The Dark Office
Robots do not need light, so the factory went dark. A decision engine does not need segments, so the room where retail decides is next. Not the whole office. One room first.
Ilyas Kurklu, CEO & Co-founder, Replenit
Nine people at the end of the line
In Drachten, in the north of the Netherlands, Philips builds electric shavers in a hall where 128 robots do the work. Nine people work that line. All nine stand at the end of it, in quality assurance, looking at what came out.
Near Mount Fuji, FANUC has run its robot plant with the lights off since 2001. Robots build robots there, about fifty a day, and the line can run for a month without anyone walking in. Gary Zywiol, a FANUC vice president, put it the way only an engineer would: “Not only is it lights-out, we turn off the air conditioning and heat too.”
Nobody turned the lights off to save electricity. They turned them off because the machines never needed them. The light was for us.
The floor I wrote about before
In “The Factory Around the Motor” I wrote about Highland Park in 1913. Factories had electricity for a decade and got nothing from it, because the line shaft still dictated where every machine stood. Ford rebuilt the floor around the work, and a car that took over twelve hours to assemble took about ninety-three minutes.
I thought that was the end of the story. It was the middle.
The line shaft floor was organized around power. The moving line floor was organized around the work. The dark floor is organized around nothing a human needs. It is the third floor, and what it removed was the last thing the first two still assumed: that a person has to see the work for the work to happen.
A room goes dark when the machine no longer needs what the humans were there to provide.
Then the factory learned to write
This year the word crossed into software. In January, Dan Shapiro, the CEO of Glowforge, published a ladder of five levels for building software with models. Level three: the model writes most of the code and you review every line. Level four: you write the spec and a team of agents does the work. Level five he called the dark software factory, “a black box that turns specs into software.” Dark, he wrote, “because it’s a place where humans are neither needed nor welcome.”
By August, Klaviyo, a public company of 2,300 people, had told every employee to reach what it calls L3 by June, meaning you run teams of agents and validate what they produce, and had built an internal system it named the Dark Factory: give it a prompt and it writes the spec, splits the problem into subsystems, runs agents against each piece and builds through the weekend.1
Klaviyo built a dark factory to write Klaviyo. Nobody has built one to run it. That gap is this essay.
Retail’s floor is the campaign room
Every retailer already has a factory floor. It does not look like one because it has carpet. It is the room where the campaign calendar lives, and it runs the way Highland Park ran in 1912: a line shaft, operators, shifts.
The line shaft is the segment. A segment exists so that a human can look at a group of customers and decide what to send them, and the whole room is arranged around that act of looking: the segment builder, the journey canvas, the send window, the Monday planning meeting, the Friday report. The operators are the CRM team, and they are good. They are also a fixed capacity, and a fixed capacity facing millions of customers has one available output. It averages.
Braze, Salesforce, Klaviyo, Bloomreach, Synerise: these stay. They are the conveyors and the arms. What none of them can do from the inside is turn off the lights, because each was built on the assumption that a human looks at the segment before anything moves. Put a copilot inside one and you get, as I wrote in “The Keyboard Problem,” better keys. The room gets faster and stays lit. Rieseberg has the sharpest line on this: “Don’t ask the model to write like a short seller, ask it to do what a short seller does.” A copilot writes like a CRM manager. A worker does what the CRM manager does, with the manager’s instruments, and is measured on the manager’s number. A tool makes the operator faster inside a process designed around the operator’s limits. A workforce makes the limits obsolete.
The goal beyond the goal
Six days ago Felix Rieseberg, an engineering lead at Anthropic, published a tutorial on ambition for people who use models. He sorts ambition into four kinds: speed, altitude, depth, range. Altitude is the one this essay needs. His example is his own: he stopped asking the model to fix individual crashes and asked it to “keep our apps crash-free.” The question he gives you is “what’s the goal beyond the goal?” His tell for when the altitude is right: if you would otherwise hold meetings about the work, it fits.
The campaign planning meeting is the tell.
His worked example is a fictional company preparing its quarterly review, with three prompts at rising altitude. Translate them for a CMO:
Get Ilyas Kurklu’s stories in your inbox
Join Medium for free to get updates from this writer.
“Tell me what changed with our customers this quarter, why, and what we should do about it.”
“Name the three decisions that matter most right now and argue for and against each.”
“Design a better way to catch a customer leaving than a calendar. Validate it against our own history. Run it beside the current process for a quarter.”
The third prompt is a dark room with the door still open. Most companies never type it, and not because the model cannot answer. Rieseberg is honest about why: “Claude, go make money” does not work by itself in 2026. Before his model could keep apps crash-free, he had to make crash reports fetchable, map every app version to a commit, and give it machines to reproduce the bug on. That is not prompting. That is building a floor. Readers of the first essay will recognize the dynamo bolted to the old line shaft.
