AI: Endgame 2 — The Future: AI Summary

This is an AI-generated summary of AI: Endgame 2 — The Future. The full paper is at https://www.lloydwtaylor.com/endgame/the-future/

Most business conversations about AI are framed as questions of adoption: how fast, how much, which uses, which risks. That frame treats AI as a new tool, a familiar problem with well-worn methods.

The essay says the frame is wrong. In competitive knowledge work, what is happening is not adoption. It is a structural change in the productive unit. The difference matters. Adopting a tool leaves the firm intact and adds to what it can do. Changing the productive unit changes what the firm is.

The Compound Entity

The productive unit is the thing whose output a firm exists to generate, measure, and manage. For most of modern business history, that was the individual knowledge worker. Everything else, from tools to capital, served that person.

That was the official story. Real work has always been richer and harder to see. It runs through relationships, informal judgment, workarounds, and the person who knows who to call when something is stuck. The individual was the unit management was organized around. It was not always the unit through which value was actually produced.

The first essay in the series described integration: an outside disturbance absorbed into a system’s machinery until it stops being outside. In the firm, the essay argues, integration means a new productive unit. It calls this the compound entity: human thinking and AI output working together to produce results neither could reach alone. Over time, the practitioner’s own way of working reorganizes around the compound.

The compound is defined by a threshold, not by its parts. Below the threshold, a practitioner uses AI as an instrument, the way a surgeon uses a scalpel or a lawyer uses a database. Above it, the firm’s training, expectations, and benchmarks have been rebuilt around what the compound produces. A practitioner who works without the AI is then at a structural disadvantage that grows over time.

Heavy use is not the test, and neither is capability. Someone can rely heavily on AI and still be the productive unit. That holds as long as the work and its standards still assume the person could produce the output alone. The compound begins where removing the AI would not just slow things down. It would take apart the system the institution has built around the work.

The clearest early sign is how training is designed. Below the threshold, training builds independent practitioners who later learn AI tools. Above it, training assumes the compound from day one. A redesign like that admits something institutions rarely say out loud. The compound, not the individual, is now the productive unit.

The threshold is becoming visible in parts of diagnostic radiology. Signs are appearing in high-volume contract review, discovery-phase litigation, and parts of financial risk modeling, though accountability there still attaches to human practitioners. The threshold is approaching in parts of drug discovery, scientific research, and management consulting. What these fields share is not how much AI they use. It is that work has started to depend on output that no individual is expected to reproduce alone.

Here is what firms have not yet absorbed. If the compound is now the productive unit, then job descriptions, reviews, training, pay, accountability, and succession plans were all built for a different kind of entity. The firm is pointing those instruments at something they were never designed to see.

The Accountability Gap

The firm and the professions that staff it built their accountability systems around the individual. Licensing certifies a person’s competence. Malpractice liability falls on a practitioner. Fiduciary duty, the legal duty to act in someone else’s best interest, is held by named people. Negligence requires a human who made, or failed to make, a decision. When something goes wrong, the system looks for the person whose judgment failed.

The compound entity has no address in that system.

The gap is not theoretical. It is producing disputes right now in medicine, law, and financial services. Suppose a compound produces a wrong diagnosis or a flawed contract. Who is responsible? The essay plays out the exchange. The practitioner says the output looked right on review and was wrong in ways review couldn’t catch. The provider says its system worked within its specifications. The firm says the practitioner holds the license. And the practitioner replies: responsible for what, exactly? I approved the output of a process I can’t audit, through reasoning I can’t inspect, trained on data whose origins I don’t know.

That isn’t bad faith, the essay says. It accurately describes the compound. The human and AI contributions can’t be cleanly separated. The output is compound, but accountability demands a separable unit.

The gap is being closed now, but not by a plan worked out in advance. It is closing case by case, through enforcement actions, professional standards, and contract terms. How it closes will decide who carries power and liability in the compound economy. One possible path is broader organizational responsibility for AI-enabled work. That would resemble enterprise liability, which places responsibility on the business as a whole, and it would be a major shift. What no assignment of blame can do is restore the old structure. The compound exists and it produces. The only question is who answers for it.

The Cognitive Supply Chain

Every firm running compound entities depends on infrastructure it doesn’t control. The AI runs on computing power owned by a few providers, using models built by even fewer organizations. Firms reach it through software they usually can’t audit or easily replace.

Firms have always had suppliers. The difference now is that the dependency reaches into the productive unit itself. Below the threshold, switching infrastructure means switching a tool. Above it, switching means replacing part of the entity the firm’s work is organized around.

AI infrastructure is concentrated in three layers: cloud computing, advanced chips, and frontier model development. In cloud computing alone, the biggest providers hold a large majority of global market share. That concentration formed within a decade, faster than regulation could respond.

