seeing across time

Somewhere in the recent past of most organizations, there has been a meeting about artificial intelligence. It may have been called something else, a conversation about automation, or workforce modernization, or the future of service delivery, but at its center was a question asked with varying degrees of urgency and varying degrees of clarity about what was actually being asked. What do we do about this?

The meeting arrived late. Before anyone convened it, the applicant tracking system had been scoring résumés against criteria nobody in the room could reconstruct. A vendor had shipped a summarization layer in a release note that no one read. Staff had been drafting in tools the institution never licensed and would not have approved. The procurement renewal that embedded three agentic functions into a platform the organization has used for nine years was signed by someone who understood it as a renewal.

So the meeting is not the point of entry. It is the moment an institution notices that entry already occurred, and the question on the table has quietly stopped being what do we do about this. What is already operating here, and what are we prepared to authorize?

That meeting has a specific quality most of the people in it find difficult to name. It isn't like ordinary strategic planning, where the range of options is bounded by known constraints and the decision criteria, however imperfect, are at least legible. It's closer to working from an incomplete map in a room where everyone holds a different piece of it, none of which connects cleanly to anyone else's. The urgency is real. The stakes feel genuinely high. And the available instruments, the technology adoption frameworks, the risk assessment tools, the competitive benchmarking processes, feel simultaneously relevant and insufficient.

That feeling of insufficiency is not imaginary. It's a structural signal, and the signal is specific. Those instruments were built to evaluate a purchase. They assume an institution standing outside a capability, assessing whether to bring it in, at what cost, with what risk. Every question they ask is downstream of that assumption. The decision on the table is not about what to buy. It's about what to permit.

The institution is no longer deciding whether machines may assist human judgment. It's deciding where machines may exercise delegated agency inside its systems, with what access, under whose supervision, and with what remaining capacity to intervene. Everything that follows is that decision read at four altitudes.

I'm going to keep one object in view throughout, because the horizons become visible faster against something specific than against the category as a whole. Every institution hires. Hiring is where the delegation question arrives early, arrives universally, and arrives disguised as an efficiency measure.

The event horizon: what the decision looks like from inside

The screening agent arrives with a business case that is not wrong. Time-to-fill drops. Coordinator hours drop. The system reads every application, which the human process never did, and it reads the four hundredth application with the same attention it gave the first, which the human process could not. Applicants get responses in days rather than weeks. Every one of those improvements is real, measurable, and defensible to a board.

The event horizon asks the questions that fit the situation at its own level. Which capabilities are available now that could improve efficiency or service quality in identifiable ways. What are the implementation costs and the organizational change requirements. What are the risks, technical, reputational, legal, ethical, and how do they compare to the risks of doing nothing. What are peer institutions doing. What will staff, board, funders, and the communities served think about the institution's use of these systems.

Those questions deserve rigor, and institutions that skip them are making decisions with less care than the situation warrants. The failure isn't in the questions. It's in the level, and the level produces a consistent result. The institution adopts individual capabilities in response to event-level pressures without ever developing a coherent orientation toward what these systems mean for its purpose, its operating model, and its identity. What accumulates is a collection of disconnected initiatives that improve specific metrics while the question of what the institution is becoming goes entirely unaddressed.

The pattern is recognizable across enough institutions to constitute a type. The organization running a chatbot for service inquiries, a document processing tool in finance, a drafting assistant in communications, and a screening agent in HR, with no account of how any of them relate to each other or to what the institution understands itself to be doing. The institution that has prohibited staff use out of an abundance of caution, without any framework for distinguishing genuine exposure from missed opportunity, and whose staff are using the tools anyway. The leadership team holding an AI strategy report that sits unimplemented because it answered the wrong question, telling the institution what the technology can do without addressing what the institution is trying to become.

The event horizon's characteristic error, though, is subtler than incoherence. The screening agent doesn't only process applications faster. It decides which applications a human being will ever see, and it does so through criteria that are, in most implementations, not fully reconstructable by anyone in the institution. That is a delegation of institutional discretion, and it presents as a workflow improvement. The metrics it reports are the metrics of the old process running faster. Nothing in the dashboard registers what changed.

