One Move, Three Faces

Four people looked hard at the same problem this year and came back with four different names for it.

A futurist who advises a hundred CEOs called it pilot purgatory — companies drowning in demos that never ship. A consulting firm surveyed nearly a thousand companies and called it an organizational failure, not a technological one. Two exponential-org theorists called it the Organizational Singularity and said the company itself is about to be rewritten. And one founder didn't name it at all — he just cut forty percent of his company in a single day and rebuilt what was left around agents.

They were not describing four problems. They were describing one, from four sides. And the reason each saw only one side is the reason most AI programs are failing: the thing they were all looking at does not come apart into pieces. It only looks like it does. We inherited a separate vocabulary for each face of it.

This essay is about the shape underneath all four. Once you see it, the failures stop looking like bad luck or immature technology and start looking like what they are: a single design mistake, made over and over, by people each holding one third of the answer.

Four vantage points, one object

Start with the futurist. Amy Webb — speaking to Fortune — talks to between a hundred and a hundred fifty CEOs a year, and what she reports is a workforce buried in output it can't use: presentations that once took a week now take a day, while the same team receives five times as many. Executives tell her they're stuck running an endless series of pilots — “enormous productivity but they're not sure what to do with that.” Her sharpest observation lands on the people steering, not the tools: no CEO was hired because they were an expert in artificial intelligence. They were hired because they were excellent executives — and AI, she says, is nearly the worst possible technology for a generalist to grapple with, because it isn't one technology, it's an umbrella for many, and planning it requires data, not gut.

That is a description of the human face of the problem. Judgment is the constraint — specifically, judgment that has not yet climbed from doing the work to governing the systems that do it. Webb feels the org problem behind it, but she has no frame for it, so she names it after the axis she can see and calls the rest a coming reckoning.

Now the consultants. Bain surveyed 951 companies and found the gap in hard numbers.

Bain's own subhead states their finding plainly: the fix is organizational, not technological. The technology worked; the value didn't arrive. It is the same finding four research organizations reached independently — the account I gave in The Problem Was Never the Technology, whose mechanism is the knowledge distance between what a tool can do and what an organization can absorb. Their prescription is worth holding onto: pay down your workflow debt before you deploy, because the single most costly mistake in AI is automating a broken process. AI doesn't fix workflow debt; it locks it in, speeds it up, and makes it far more expensive to unwind. The question to ask before any AI program, they write, is not “where can we apply AI” but “if we were designing this process from scratch today, what would it look like?”

That is the organization face — authority, structure, the operating model. And notice: Bain's data also surfaces the third face without a frame for it. Their number-one barrier isn't budget or skills or buy-in; it's that companies cannot reliably reach their own data. That is a technology-and-ownership problem showing up in an organizational survey. They diagnose it in org language because that's the vocabulary they brought.

Then the theorists. Salim Ismail — who co-authored the Exponential Organizations framework — and Peter Diamandis published the far edge of the same argument. Their provocation is a single question Ismail puts to every CEO he advises: is there a high-margin line of business two people with AI agents could replicate in sixty to ninety days? Seventy percent say yes. Their method for responding is precise and, notably, borrowed from engineering: rebuild two high-volume workflows with AI, run them beside the old system, and once the new version wins, retire the original — repeating until the AI-native operation has replaced the legacy core. What survives the rewrite, they say, is the legal entity, the data, the learning systems, the purpose, and human judgment. What disappears is the static org chart, the five-year plan, and most of middle management.

That is the organization face again — but seen as a destination and a migration method rather than a present failure. It is, deliberately, the optimist's version of what Bain measured and Webb feared.

Three observers, three axes: Webb sees the human, Bain sees the organization and half of the technology, Ismail sees the destination. Each is standing on one strand and describing the whole rope by the one strand in their hand.

The one who did all three at once

Which brings us to the founder. In February 2026, Jack Dorsey cut Block from over ten thousand people to under six thousand — roughly forty percent — in a single day, from a position of financial strength, and tied the move directly to AI. But the cut is the least interesting part. What matters is the sequence around it.

He built the capability first: Block's internal agent had been in production for more than a year, saving engineers eight to ten hours a week, with the large majority of recent code AI-authored by the time of the announcement. He co-wrote a thesis with board member Roelof Botha arguing that the org chart itself was obsolete: hierarchy is a two-thousand-year-old information-routing protocol, and the machine can now carry it. Then he redesigned the roles around that claim — individual contributors, directly responsible individuals, player-coaches — dissolving the reporting-and-coordination layer of middle management into the systems that now hold alignment. And he treated the whole thing as an experiment to be debugged, expecting trial and error as he learned what worked.

This was neither a layoff dressed as a strategy nor a strategy that skipped the work. It is the technology, the organization, and the human re-formed as a single move. Dorsey is the outlier the other three were implicitly asking for — the person who saw all three faces because he happened to hold three kinds of latitude at once: the systems-architect cognition to see the object, the founder's authority to act on it, and the financial strength to do it before he was forced to.

That is exactly why he is rare, and why the lesson is not “find a better CEO.” The honest reading of Webb's observation is not that executives are dim about technology. It's that organizational design quietly changed categories. It used to be a leadership-and-people discipline. In the agent era it is becoming an architecture discipline, and those select for different cognition. Almost no one holds all three in one head. Dorsey does. The rest of us need another answer.

Why it is one object, and why now

Here is the claim the argument turns on. The human, the organization, and the technology are not three problems that happen to interact. They are three faces of one design decision — and in the agent era, that decision has a name and a location.

