Lights out: när kontoret inte längre behövs

Picture a factory. It is maybe three o'clock in the morning. Right. Outside it is raining. It is cold, just miserable. But inside, inside the production line is screaming along at absolute maximum capacity. Just nonstop. Yeah, nonstop. Robots are welding, assembling, painting, moving parts with sub-millimeter precision. But here's the thing. It is pitch black inside. I mean, you literally could not see your hand in front of your face. There is not a single light bulb turned on. Because robots do not need to see. They do not need comfort. They do not need coffee breaks. And they certainly do not need light. They just need data. Exactly. And this is the concept of lights out manufacturing. It has been the holy grail of the industrial world for a really long time. It is that spooky, hyper-efficient dream where the factory just runs itself. But the reason we are doing a deep dive into this today is not because we care about car parts or widgets. No, we're looking at something much more disruptive. We have got a stack of research here that suggests this exact concept, this lights out idea, is leaving the factory floor and entering the office building. It is coming for the service sector. Which is just a wild thought. We are talking about the end of hiring people to do tasks and the beginning of building service factories. And just to set the stakes here for you listening, because I think people might hear automation and just roll their eyes. One of the sources we are looking at makes a comparison that completely stopped me in my tracks. The comparison between the old consulting model and the new way. Yeah. It compared a complex consulting task, something that requires actual analysis, not just data entry. Traditionally, this problem took about 40 billable human hours to solve. A full work week for a very smart person. Right. A full week. And in the service factory model. An AI agent solved it in five minutes. Wow. But the kicker is not the speed. I mean, speed is nice, but the kicker is the cost. It went from thousands of dollars in billable time to... Pennies. Yes. Pennies. The source mentions a few crowns because it is a European case study. But yeah, effectively pennies. It is the difference between buying a used car and buying a cup of coffee. So we are not just talking about a discount here. We are talking about a total collapse of the old cost structure. Or a total reinvention of the profit structure, depending on which side of the table you are sitting on. This is the shift from labor arbitrage to intelligence arbitrage. Okay. Let us unpack that because those terms are really the foundation of everything else we are going to talk about. Okay. Let us unpack that because those terms are really the foundation of everything else we are going to discuss today. Labor arbitrage is the business model we have all lived with for the last 30 years. It is. It is the reason your call center is in Manila or your back office accounting happens in Bangalore or Bucharest. It is purely financial. You have a company in a high cost city like New York or London. You have a task that needs doing. So you find a human being in a lower cost geography to do it. And the business model is just the gap. The profit is the difference between what the client pays me and what I pay the worker in Manila. And the business model is just the gap. The profit is the difference between what the client pays me and what I pay the worker in Manila. It is simple math. But, and this is the key point our sources highlight, it has a fatal flaw. It is heavy. Heavy in terms of management. Heavy in every single sense. It is linearly attached to human beings. If you land a massive new client, say a huge retail chain, these are invoices processed. You cannot just flip a switch. You have to hire 50 people. You need an office for them. You need HR to manage them. Right. You need middle managers to manage the people managing the people. Exactly. And people get sick. They quit. People have bad days. So your margins stay thin because you are always carrying the weight of a growing biological workforce. You are moving the work geographically, but you are not eliminating the effort. Which brings us to the new model. Intelligence arbitrage. I love this term because it sounds sophisticated, but it's actually brutally simple. You are not hiring cheaper humans anymore. You are buying tokens. Right. You are buying units of AI processing from the big model providers. Right. And here is the economic reality. And here is the economic reality. And here is the economic reality that is driving this entire shift. Tokens are cheaper than people. And unlike people, tokens get cheaper every single month. That is the wild part. I mean, in the human world, wages generally go up. Inflation, cost of living, experience. People get more expensive over time. But in the AI world, the cost of intelligence is in a freefall race to the bottom while the capability is racing to the top. So we are moving from billable hours, where time is money, to instant execution, where the cost is negligible. We are moving from renting. We are moving from spending time to selling outcomes. Usually, when we talk about this stuff, it feels very