Once firms build operating advantage from authority movement, the late majority does not get to copy it by downloading an app. It has to reconstruct the institution. Another counterweight is even harsher. Recent work from NBER authors using a representative international survey of almost 6,000 executives reports muted realized gains and modest expectations, with many firms reporting no productivity impact so far and only small expected output boosts over the next few years from AI adoption. That is the productivity paradox showing up again, and it is a warning against hype.

That evidence does not refute the agency thesis. It strengthens it. If the median firm is using AI for small, peripheral tasks, then measured gains will be small. If AI does not touch authority, then AI does not touch the conversion of time into money. The paradox is not proof that AI is weak. It is proof that most firms are stuck at the tool layer. The prediction no board wants to hear Here is a prediction that would be embarrassing if wrong. By the end of 2027, at least one large publicly traded industrial enterprise will disclose a material incident where an AI driven automated action caused a customer, safety, or compliance failure, and the postmortem will trace the root cause to missing permission boundaries and missing audit logs. That incident will not be caused by “bad AI.” It will be caused by an enterprise that treated the word agent as a marketing label, let authority move without making it explicit, and then discovered that legitimacy cannot be improvised after the bill arrives. The fastest way to falsify this prediction is to build the missing architecture at scale before the market forces it. The slowest way is to wait for others to learn the lesson in public. Two diagnostic paragraphs, for leaders who want the truth cleanly, belong in the mouth of a COO and a CFO, and they do not require any new tooling to ask. When a recommendation appears in your business, who has the right to act on it, and does that right exist as an explicit boundary the organization can audit later. Can you name the constraints that define safe action in that workflow, and are those constraints executable, or are they enforced through meetings and escalation. If a bad outcome occurs, can you trace accountability without a political fight. Can you show what evidence the decision relied on without reconstructing a narrative after the fact. If you cannot, you do not have an AI problem. You have an authority problem. When the enterprise is slow, where does the time go. Does it go into physical reality, like curing, shipping, maintenance, and real lead times that cannot be compressed. Or does it go into permission, alignment, and internal consent rituals that exist because the enterprise cannot trust its own systems. When your best operators say “we spend all day coordinating,” are they describing necessary coordination, or coordination that substitutes for legitimacy. If a competitor could remove one third of that coordination burden without breaking control, would your cost curve, service level, and learning rate stay competitive. If you cannot answer that, you are not early. You are exposed. The ending is not advice. It is consequence. Travis’s chart is the right kind of reality check because it tells the truth about scale. Most of the world is not in this yet.

The enterprise consequence is that this does not buy you time. It narrows your margin for error. In the early phase of a platform era, tools are cheap and capability is scarce. In the later phase, tools are still cheap and the scarce asset becomes the operating model that can convert them into authority without losing control. That is why there is no second place in enterprise agency. Not because second place cannot buy models. Second place can buy models. Second place cannot buy back the months and years it spent negotiating permission while a competitor moved authority into auditable structure, closed loops, and began compounding. The chart will fade from the feed. The repricing will not. When permission stays vague, politics becomes the control system. References This piece draws on Travis Johnson’s February 2026 LinkedIn post and the associated adoption graphic that quantifies interactive AI use by segment and scale, including the 6.8 billion gray zone and the small paid and coding slivers, because it frames the bubble problem cleanly and forces an honest discussion of who is actually using what. It uses Michael Carroll’s internal operating doctrine on authored longform, provenance discipline, authority, and the definitional bar for agents, because the argument hinges on definitions becoming architecture and on authority as the unit that compounds value. It triangulates enterprise level adoption claims using Stanford HAI’s AI Index 2025 reporting sharp year over year increases in reported organizational AI use and in reported generative AI deployment, alongside OECD firm adoption data showing rapid but uneven uptake across economies, because measurement variance itself is part of the governance problem. It grounds the “tools do not equal returns” tension in Brynjolfsson, Rock, and Syverson’s Productivity J curve work on general purpose technologies and intangible complements, and it explains lock in dynamics through W. Brian Arthur’s increasing returns and path dependence, because that is the mechanism behind “no second place.” It defines human agency using Bandura’s agentic perspective and cross checks the moral and institutional stakes using the capability approach summary, because enterprise agency that collapses human agency produces brittle compliance, not durable performance. It validates the “we are early” internet comparison using ILO reported global internet usage in 2005, because the historical analogy is only useful if the baseline is real. It pressure tests the hype with recent NBER survey based evidence that many firms report limited realized productivity impacts and modest expectations from AI so far, because the strongest version of this argument must survive the paradox rather than talk around it.