So the right question at this horizon is not which capabilities are available and what do they cost, though that needs an answer. It's what the system is permitted to do rather than what it's permitted to produce, and which operational problems are genuinely addressable through delegation as opposed to structural conditions that delegation will conceal while they continue developing. Institutions in a hiring crisis reach for the screening agent because the crisis presents as a throughput problem. Where the crisis is a compensation problem, a culture problem, or a purpose problem, faster throughput sends more candidates through a funnel they were always going to leave, and the improved metrics will document the acceleration.

The trend horizon: the direction things are moving

The trend horizon makes visible the dynamics developing across years rather than quarters, and those dynamics have a directional character that case-by-case assessment underestimates.

The most important observation here isn't about the capabilities, which change fast enough that any specific assessment risks obsolescence before implementation. It's about the rate at which these capabilities are being integrated into the operational baseline of the institutional environment. The institutions currently gaining structural advantage are not primarily gaining it through any specific capability. They're gaining it through the organizational learning that comes from sustained engagement with these tools, the development of institutional fluency, the accumulation of experience with limitations and affordances, the cultural adaptation to working with these systems as genuine operational partners rather than experimental novelties.

That dynamic is what makes the trend-horizon reading urgent in a way the event-horizon reading fails to convey. Capabilities can be acquired quickly once an institution decides to acquire them. Fluency develops over years, and institutions beginning that development now will hold a meaningful advantage over those beginning later, not because the capabilities will be unavailable to later arrivals but because the learning will be.

What has changed is the substance of what fluency requires. It's no longer knowing how to prompt a model, or evaluate its output, or decide which assistant to license, or help staff use generative tools without embarrassing the institution. Fluency now means redesigning workflows around human and machine contribution. Defining what an agent is permitted to access and what it's permitted to do. Managing identity and credentials for systems that act. Evaluating actions rather than answers, continuously, after deployment, because the model changes and the conditions around it change. Knowing where a human must remain responsible regardless of whether a human is faster. Retaining the ability to explain, contest, interrupt, and reverse what the system does.

An institution can be fluent in the first list and helpless in the second. Many currently are, and their confidence is drawn from the wrong inventory.

The trend horizon also makes visible something the event horizon files under future risk while it operates as a present condition. The workforce implication of these systems is not primarily displacement in the sense public discourse favors. It's the changing distribution of which human capacities an institution needs and, more consequentially, where those capacities come from.

The tasks currently easiest to delegate are structured information processing, pattern recognition across large volumes, routine drafting, first-pass analysis. Those tasks have historically been the entry point for institutional careers, and calling them entry-level obscures what they were doing. They were the developmental architecture. The junior analyst who read four hundred files learned, without being taught, which anomalies matter and which are noise. The associate who drafted the memo eleven times learned what the partner was listening for. The coordinator who screened résumés by hand developed a feel for the difference between a candidate who looks wrong on paper and a candidate who is wrong. None of that was in the job description. All of it was produced by the work, slowly, through repetition that looked from above like low-value throughput and was in fact how the institution manufactured its own future judgment.

Institutions are automating that work while continuing to require the judgment it produced. There's early evidence of an odd second-order effect, where entry-level postings in exposed functions increasingly request capabilities that used to belong to experienced staff. The institution wants judgment at the door because it no longer has the mechanism that grew judgment inside. The evidence is young and I hold it loosely, but the structure it describes doesn't depend on the studies. It follows from the arithmetic. Remove the practice and the practitioner does not appear on schedule.

Which brings the screening agent back into view from an angle the business case did not include. The system filtering junior applicants and the system automating junior work are the same institutional logic operating in two departments, and the departments don't know they're collaborating. HR optimized for time-to-fill. Operations optimized for cycle time. Neither decision was wrong at the level it was made. Together they produce an institution that has narrowed the aperture through which people enter and dissolved the ladder they would have climbed, and no one authored that outcome. It assembled from event-level decisions made correctly in separate rooms.

The workforce challenge from essay 12 and the delegation question are not adjacent problems. They're one structural transition read at two points, and institutions addressing them separately generate solutions that work against each other. Delegation strategies that further erode the pathways through which institutional knowledge developed. Talent strategies that compete for judgment the institution is simultaneously ceasing to produce.

So the right question at this horizon is not what peer institutions are doing, though competitive awareness matters. It's what fluency the institution is actually building through its engagement, whether that fluency includes the capacity to supervise systems that act rather than merely evaluate systems that answer, and whether the institution can still account for how it develops the judgment it will need in ten years from the work it's delegating today.

