The Back of the Watch

You cannot convert an analog watch into a digital one. Not by swapping a gear, not by adding a chip to the movement, not one part at a time. The two keep the same time and share not a single component in how they keep it. One is a machine of gears and springs; the other is a different machine entirely. To get from the first to the second you do not retrofit. You design the second and build it.

Most organizations are about to attempt the retrofit anyway — because it looks responsible, and because the alternative looks like starting over. Something is arriving at the front door, and the cheap way to handle it will pass every review right up until it stops the clock.

The thing arriving is the agent. Not the chatbot a company deploys on its own site — the agent acting for the customer, sent to transact on their behalf: to compare, to apply, to buy, to dispute. The industry has started to call the discipline for handling it Know Your Agent, the way it has long said Know Your Customer. It is real enough now to have shipped standards, vendors, and regulators circling it. The principle is easy to state: autonomy without identity is ungovernable, and autonomy must not dilute accountability.

So far this reads as an identity problem, and identity problems get solved with features. Detect the agent. Flag it. Issue it a credential. Add a field. That work is necessary and it is coming, and none of it is what this essay is about.

Because underneath the identity question sits a larger one that no feature touches. Every organization built its customer journey on a single assumption it never wrote down: that a human is on the other side. Authentication assumes a person with a phone. Disclosures are written to be read by a person who can be bored or confused. The service queue routes to a resolution shaped for a person. Pricing, rate limits, the escalation path, the fraud model, the tempo of the whole thing — all of it presumes a human intent and a human pace. That presumption is encoded in a thousand places and stated in none.

Know Your Agent, done as a bolt-on, adds one gear that says this one is an agent and hands it to a movement whose every other gear still assumes a person. The system now sees the agent, and then does the human thing to it anyway, faster. You have not governed the agent. You have detected it and mis-served it at speed.

That is the trap — the retrofit the opening just named, arriving at the front door in practice. An economist has spent thirty years measuring exactly this gap, between what a technology can do and what an organization actually gets from it.

The gap, named

A piece in Forbes by Gallup’s Vibhas Ratanjee put the argument in the open in August 2026, built on thirty years of Erik Brynjolfsson’s work. Readers here have met Brynjolfsson before — in The Seam, where his account of the limit on any single mind explains why authority was ever spread across an org chart, and in What’s Missing, where his payroll study finds the AI employment effect first at the youngest edge of the workforce. He is useful a third time now, for a third reason.

Brynjolfsson, at the Stanford Digital Economy Lab, has spent his career measuring one gap: what a technology can do, against what an organization actually gets from it. Asked how wide that gap is today, he did not soften it. The biggest, he said, in world history. And he does not mean it as a technology problem. His life’s argument is that a powerful general-purpose technology delivers nothing on its own. It demands co-inventions — the redesigned process, the retrained people, the changed incentives. And those are real investments that show up in no budget, because they are intangible. As he puts it, they are called intangibles because they are not measured.

That is the 70/20/10 receipt in Orientation read from the far side. There, the value of a transformation sits mostly in people and process while the budget pours into technology. Brynjolfsson names the reason the budget runs backward: the tech spend wins because it is countable, and the co-invention it beats is not.

The dip that buys nothing

The unmeasured spend has a shape. Brynjolfsson and his colleagues named it the productivity J-curve. Output dips early, while the intangible rebuilding is underway and uncounted, then rises once the gains arrive. Anyone who has run a turnaround knows the J as a discipline rather than a theory: you take the hit on purpose, spend against it, and the curve turns. The dip has a shape, an owner, and a deadline.

The failure is not refusing the dip. It is entering it and skipping the rebuild it was meant to pay for. A J only bends back up if the down years bought something — the redesigned process, the moved authority, the new skill. Skip that and there is no upstroke. There is only the spend. Brynjolfsson’s own field estimate is blunt: most executives he talks with discuss the structural work; maybe a quarter to a third do it.

The dip now has a public price. MIT FutureTech scored every company in the S&P 500 on how deeply AI runs through the business. Among companies outside tech, the ones that added AI to processes built for people run profit margins two to three points below companies that never started. The few that rebuilt the work sit well above them once size and sector are accounted for, though that group is nineteen firms and the authors call it an association rather than proof. Read against the J, the shape is plain enough. The company that bolted AI on is not on a slower road to the same place. It is sitting in the dip, paying for a rebuild it never started.

Why you cannot fix one gear

The mechanism is where the essay gets its name.

Brynjolfsson reaches for the same object this essay opened on. Try to convert an analog watch to digital by swapping one gear at a time, he says, and the lesson is immediate: it is an integrated system, or it is nothing.

But conversion is not quite the failure organizations run into when they retrofit anyway, and the difference matters. The opening image names the wrong kind of machine — no gear-by-gear path across, no matter how carefully you swap. What actually stalls a retrofit already underway is narrower and meaner: seizure. Not a good gear replaced badly, but an old gear that locked in place and stopped the two around it. Same lesson, from the other direction. In a real movement the gears are meshed. Turn one and all three turn. Seize one and the other two stop, however sound they are.

The seizure has been measured. Across more than half a million developers on GitHub, commits rose 240 percent once autonomous coding agents arrived. Releases rose 30 percent. Review and shipping did not speed up, and a product ships only as fast as its slowest step. The agent turned the gear that was already turning. The one that kept the time did not move.

Those three gears are Human, Organization, and Technology. It is the argument for HOT held in one object: the three move together or the watch keeps no time.

