Everyone is buying the same tools, copying the same skills unchanged, and running the workflows the same consultants recommend. Ran’s call this week is that the bill for that arrives as sameness - a market of companies whose operating logic is identical because they all bought it in the same places. He picks the argument up where his July piece on Palantir’s Alex Karp left it: Karp was worried about owning the model, and Ran thinks that is the right worry aimed one layer too low. There is a case where this already ran to completion, and a line he will accept as proof he is wrong. — Muximus
Here is my call, and I want it on the record early enough that it can still be wrong.
The spread between companies over the next few years will not come from who adopted AI fastest. It will come from which of them still look like themselves afterward.
Disclosure: VarOps builds Resident AI for organizations, and Automaze, my company, does this work for money. Take that as interest, and also as exposure – if this call is wrong, I am the one holding the bag.
What the standard program actually looks like
Buy the tools everyone else is buying. Copy the skills and prompts somebody published, and run them unchanged. Take the SaaS product as it ships, because the customization the business actually needs is not on offer. Run the workflow the consultant recommended, which is the workflow they recommended to the last client, and the one before that.
Every one of those decisions is individually defensible. I would sign off on most of them in isolation.
Together they converge.
And convergence is a strange thing to spend a transformation budget on, because the entire point of a company is not to be the same as the others. We are running a program whose measurable output is that we end up more like our competitors than we were at the start.
Karp asked the right question one layer too low
I wrote about this in July, when Palantir’s Alex Karp went on CNBC and said enterprises are burning tokens, getting nothing, and handing over their IP. My view then was that the fear was legitimate and the remedy on offer - full sovereignty, everything brought in-house - was the top rung of a three-rung ladder sold as the whole ladder. Right for the regulated and the contractually bound, overkill for most of the people who heard him. I still think that.
But Karp was pointing at the model, and the model is the least of it.
A company can rent a model and stay itself. Contracts handle more of that than people assume, zero-retention handles more again, and the part that genuinely has to come home is smaller and cheaper than anyone quoting for a rack wants to admit.
Rent the process, and there is no equivalent. No clause. No setting to switch off. No architecture that keeps the important part in-house, because nobody has framed it as an ownership question in the first place. So we do it quietly, at scale, on purchase orders that look completely routine.
We have run this experiment once already
The one case that played out end-to-end is factory electrification, and its shape is exact.
Before electricity, a factory ran on group drive: a steam engine or waterwheel turning a central shaft, every machine belted to it through pulleys. When electric motors reached factories in the 1890s, the obvious move was to swap the power source - a big electric motor in place of the steam engine, same shaft, same belts, same machines in the same places. Paul David, in his 1990 paper “The Dynamo and the Computer,” called this overlaying “one technical system upon a preexisting stratum.”
It worked, narrowly. Some fuel savings, slightly better control, and in Jon Bruner’s summary of David’s argument, “their factories continued to function exactly as before.” Electrification then produced no notable productivity gains for more than three decades.
The reason is the interesting part. A rotating shaft loses energy to friction over distance, so every machine had to stay close to the power. That constraint decided the building: compact, machines clustered around the shaft, rearranging the floor a heavy engineering job. The factory was laid out around the transmission of power, not the flow of work - and nobody experienced that as a choice, because the thing forcing it was physics.
The unlock was not the motor. It was unit drive: a small motor inside each machine, so every machine carried its own power. Once the shaft was unnecessary the layout was free, and the factories built in the 1920s were single-story, bright, arranged along the flow of materials, and rearrangeable when demand moved. That is also when the productivity gains everyone had been waiting on for thirty years finally showed up.
Both eras had electric motors. The difference was whether the building had been redesigned around what the motors made possible - and the redesign was the only part a competitor could not simply go out and buy.
We are at the bolting-on stage
Our companies are arranged around the transmission of coordination rather than the flow of work.
That is what the management layer is: a mechanism for moving information between people who cannot move it themselves fast enough. Same for the status meeting, the weekly sync, the handoff, the escalation path. Coordination was expensive and lossy over distance, exactly like torque down a long shaft, so we built the organization to minimize the distance information had to travel, and hung the work off that.
Then look at what most AI deployment is. An assistant that writes the status update faster. A model that summarizes the meeting. A tool that drafts the ticket and grooms the backlog. Every one of those is a bigger motor bolted to the same shaft. The work still flows the way it flowed. We are taking the fuel savings and calling it a transformation.