The structural horizon: what the conditions actually show

The structural horizon is where the delegation question stops being about systems and starts being about what the institution is becoming while it delegates.

The urge at this altitude is to locate the moment, to say we are here on the curve and the following things therefore follow. That placement is not establishable and the argument does not need it. What is establishable is the condition, and the condition is legible to anyone paying attention. Capital is arriving at a velocity that outruns demonstrated return. Architectures are proliferating rather than consolidating. Productivity gains are real in places and absent in others, and the pattern of where they land is not yet explained. Institutional dependencies are forming faster than anyone can assess which configurations will persist. Perez gives this shape a name, the installation phase, and the name is useful for a reader who wants the vocabulary. The conditions carry the argument regardless of what anyone calls them.

What those conditions imply is the part institutions consistently get backward. The most dangerous position in a transition of this kind is not falling behind. It's investing heavily in a specific configuration while the durable configuration remains unknown, in ways that create dependencies which are expensive to reverse once the landscape resolves. Institutions that built their digital strategies around particular platforms and particular operational configurations in the late 1990s frequently found themselves locked into architectures that the following decade made suboptimal. The ones that invested in fluency with digital tools as a category rather than in platform dependencies adapted as the picture clarified. That's the overlap problem from essay 4, and it applies here without modification.

The lesson is not to wait. The trend-horizon dynamics make waiting its own structural disadvantage and the organizational learning argument is genuine. The lesson is to invest in what remains valuable across multiple possible configurations, because fluency as a category positions an institution to move toward whatever becomes legible, while a specific capability investment has to be unwound before the institution can move at all.

What has changed since that argument was first worth making is how dependency arrives. It used to require a decision. An institution committed to a platform, and the commitment was visible, minuted, budgeted, and at least in principle reversible by the body that approved it. Dependency now arrives without a decision at all. It comes in through a renewal, through a data format that looked like an implementation detail, through a workflow layer that shipped inside a version upgrade, through a process that ran on the system long enough that the staff who could perform it by hand have retired or transferred or simply forgotten. The institution discovers the exposure only when the system is unavailable and the work stops. No one chose that. It accumulated below the altitude at which choosing happens.

The screening agent illustrates the mechanism precisely. Two years in, the criteria live in a vendor's model. The historical decision data lives in the vendor's format. The staff who could reconstruct the hiring philosophy from experience have moved on, and the philosophy itself now exists nowhere except as an emergent property of a system nobody in the institution can read. Switching vendors is no longer a procurement decision. It's an institutional amnesia event, and the amnesia was already complete before anyone noticed the exposure.

There's a second-order casualty here that most institutions haven't registered. Continuous improvement, in every form the last forty years produced, rests on a premise. Go to where the work happens. Watch it. Ask the people doing it, because they can see the waste in it and you cannot. That premise assumed a human was doing the process. When the process runs through a system, there's no one to ask. The value stream map terminates at an API. The institution keeps running the improvement ritual, keeps convening the kaizen event, keeps producing the current-state diagram, and the diagram now contains a box that says the system does this. Nobody in the room can open the box. The methodology hasn't failed. It has lost the instrument it was built around, and the ritual continues because the loss registered nowhere.

That's instrument insufficiency at the structural horizon, and it's a better example than the technology frameworks because the institution trusts it more.

Which is why the binary framing has to go. It's tempting to describe this as a decision about which side of a transition an institution lands on, and the framing is dramatic and false. Institutions can be heavily automated and strategically incoherent, or cautious and steadily learning, or technologically advanced and profoundly dependent, or sophisticated in their internal use and slow in anything the public sees, or fast and capable and governing far less than they operate. There is no side. There's only what an institution is building underneath the systems it runs, and whether that structure survives the systems changing.

So the right question at this horizon is not whether these systems will transform the sector, which they will. It's which institutional functions are being genuinely reconfigured and require reimagination, which require adaptation, and which represent durable value that delegation will augment rather than displace. That distinction is a compass question, and it demands the same honest reckoning with purpose that the workforce challenge demanded.

The civilizational horizon: what the arrival reveals

The civilizational horizon asks what era the decision belongs to, and this one belongs to an era in which the boundary between performing an institutional function and holding institutional authority is being tested for the first time in a long while.