What has to be true before the lights go out
A dark factory is not a factory with robots in it. Fremont has robots. A dark factory is a floor that meets three conditions, and the same three decide whether any room can be dark, including rooms with carpet.
First, the output is measured without anyone looking. The robot arm has tolerances. The software has tests. Rieseberg’s rule is the same rule in a different coat: “Do not ask the model to judge itself.” Insist on an outside judge. For a commercial decision the outside judge is a holdout: the customers the machine was not allowed to touch, against whom every decision it did make is scored. Verification, not intelligence, is what gates darkness. A room full of brilliant decisions nobody can score has to keep its lights on.
Second, the work is formal. Written down far enough that the machine can hold all of it. For software, the spec. For a commercial decision, memory: what we know about the customer, what we know about the product, what the brand will and will not do. Retail did this to itself over fifteen years of segment builders and campaign calendars. The work is already formal. It is just executed by hand.
Third, the volume is past what a hand can hold. A strong CRM team keeps a few dozen segments in its head. The customer file holds millions of people.
Where all three hold, the room goes dark. Where one fails, you get Fremont.
Humans are underrated
On 13 April 2018 Elon Musk wrote: “Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.” Tesla had built what he called a crazy complex network of conveyor belts for the Model 3 line. It did not work. They tore it out and put people back.
Tesla did not fail on intelligence. It failed the second condition: assembling a brand new car at volume was not formal enough for the machines to hold, so the machines held a mess. Nobody should read this essay as permission to turn the lights off in a room nobody understands.
And nobody should read it as the thing I argued against in “From Operators to Owners.” A dark room does not empty the building. FANUC is a company of thousands; the people moved from the floor to designing the next robot and the next floor. Philips’s nine are not leftovers. They stand where the owner stands, at the end of the line, asking the only question that matters: did we land? What never goes dark is the brand, the spec, the judge, and the decision about what the machine is optimizing for. A dark room with the owners fired is not a dark office. It is a dark company.
Our dark room
Retail has done this once before. Around 2009 Tesco opened a supermarket in Croydon that no customer would ever enter. Pickers walked the aisles with tablets. The industry called it the dark store, and it called it a format, not a layoff. The dark store took the shopper out of the building. The dark office takes the operator out of the decision. Dark factories, dark stores, dark offices.
Here is ours. We built Maestro, an AI CRM Manager, to be the first dark room in retail, and I will describe it the way I would describe a floor. It holds three memories: the customer, the product, the brand. It reads behavior as intent, which is what we mean by Theory of Mind, and which is why its output can be restraint as well as a message. Hundreds of specialist agents argue each customer’s case against those memories, the brand’s constraints sit as a gate the argument cannot pass, and what comes out is one Golden Decision per person per day, sometimes silence. It owns the workflow end to end, replenishment, cross-sell, churn, winback, instead of predicting something and handing it to a human to act on. It is accountable for the outcome, not the activity. And every decision is scored against the holdout and fed back, so the room is smarter on Friday than it was on Monday. The channels do not change. The decision is dispatched into Braze, Salesforce, Klaviyo, Bloomreach or Synerise as a ready instruction. Nobody has to look at a segment for a decision to happen. That is the whole definition of dark.
At L’Occitane that room added 235% to post-purchase revenue. At Mumzworld it was live in seven days on the Braze stack the team already ran, and the team stayed where Philips put its nine: at the end of the line, reading the outcome. This is commercial intelligence, not marketing software, and retail is where it is proven first, not where it ends. The three conditions hold in banking, in airlines, in insurance. Retail just formalized its work earliest.
Forty-six days
Rieseberg ends his tutorial where I want to end this one, on self-efficacy. Roger Bannister ran the four-minute mile in May 1954 and was, for forty-six days, the only human alive who had done it. This April in London, Sabastian Sawe ran the first official marathon under two hours, and Yomif Kejelcha, eleven seconds behind him, ran under two hours in the same race.
Nobody got faster that afternoon. They had watched someone do it.
So, is a dark office possible? Not the whole office, not this year, and not in any company that skips the three conditions. One room is possible now, and in retail it is the room that decides what each customer gets, because that is the one room where the work is already formal, the volume is already past the hand, and a holdout can keep score. What stands between a retailer and that room is not the model. It is the belief that the room can be dark at all, and belief moves the way it moved on the track. One company does it, and the forty-six days begin.
The owner of the 1900 factory asked where the dynamo should go. The owner of the 1913 factory asked where the work should go. The owner in 2026 has one question left: what in this building still needs to see?
Ilyas Kurklu is the CEO and co-founder of Replenit, which builds AI workers for retail. Maestro, its first role, is an AI CRM Manager that decides one to one for every customer on the brand’s existing channels. This essay extends “The Factory Around the Motor” and answers the objection left open in “From Operators to Owners.”
Diagrams by Replenit. Amsterdam, 2026.
