The Gold Rush Was Real

That sentence aged well because it was not about models. It was about the interval between signal and action. It was about how modern enterprises have learned to substitute explanation for movement. How they can run the report, see the deviation, agree it matters, and still fail to convert that agreement into a change on the floor, in the schedule, in the maintenance plan, or in the promise made to a customer. The trap was never that AI would not work. The trap was that it would work in the one place executives already feel safe. In the insight layer. The trap was that it would accelerate the production of options while the enterprise stayed slow at choosing and committing. The trap was that more “answers” would increase meeting load, not reduce it, because every answer creates a new argument about priority, ownership, and permission. What would have to be true for this outcome to keep repeating. It would have to be true that the limiting factor is not intelligence. It is authority. It would have to be true that most organizations do not fail because they cannot see. They fail because they cannot act without supervision. It would have to be true that every manual approval step exists because permission has not been made explicit, scoped, and trusted. The post pointed at this without using the word that makes people nervous. Authority. It used “capacity and capabilities to act on the insights.” That phrasing was the tell. It was a recognition that insight does not create value. Only a change in what the enterprise does creates value, and changes only happen when someone has the right to touch the system. Time is no longer on our side. What you got right, and why it still matters The strongest claim in the original was the line that many people read and then immediately underestimate. Insights will become table stakes, but they will still need prioritization. That is the honest part. It admits that the future does not remove managerial work. It changes the kind of managerial work that matters. In 2023, most leaders still believed the advantage would come from having better models or better prompts. That belief was reasonable at the time because the technology was visibly improving and the enterprise was visibly behind. But the document argued that even perfect insight does not decide. It does not schedule. It does not release the work order. It does not authorize a deviation disposition. It does not change the supplier signal. It does not approve the overtime. It does not make the promise to the customer. It only produces another candidate for action. The last two years have largely validated that line. The more organizations deploy AI for analysis and summarization, the more they discover that their real bottleneck sits in the governance path that follows. Not governance as a virtue. Governance as the permanent

architecture of delay. The organization can be proud of its controls and still be slow. It can be slow and still call the slowness “diligence.” The post also named something most people avoid because it feels uncomfortable to say in public. Knowledge loss. That was not a generic lament. It was an operational constraint. When the person who knows the line retires, the line does not retire. When the planner who knows how to read the noise leaves, the demand signal does not become cleaner. When the quality engineer who knows which problems are real goes away, the defect data does not become less political. Knowledge loss is not just lost information. It is lost judgment about what information matters and what it means. That recognition has only gotten more urgent. Generative AI makes it easy to generate plausible explanations. That makes judgment more valuable, not less. When the enterprise loses judgment and gains explanation engines, it becomes more confident and less correct. That is the kind of trap that does not feel like failure until the market punishes it. The document also insisted on contextualized data and “ground truth” for end users. That is a mature instinct. It implies that data without provenance is noise. It implies that an answer without traceability becomes political. It also implies that if the person on the floor cannot trust what the system says, the system becomes another dashboard, and dashboards are where good intentions go to die. Where the post overreached, and why that matters now Two lines in the original read like certainty, and the market has punished certainty. One is the claim that being second equates to being irrelevant. The other is the claim that Special Purpose Intelligences will offer an insurmountable edge over linear human performance. The “second equals irrelevant” line is emotionally true in some categories, and false in others. In software markets with network effects, second place can be fatal. In many industrial markets, second place can be profitable if it arrives with reliability, integration, and a credible service model. Plants buy risk reduction, not novelty. They do not reward the earliest demo. They reward the first implementation that does not break production. So the stronger version of that line is not about being first. It is about being early enough to build operating muscle before the standards, procurement rules, and internal politics harden. The penalty is not irrelevance. The penalty is that you become a buyer of packaged capability rather than a designer of your own operating advantage. You still “adopt.” You just adopt on someone else’s terms. That distinction matters because it explains why so many firms will use AI and still not gain an edge from it. The “insurmountable edge” line also needs tightening. Intelligence alone is not an edge in physical operations. Constraint is an edge. Permission is an edge. Safety is an edge. Integration is an edge. In the physical world, time is not only compute. Time is travel. Time is changeover. Time is mechanical response. Time is thermal inertia. Time is human attention. Time is the audit trail you need after something goes wrong.