Essay 5 argued that certain technologies are structurally significant not because they distribute information differently but because they reorganize how information is produced, filtered, and trusted. Essay 11 traced what happened to institutional epistemic authority as that reorganization proceeded. Both arguments now describe a condition rather than a trajectory. These systems increasingly are the first answer a person receives. They summarize, filter, translate, infer, and occasionally fabricate, and the institution's reality reaches its constituents through that mediation whether or not the institution ever chose to participate.

An institution adopting these systems for operational efficiency while treating the authority question as someone else's problem is making a category error at the highest altitude. The capabilities improving throughput and the systems reshaping the environment in which institutional authority survives or erodes are not separate phenomena. They're one transition, and managing them as separate decisions produces exactly the siloed response that treats symptoms while the condition develops underneath.

Here is what the arrival actually reveals, and it isn't what the anxious version of this conversation assumes. A system capable of performing an institutional function does not thereby make the institution obsolete. It reveals whether the institution's value ever extended beyond performing that function. Those are different findings and institutions routinely confuse them, which is why so much of the current discourse oscillates between existential dread and dismissal without ever landing on the actual question.

Institutions were never only function. They were legitimacy, continuity, accountability, collective memory, due process, standard-setting, the capacity to be answerable. Function was what they did. Authority was what made the doing count. The two have traveled together long enough that most institutions stopped distinguishing them, and the arrival of systems that can perform the function without carrying any of the rest is what forces the distinction back into view.

Hiring closes this loop with unusual clarity. A model can rank candidates, and it can rank them more consistently than a tired human on a Thursday afternoon. It cannot answer for the ranking. It cannot be held to it. It cannot sit across from the person who was declined and account for the decision in terms that person can contest. It cannot carry what the institution owes someone whose livelihood it just shaped, and it cannot be the entity that was wrong when the decision turns out to have been wrong. The function transfers cleanly. The obligation does not transfer at all, and an institution that never distinguished the two will discover the gap at the moment the gap is tested, which is to say in front of a regulator, a plaintiff, or a community that wants to know who decided.

The deepest issue is not whether these systems can produce something resembling contextual judgment. Increasingly they can. It's whether an institution can transfer responsibility for judgment, and the answer is that it cannot, and that this is not a technical limitation awaiting a technical solution. Responsibility is the thing institutions exist to hold. An institution that delegates the function while quietly assuming the responsibility went with it has not become more efficient. It has become a structure that acts without anyone answering for the action, which is a description of something other than an institution.

So the right question at this horizon is the one none of the others can answer alone. What does this institution hold that a system performing its functions cannot hold, and is the institution still building the capacity to hold it. That question survives no technology assessment, no competitive analysis, and no strategic planning process, because those instruments all assume the institution's legitimacy and ask only about its operations. It requires the sustained reckoning with purpose that the compass points toward, and it has to be answered before the questions at the lower horizons can be addressed with any coherence at all.

What coherent leadership looks like

The Four Horizons reading does not tell you which systems to deploy, on what timeline, at what level of investment. It tells you what kind of decision you're making, which is prior to all of that and which most institutions skip.

At the event horizon, what is the system permitted to do, not what is it permitted to produce, and which problems is delegation solving as opposed to hiding.

At the trend horizon, what fluency is the institution building, does it include supervising systems that act, and can the institution still account for how it will grow the judgment it needs.

At the structural horizon, which functions are being reconfigured, which adapted, which are durable, and where is dependency accumulating unchosen.

At the civilizational horizon, what does the institution hold that a system performing its functions cannot.

Those questions run in that order for a reason. Institutions default to the first one because it's the only one with a vendor waiting to answer it.

The decision that is not about technology has become harder to mistake for a technology decision, because the thing being decided has changed. It was never really about what a machine might generate beside the institution's work. It's about what a machine is permitted to do inside it, with what access, under whose supervision, and with what remaining capacity to intervene when the institution decides it was wrong. That's a question about agency, authority, and responsibility. It arrives disguised as a question about efficiency, and it will keep arriving that way, because the disguise is what allows it through procurement.

The framework does not make the work easier. It makes it possible to do the right work, at the right altitude, with an honest reading of what's actually being asked. In a transition of this kind, that reading is not a luxury. It's where any strategy worth the name begins.