The essay compares this to the early twentieth century, when factories stopped running their own steam engines and plugged into the electrical grid. It looked like an efficiency gain, and in many ways it was. Over time, though, firms became dependent on infrastructure they did not control, could not inspect, and could not easily copy.

AI dependency carries higher switching costs than electricity, because the infrastructure is built into the compound’s way of thinking and working. For deeply integrated firms, changing providers can mean rebuilding much of the workflow, testing, governance, and practical experience built around the compound. The productivity gained from integration is held hostage to the relationship, which gives the provider leverage over the firm’s basic productive architecture. A firm can replace its law firm or its logistics partner. It can’t easily replace the cognitive infrastructure its productive units run on.

So a compound-era firm’s position depends on two things. One is what its compound entities can produce. The other is the terms of its relationship with whoever supplies their infrastructure. That second relationship has no standard model yet.

The Plasticity Split

Once the productive unit changes, firms differ in how costly the change is to absorb. The essay borrows a term from the first paper: phenotypic plasticity. It means the ability to reorganize quickly, without falling apart, when conditions change faster than ordinary adaptation can keep up. Plasticity is not a separate cause of the shift. It decides how much of the old organization must be torn down before the firm can work around the new unit.

Highly plastic firms share several features that look inefficient by the standards of the old optimum. Decision authority sits close to the work: with the practitioner, the team, the engagement. Leaders specify outcomes, not procedures, since the procedure that worked before may now be exactly wrong. Redundancy is kept on purpose, with overlapping skills and generalists working alongside specialists. And some internal inconsistency is tolerated, because consistency assumes that what worked before will keep working.

Some professional partnerships show this structure before bureaucracy sets in. So do platforms that route human judgment through distributed networks, and research organizations, where the productive unit has always been a person plus an instrument.

The firms facing the highest cost are those most tightly bound up with regulation, credentials, legacy systems, and formal accountability. Think of large financial institutions with multi-year approval cycles, or health systems whose credentialing assumes the practitioner is the unit. Think of law firms built on partnership and precedent, or schools whose accreditation was designed for training independent practitioners. These are not unprepared organizations. They are “exquisitely functional organizations, built for conditions that are reorganizing around them.”

The biology offers no prescription. The organism that survives fast change is the one most able to reorganize, not the one best fitted to the old world. No fix is available that doesn’t begin by dismantling the optimization. Firms that built their edge through decades of refining processes now face that same structure.

The Management Problem

Since the scientific management movement of the early twentieth century, management theory has rested on one assumption. The productive unit can be observed. You can watch people work, audit processes, and trace results to decisions. All performance management depends on that.

The compound has a component that is only partly observable. Firms can see an AI system’s inputs, outputs, and logs, but usually can’t reconstruct how it reached a specific result. So when the compound succeeds or fails, no one can say which part did what. Management demands separable credit and blame. The essay says this is not a data problem that better dashboards will fix. It is architectural.

Four management functions break. Performance evaluation breaks first. It over-credits the person when the AI does most of the work. It under-credits the person when the real skill is running the compound well, a demanding skill that output measures can’t see. Capability development breaks next. Training someone for the compound requires knowing what the human contributes, which requires seeing inside the compound. Failure diagnosis breaks the same way, since fixing an error means locating it in a process that can only be partly reconstructed. Succession breaks last and least visibly. When an experienced compound practitioner leaves, their trained relationship with specific AI systems and their workflow habits leave too. None of it was ever written down where a successor could read it.

A management theory that starts from the compound as the productive unit has not yet come together; most writing on AI in the firm is about tool adoption. The essay compares the moment to the shift from craft to industrial production, which required a whole new management theory. For now, firms apply the old theory to a new entity. It produces readings, but they don’t measure what actually decides competitive outcomes.

The firm can’t see this from where it manages. The essay stresses that this is structural, not a failure of intelligence. Firms have always struggled to see how work really gets done, and AI adds technical opacity on top of that old social opacity. Worse, the instruments the firm uses to assess itself increasingly run through AI-integrated processes. This is the turn the section rests on: the observer is not watching the change from outside. It runs on the same foundation the change is reorganizing.

If this account holds, some firms will rebuild accountability, management, training, and power analysis around the compound. They will gain a structural advantage over firms still using frameworks built for individuals. That advantage will come not from solving a strategic puzzle, but from correctly identifying what kind of entity the firm had already become. The difficulty is that firms must recognize the change using instruments built for the entity being displaced.

The essay also sets a boundary on its own claim. None of this applies just because people use AI, even heavily. The argument begins only where an institution has rebuilt work, training, and expectations around output the practitioner is no longer expected to produce alone.

The essay’s closing:

Where the compound threshold has been crossed, the firm’s management instruments remain calibrated for an entity that is no longer the productive unit. What they are pointing at now — compound, integrated, dependent, unobservable in its internal operation — is something those instruments were not designed to see.