The post was right that models would become able to reason better, see better, and coordinate tools. But “exponential speed” is only valuable when the system is allowed to move at that speed. In most enterprises, the slowest component is not inference. It is approval. It is the human gate that exists because nobody can defend letting the system touch the real levers. That is why the most important missing phrase in the original is permission architecture. The document gestures at “safe access interfaces,” but it does not name the core issue. Who is allowed to do what, to which systems, under what conditions, with what evidence, and with what logging. Until that is explicit, the best AI in the world is trapped inside suggestion mode. That is not a technology failure. It is an authority failure. Insights do not create value. Only committed action does. The hidden mechanism you implied, but did not name If we revise the post with what we now know, the mechanism becomes clearer and more testable. The advantage does not come from generating more insights. The advantage comes from compressing the time between signal and authorized action, without increasing incident rate. That sentence is falsifiable. You can measure the time from detection to disposition on quality issues. You can measure how long it takes to approve a schedule change when demand deviates. You can measure how many human touches are required to approve a maintenance deferral or a parameter adjustment. You can measure the rework caused by decisions made without local context. You can measure how often a decision is reversed because nobody trusted the evidence. Generative AI makes every one of those measurements more relevant because it increases the number of signals and the number of plausible actions. If the organization cannot decide, it gets slower as it gets “smarter.” That is the paradox many executives are now living. More options. More meetings. More alignment rituals. More slide decks. More pilots. The same cycle time to act. This is where your later teaching on agency becomes the upgrade. If it cannot shape an outcome, it is not an agent. That definition turns the conversation away from marketing and toward architecture. It also forces an uncomfortable admission. Most “agents” being sold into operations are not agents. They are copilots. They produce text and suggestions. They do not hold authority. They do not own a bounded decision right. They do not carry an audit trail that survives a serious incident. The revised post, written today, would say it plainly. Agents do not create value. Authority does. The technology only matters insofar as it can be trusted with bounded authority. If it cannot shape an outcome, it is not an agent.

Two questions that diagnose the real bottleneck When an AI system produces a recommendation that would change production, who can authorize it. Not in principle, not in a policy document. In the record. In the system. Can that authorization be executed without a meeting. Can it be executed without a chain of emails. Can it be executed without the decision being re-litigated by people who were not in the room when the evidence was generated. If the answer is no, then the bottleneck is not data, and it is not the model. The bottleneck is that the organization cannot encode trust. It still relies on proximity, seniority, and social proof to move work. That is why “insights” pile up and “action” stalls. When something goes wrong after an AI-influenced action, can you reconstruct why the action occurred. Can you show the evidence that was used. Can you show what permissions were invoked. Can you show who approved what, and what the system touched. Can you do it fast enough that legal, safety, and the board accept it as a controlled system, not a reckless experiment. If the answer is no, then the organization will keep every AI system in read only mode, regardless of how advanced it becomes. That is not caution. That is self-preservation. A counterexample worth taking seriously There is a path that partially breaks your original thesis. Some organizations can buy their way into competence without being early. They can adopt proven solutions once the market hardens. They can ride vendor consolidation. They can use a standard platform, accept generic workflows, and still improve cost and service. This counterexample matters because it is real, and it is attractive to boards. It reduces perceived risk. It avoids building internal capability. It keeps accountability legible. It also allows leaders to claim progress because the procurement artifact exists and the deployment plan looks professional. The problem is what that path cannot buy. It cannot buy an organization’s ability to decide at speed in its own context. Packaged solutions can improve baseline performance, but they rarely give a firm a durable edge because competitors buy the same package. The edge shows up when the enterprise can run its own loops. When it can convert local evidence into authorized action without turning every deviation into a governance ceremony. So your original “second equals irrelevant” line was too absolute, but it had a kernel that remains sharp. Late adopters can still improve. They just tend to improve in a way that is shared by everyone else. The revised claim, tightened to what the market proved