theoretical, like, oh, in 10 years, robots will do our taxes. But the research highlights a company called StrongDM as a proof of concept. This is not a prediction. This is a case study of a company running right now. Yeah. StrongDM is really the canary in the coal mine here. They work in software development. And the numbers here are just, well, they do not look real at first glance. They have three engineers. Three. Running. Running their entire production environment. Entirely. And to be clear, these are not three engineers typing furiously 18 hours a day, chugging energy drinks. That was my first thought. I pictured three guys doing the work of 50, just completely burning out. Not at all. None of those three engineers are writing code day to day. None of them are manually reviewing code either. So what are they actually doing? They are orchestrating. The system is a collection of AI agents. These agents follow instructions laid out in Markdown files. For the listener who is not a coder, Markdown is basically just a textbook. It is not a text file, right? It is just formatted text. Exactly. Think of it like a very detailed recipe card. The engineers maintain the recipe. The agents cook the meal. It is a lights out software factory. The humans design the logic and the machines execute the labor. Okay. I can see how that works for code because code is structured. It is logic. This, then that. But the source is undue. This model is coming for administration. For the messy stuff, invoices, tax filings, logistics, customer support. Chaotic reality of everyday business. Exactly. I mean, have you ever seen the inbox of a logistics manager? It is not structured data. It is an email from Dave with the subject line urgent and a blurry photo of a coffee stained receipt attached. How do you turn that into a lights out factory? That is the million dollar question right there. You cannot just throw AI at a messy inbox and hope for the best. That sounds like a recipe for hallucinations. It would be a total disaster. You would have the AI paying the wrong vendors or inventing tax codes. So the sources break this down. It is a really elegant three-step architecture. If you want to build a service factory, this is the blueprint. Let us walk through it. What is step one? Step one is input equals specifications. Which means no messy emails from Dave. Exactly. In a traditional firm, a client throws chaos over the wall. Right. They say, hey, sort this out. A human can figure that out using intuition. A factory cannot. In this model, the client's needs must be converted into stripped step by step specifications. So handle supplier invoice is not a vague request anymore. It becomes a rigid set of rules. Like if the date is past 30 days, do X. If the VAT is missing, do Y. You have to structure the chaos before it enters the machine. If you do not have good specs, you do not have a factory. You just have a very fast, very confused robot. Okay. So we have clear rules. Step two is something called digital twins. Now, usually I hear this term used in aerospace, like testing a jet engine in a simulation before you build the real thing. It is the exact same principle. But applied to administrative work. Think about safety. You do not want an AI agent, no matter how smart it is, logging into your real bank account or your live ERP system and making irreversible changes based on a guess. Right. Oops, I just wired a million dollars to the wrong account. Exactly. So these service factories create a digital twin. It is a simulation of the client system. The agent does the work there first. It books the invoice in the simulation. It reconciles the account in the simulation. Nothing touches the real world until it is verified. Which brings us to step three. How do we verify it? Because if a human has to go into the simulation and check the work, we have just reinvented the old model. We are right back to billable hours. Precisely. This is where scenarios replace reviews. This is the secret sauce of the whole operation. Instead of a human manager spot checking the work, the system runs automated scenarios against the agent's work in the digital twin. Give me a concrete example of that. What does a scenario look like in practice? OK. Let us say an agent processes that blurry invoice from Dave. It extracts the data and books it in the digital twin. The system instantly fires off a battery of questions. Does the total sum match the line items? Is the tax calculated correctly for this specific region? Is the vendor on the approved list? Is the date valid? So it is essentially grading its own homework. Constantly. And incredibly fast. If the answer is yes to all scenarios, the transaction is pushed to the real live scenario. The client just sees the work done. If the answer is no, say the tax is wrong, the agent retries. And if it still fails? Only then does it alert a human orchestrator. This is management by exception. You are not managing the work. You're managing the failures. This obviously changes the humans involved. If you do not need an army of administrators to process the invoices, who are you actually hiring? The source mentions we were going from 100 admins down to just three specific roles. Yes. And these roles are fascinating. Because