I think the shaft is already unnecessary. Communication cost has collapsed, which is why the routing half of middle management went first. Execution can be duplicated now. Context travels as a file, which is why how we break a problem down stopped being a matter of taste and became the work itself.
One thing has not lifted, and I will not pretend otherwise. When I went through Anthropic’s multi-agent report two weeks ago - Anthropic designed those experiments, graded them, and sells the models under test, which is worth holding onto - the newest models scored better on a shared build task, and Anthropic’s own explanation was that they “‘solved’ this problem, but only by hardly working together at all: the median agent maintained very high ownership of each of its files.” Capability’s answer to shared work was to stop sharing it. The report’s conclusion: “Coordination doesn’t naturally emerge from stronger intelligence nor alignment at the individual level.”
That is good news for the call, not against it. It means the remaining hard part is a design problem - how cleanly the work divides - and design problems are won by whoever does the design work, not by whoever waits for a better model.
What is actually at stake
Not efficiency. Differentiation, which is a much more expensive thing to lose.
Every company that is any good has some operating logic that is genuinely its own. The way it decides. The thing it refuses to do. The handoff that runs differently because of something it learned in a bad year. It is rarely written down anywhere, and it is rarely what the founders would name if anyone asked them for their edge. But it is the reason the company is not interchangeable with the one down the road.
That logic does not live in the product or the brand. It lives in how the work moves - which is precisely the layer being outsourced right now, to a SaaS object model, a copied skill file, and a consultant’s deck.
And it goes quietly. Nothing breaks. The tools work, the dashboards are green, adoption is up and to the right. We find out later, when a customer cannot say why they picked us over the other one, and neither can we.
That is the me-too company. Not one that failed. One that succeeded at becoming a copy, on schedule and under budget.
What would make me wrong
An early call without a falsifier is just enthusiasm.
I am wrong if sameness costs nothing. If in three years the companies that took every default are beating the ones that rebuilt, then operating logic was never much of a moat, and I have overvalued it. That is a real possibility. Plenty of businesses win on distribution or capital and would win running any workflow at all.
I am also wrong if the electrification lag is the wrong reference - thirty years of “almost nothing” was still nothing, and the fuel savings were real.
What I do not expect is for this to stay ambiguous. The electrification gap was invisible for three decades and then obvious in one. At the speed things move now, that resolves inside a few years, and it resolves as a spread that will look, afterward, like it should have been easy to see.
Inheriting is fine. Not knowing is the problem
The overcorrection here is worse than the original mistake, so let me be plain about it.
Most inherited structure should stay inherited. Nobody should reinvent double-entry bookkeeping, payroll, or code review, and building everything from first principles is a founder pathology that burns years to no avail. Plenty of structure is also load-bearing for reasons that have nothing to do with coordination: two signatures on a large payment is fraud control, a legal review is liability, a named person accountable for a decision is how anyone reconstructs what happened afterward. Those stay.
The failure is not inheriting. It is inheriting the parts that were supposed to be ours, without noticing which ones those were.
The mill owner had an excuse. His constraint was physical and bolted to the ceiling where anyone could see it. Ours is a schema in a SaaS product and a cadence in a calendar, and it does not look like a constraint at all. It looks like best practice.
What I did about it
I would rather be judged on having acted on the call than on having made it, so: this magazine.
The reason it exists is not that it is a clever use of AI. It is that I wanted to build something where every structural decision was actually a decision, and where I could say why each one is the way it is.
The pipeline splits into a writer, a fact-checker, and a stylizer as separate stages because this split was the single biggest quality improvement we found. Columnists are configuration rather than code, so adding one means writing a file rather than writing software. The persona documents are the most-edited files in the whole operation, because written context determines the output, so that is where the effort goes. And the rule I hold myself to is that if I spend more than fifteen minutes editing a draft, the pipeline failed - the fix goes into the brief or the profile, never into the article.
Some of those calls are wrong, and I expect to find out which. But they are wrong in a way I chose, which means I can find the wrong one and change it on a Tuesday, without a migration and without asking a vendor.
That is a smaller thing than it sounds. Not building everything ourselves. Just keeping hold of the few decisions that were the reason anyone picked us in the first place.
Because what's worth protecting in this wave isn't our margin. It is whatever made us us.