The most accurate update to your 2023 document is not a new prediction about model capability. It is a sharper statement about where the contest moved. Generative AI did, in fact, lower the cost of producing insight and explanation. It did, in fact, make it easier to query data, summarize issues, generate options, and scale support functions. It did, in fact, increase the velocity of analysis. It did not, by itself, compress the time between signal and action in most enterprises. That compression requires permission. It requires evidence discipline. It requires bounded authority. It requires audit trails that stand up after failure. It requires interfaces that let end users see not only the answer, but the provenance and the scope of what will happen if they accept it. Your original list of what organizations need to do. Better data gathering. Systems that contextualize data. Assistants that provide ground truth. Safe access interfaces. Automation and reasoned action. That list remains directionally right. The upgrade is to reorder it around the thing that creates compounding advantage. The new order is not about tools. It is about decision rights. If you can encode authority with evidence and enforcement, the technology becomes useful. If you cannot, the technology becomes louder. That is why the post reads less like a forecast now and more like a diagnosis that was missing a label. The label is authority. The edge is not intelligence. The edge is trusted authority at system speed. A prediction that risks embarrassment By the end of 2027, boards in regulated and safety critical industrial businesses will treat “permission architecture plus auditability” as a gating requirement for any AI system that can touch production outcomes. Firms that cannot produce a clear authority map and an incidentready audit trail within fifteen minutes will confine AI to summarization and recommendations, and they will call that boundary “responsible AI” even when it is really an inability to trust their own system. If that does not happen, then the enterprise has found a way to grant machine-driven authority without needing explainability, traceability, or enforceable permission. That would be a genuine break from how industrial accountability has worked for decades, and it would reorder the risk calculus across manufacturing, energy, and logistics. Why this matters. Not as advice, as consequence The post framed a gold rush. Two years later, the gold is not rare. The scarce resource is not access to intelligence. It is the ability to convert intelligence into action without creating chaos.

Organizations that cannot encode trust will become meetings. They will become committees. They will become alignment rituals. They will become professional explainers of why they cannot move. They will adopt every new tool and still be slow where it counts. Organizations that can encode trust will compress the time between signal and action. They will build a record of decisions that survives failure. They will reduce the number of people required to make routine decisions. They will keep humans where judgment is needed, not where permissions were never clarified. Your 2023 post was a warning that the competition would determine who survives and who fades. The revision is harsher and more specific. Survival will belong to the enterprises that can grant bounded authority with audit grade evidence. Everyone else will still “use AI.” They will just use it as a mirror. The better question now is not who will adopt AI. The better question is who can afford to keep authority implicit. Time is still not on our side. References This review draws directly on Michael Carroll’s LinkedIn post from two years ago and the attached two-page document, “Who will get Left Behind in the Generative AI Gold Rush,” because it captured early that insight would become common while action would remain scarce, and because it named knowledge loss and prioritization as the true constraints rather than model capability. It also draws on Carroll’s later operating doctrine that definitions become architecture, that an agent must be able to shape an outcome, and that authority is what allows work to move without supervision, because those ideas tighten what the original implied but did not label. For external ballast on why this bottleneck persists, it leans on Herbert Simon’s work on bounded rationality, Douglass North’s work on institutions and transaction costs, Norbert Wiener’s cybernetics as control rather than reporting, and Judea Pearl’s ladder of causation as the line between association and intervention. It grounds the governance and accountability implications in the NIST AI Risk Management Framework and the EU AI Act timeline, and it uses recent empirical and journalistic reporting from the Stanford AI Index, NBER field studies on measured workplace outcomes, and Gartner and Reuters coverage of project discontinuation and “agent” hype to validate that adoption is rising while durable advantage depends on permission, auditability, and the time between signal and authorized action.