The gear that seizes

Return to the agent at the door. The reason Know Your Agent as a bolt-on stops the clock is that the gear it adds is a T gear — detection, enforcement — dropped into a movement whose O gear never moved. The organization has not decided what an agent is permitted to be: what it may do alone, what it must hand back to a person, whose authority it acts under. Without that decision, detection produces nothing. You can now see the agent, and you have not said what it is allowed to be, so the system treats it as the thing it was always built for — a customer. And serves it as one.

That unmoved authority is the Organization axis, and its whole job is to draw the line before any agent runs. Done as redesign, that line gets drawn for a new kind of actor: the org decides the agent’s permissions, and the platform enforces the fork on every action — the Technology axis, the layer everyone overfunds and least understands, where power added to an ungoverned movement just runs the old logic faster. Agent-versus-human stops being a flag read at the door and becomes the distinction the movement is built around.

The fork is not really a technical choice. It is an authority choice wearing a technical costume.

Deciding what agents may be reassigns who owns the customer relationship when half the traffic is not human. It changes what the service organization is for. It changes whose sign-off governs a transaction no person initiated. Firms will bolt it on instead — not from ignorance, but because the redesign costs someone standing, and the widget lets every existing gear keep turning while the movement quietly stops keeping time.

The same seizure runs backward, too, in systems that have already stopped. Ratanjee tells of a manager delighted with an AI rollout — reporting automated, hours saved — who drew the line at the core process. That one needs verification, he said; that is white-glove work, we check it ourselves. His team described the step differently. The verification was a residue of how the systems had been built years earlier, when the output genuinely could not be trusted and someone senior had to sign off. The systems were replaced. The checking was not. It hardened into a role, then a standard, then a definition of quality. The manager could not see it. The team could see nothing else. The Seam explains why the check was there to begin with — authority was distributed because no one mind could hold every decision. And Knowledge Distance already named what happens next: automation does not fix a broken process, it accelerates one. The agent at the door is that same story, told before it happens.

What an agent buys in a movement like that has now been counted, too. A 2025 survey of 5,512 Korean workers found a correlation of 0.008 between the hours AI saved them and the change in what they produced. The hours were real. They went, largely, to leisure on the clock. Nobody decided that. The organization simply had nowhere for the hour to go, so the time went where the old arrangement sent it.

The most successful seize hardest

The instinct to preserve is not a flaw. In a well-run company it is the correct reflex, earned over years of being right about how the work should be done. Which is why Brynjolfsson calls it hardest for the most successful companies. The preservation budget is largest where the balance sheet looks strongest. The firms that will not survive this are not the ones that moved too slowly. They are the ones that were doing well enough, for long enough, that nobody thought to open the back of the watch.

His own evidence agrees. The Stanford lab’s Enterprise AI Playbook ran five months across forty-one organizations and fifty-one deployments that had produced real value. Asked what was hardest, seventy-seven percent named the invisible work — change management, data quality, process redesign. Not the model. Not the infrastructure. The authors went looking for what separated the wins from the failures, and the answer ran to one sentence: the difference was never the AI model. It was always the organization.

That is Orientation restated by a stranger. Two people who have never spoken, one finding.

Three moves before the next approval

None of this is fated, and the first moves cost nothing.

Those three moves are an audit. They find the seized gear and let you price a number that, right now, no one has. The return on a transformation is figured against the wrong denominator: the implementation spend, when the true cost is that spend plus the preservation spend running against it. Do the arithmetic honestly and break-even projects turn negative, but almost no organization can do the arithmetic, because the second figure has never been entered anywhere. It is intangible in Brynjolfsson’s exact sense: uncounted, and therefore uncountable until someone decides to count it. And deciding to count it means naming the interests the preservation spend protects, which is the conversation the whole arrangement was built to avoid. So the denominator stays wrong — not because the number is small, but because no ledger was ever built to hold it.

The benefit side has the mirror problem. JPMorgan values each AI project by its expected return and leaves out the four hours a week its internal assistant saves 150,000 people. In Jamie Dimon’s words, “That’s not in an NPV.” He is right to leave it out. Saved time is not a return until something stops being paid for, which is Move 2 by another name. The most careful public attempt to fix the arithmetic, Social Capital’s September 2026 deep dive on AI returns, gets further than most. It prices the finished task: every attempt, the failed runs, every hour a person spends checking the agent’s work. It still has no line for the old way kept running beside the new one. That missing line is the preservation budget.

The danger is not that companies fail to invest. It is that they spend enormously, watch nothing move, and conclude the technology fell short. What actually happened is that every dollar of transformation was matched, quietly, by a dollar nobody approved, defending a design nobody chose.

What the audit hands to

An audit finds the gear. It does not free it, and the watch is the reason. You cannot convert an integrated movement one part at a time. Finding the step is not the same as rebuilding the process so the step can move without stopping every gear around it.

That rebuild is what HOT is for, and it answers the question Brynjolfsson ends on. He concludes the difference was always the organization, then admits the research does not tell you which part holds, or why it holds so reliably in firms with every reason to move. The watch answers the shape of the question. It is a mechanism, not a part. HOT answers the rest by naming the gear. H aims, O couples, T drives, and the movement keeps time only when the three mesh under load. In a company with years of being right built into it, the gear that seizes is almost always O, because that is where the rightness hardened into standard. Almost always is not a diagnosis, which is why the audit comes first — from the outside, a stopped movement looks identical whichever gear locked.

Brynjolfsson’s own prescription is that a creative CFO and CEO will reinvent their measures to count the intangibles. That is the necessary first move, and the three above are how you make it. But measuring a seized gear does not turn it. The line only clears when the movement is rebuilt around what the organization now needs to permit — the redrawn boundary between what an agent may propose and what the organization will allow. Which is the work the down years of the J were always for.