Every agent operates at a seam: the line between what it may propose and what the organization will permit. I have been calling this line the seam for a while, and building out its mechanism one axis at a time. An agent drafts a refund, recommends a price, routes a case, writes to a ledger. Something has to decide whether that proposal becomes an action. Look closely at that single line and you will see it is three things at once, inseparably:

It is a human judgment — someone decided what “good” looks like here, and what must escalate. It is an organizational authority — the firm granted this class of agent this much scope and no more. And it is a technological enforcement — something in the system actually holds the line at the moment of the write, or it doesn't. Move any one of the three and the other two move with it. Widen the human's tolerance and you have changed what the org permits and what the technology must enforce. Harden the enforcement and you have changed the authority and the judgment it expresses. They are one object seen from three sides.

This is the part that is new to the agent era, and it explains why the old playbook fails. When the org chart was a human hierarchy, you could treat people, structure, and tools as separate projects, because the coordination between them was itself done by people — the fuzziness absorbed the seams. Agents collapse that. When a chunk of the organization is actually software sensing, deciding, and writing to systems of record, then architecting the company and architecting the system stop being two activities. The technical framework stops being a metaphor for the org and becomes its literal design language.

That is why the frameworks that fit this moment read like engineering. Backcasting — defining the end state and working backward — did not come from management theory; it came from energy-systems planning — John Robinson coined the term in 1982 — and became a formal method in sustainability planning well before anyone applied it to a company. The strangler-fig migration Ismail and Diamandis prescribe is a software-architecture pattern from 2004. Neither was invented for org design. Both were imported, because in the agent era the organization has become the kind of thing those methods were built for: a system with interfaces, migrations, and a live cutover.

Three strands, held under tension

If the three faces are one object, then the failure mode is not doing any one of them badly. It is doing them in sequence, as three separate projects, handed off between three separate owners. A strategy firm redesigns the org and stops at the seam. A build shop wires the technology and never reaches the judgment. A change-management program prepares the people for tools that were never governed. Each hands its strand to the next, and the tension is lost at every seam.

A fifth observer named the mechanism the other four only gestured at. Kevin Buehler spent thirty-two years at McKinsey advising the largest financial institutions in the world; he now works inside an AI-native firm rebuilding those workflows from within. Writing in Barron’s, he put the failure in a sentence any operator will recognize: speed up one machine on a factory line and you do not get a faster factory — you get a new bottleneck somewhere else on the line.

His example is a deal. The confidential memorandum was the obvious thing to automate, and a model now drafts it in a fraction of the old time. But the deal does not close weeks sooner. The memo is only as good as the financials underneath it, and those are still assembled by hand. Fix the financials and the constraint moves again, down to the buyer list.

The chain never ran at the speed of its fastest link. It ran at the speed of its slowest one, and automating a single step just relocates where slowest lives.

That is the value leak, measured from inside the work: not one step done badly, but a sequence worked one strand at a time, so the constraint slides down the line faster than any single gain can catch it. That is precisely where the ninety-five-percent-return-nothing number comes from. When two-thirds of the C-suite cannot say who owns the decision, that is the seam going ungoverned, measured. Every failure the four observers described is what a single-axis move looks like from the outside, named after whichever strand its author could see.

The alternative is not “do all three.” That still sounds additive — three things to remember to include. It is one move: three strands held together under tension, like a rope. A rope is not three strands laid end to end; it is three strands twisted so the load runs through all of them at once. Pull on it and no single strand bears the weight alone. That is the design principle, and it is why it cannot be handed off in sequence. The moment you separate the strands to work them one at a time, you have a bundle of string, not a rope.

Automate the coordination; protect the judgment. The technical framework tells you how to do the first. It exists to clear the space for the second, which it can never do for you.

Notice that the optimists and the skeptic agree on this dividing line, from opposite directions. Ismail's list of what survives the rewrite — data, learning systems, purpose, human judgment — is the same boundary Webb draws when she warns about offloaded thinking and the vanishing sense of ownership over the work. One draws it from the opportunity side, one from the risk side. They agree on where the line is. They disagree only on how gracefully you cross it.

What this asks of whoever runs it

If the move is one object with three faces, then the question “who can actually run it” has a structural answer, not a heroic one. Dorsey ran it because he held all three faces in one head — and that is rare enough that building a strategy around finding his equivalent is not a strategy. The replicable answer is a team whose members each carry one face with real credential, and who hold the three together under tension rather than passing them down a line.

That is the honest response to Webb's “CEOs don't have the skill set.” It was never a criticism; it was an observation about category — the one I made in Who Holds the Line, where the AI-leadership role turns out to be assembled from four older disciplines and no single one produces the whole. No single person is supposed to hold judgment, authority, and enforcement at equal depth — the roles that produce each of those were selected for different things.

Buehler’s own prescription points the same way from the finance side: the redesign has to be driven by senior people fluent enough in the technology to know what it can and cannot do — evaluated on that fluency, not exempt from it — because they are the ones with the authority to redraw the work. That is the human strand and the organizational strand refusing to separate. A better generalist is the wrong answer; the right one is a triad, in which the human strand, the organizational strand, and the technical strand are each owned by someone who has actually done that work, and the three are woven, not sequenced.

Which is the same test this essay has been circling. The rope is not proven by naming its three strands. It is proven by a load that runs through all three at once and holds — a single piece of work in which the judgment call, the authority grant, and the technical gate were decided together, in the same room, as one decision argued from three sides. The composition tells you the move can be run. Only the delivered work tells you it was.

This essay is the whole move, seen from above. Each strand has its own mechanism, told in depth: the human — judgment climbing from operator to governor; the organization — where authority is drawn and redrawn; the technology — the gate that holds at the write. They meet at the seam. For the failure this one prevents, see The Problem Was Never the Technology; for the two ways out, Escaping the Innovator's AI Dilemma; for what only the people can supply, The How Was Never the Point.