they represent a completely new career path in the service sector. Who are they? First, you have the system builders. These are the technical crew. But they are not your standard IT support guys fixing printers. These are engineers building the digital twins, the routing layers, the evaluation systems. They are building the factory floor itself. Okay. So they are the architects. Then you have the orchestrators. Think of them as the plant managers. They do not micromanage the agents. They do not look at every single invoice. They manage the token budget. They watch the dashboards. They check the specifications are being followed. They only step in when the factory line jams. So they are managing the system, not the tasks. Exactly. And the third role is arguably the most critical for actually selling this stuff to the real world. The domain translators. The translators. These are the bridge to the client. Because remember, the client's reality is messy. The client works in dentistry or logistics or e-commerce. They do not speak markdown specifications. They do not know how to prompt an AI. Right. A trucking company knows trucking. They do not know data schema. The domain translator understands the client's business deeply. They know what a customs declaration actually looks like in the real world. And they translate that messy reality into the strict specs the factory needs. So if I am running a factory for dentists, I need someone who actually understands how a dentist office runs to write the rules. Precisely. You cannot code that empathy or that operational nuance. Yeah. You need a human to define the what so the machine can do the how. This sounds incredibly efficient. But let us talk about the money. Because if I am a service firm, let us say I am an accounting firm, and I used to bill by the hour. I loved it when things were slow. Slow meant more hours. More hours meant more revenue. The perverse incentive of the hourly billing model. Right. But now, the work takes five minutes. If I bill for five minutes, I have just destroyed my revenue. I am going out of business next Tuesday. If you stick to the old billing model, yes. You are right. But now, the work takes five minutes. If I bill for five minutes, I have just destroyed my revenue. I am going out of business next Tuesday. If you stick to the old billing model, yes. You are right. But now, the work takes five minutes. If I bill for five minutes, I have just destroyed my revenue. I am going out of business next Tuesday. If you stick to the old billing model, yes. You are dead. This shift requires a completely new business model. You cannot bill for time anymore. You have to bill for results. Outcome-based pricing. Right. The source lays out option A. You charge, say, a flat fee per handled invoice. Let us say a dollar per invoice. The client pays for the completed transaction. And here's the magic trick. If the AI gets cheaper. You keep the difference. This is the fundamental inversion of the model. In the old model, if you became more efficient, you billed fewer hours, so you made less money. In the factory model, if you become more efficient, or if the cost of tokens drops, your margin goes up. You capture the efficiency, not the client. That is a massive shift in psychology for a business owner. There was a second option mentioned too, right? Managing an intelligence budget. Yeah. That is a bit more transparent. In that model, the client pays the API fee. They pay the API costs directly. They pay OpenAI or whoever for the tokens. And the factory charges a management fee for the orchestration and infrastructure. It is lower margin, but it is safer for some clients who really want transparency. But the financial goal, the target revenue mentioned in the source is absolutely wild. It is the kind of number that makes investors sit up straight. Traditional service firms like law firms or accounting firms, they usually top out at revenue of about $600,000 per employee. That is considered elite performance. Sure. That is a great business. But these AI native factory firms, they are aiming for $3 million to $5 million per employee. That is a massive leap. That is software company margins in a service business. Because they have completely decoupled revenue from human effort. That is the very definition of scalability. Okay. I am going to play devil's advocate here. If this is so profitable and the technology is available, why does not the client just do it themselves? Why does not the trucking company just buy a chat GPT subscription, fire their accountants and keep that $3 million margin for themselves? That is the big question, isn't it? Why do you need the middleman? Why do you need the factory at all? Exactly. The sources identify a massive moat here, a protective wall that prevents clients from doing this easily. And it is not because AI is hard to use. It is the brownfield problem. Brownfield, meaning building on old, dirty land. Metaphorically, yes. Most companies are not brand new. They are brand new startups. They have legacy systems. They have 15 years of spaghetti code, weird patches, workarounds and undocumented databases. Oh, I have seen those systems. We all have. The ones where you have to click three times, hold down the shift key and pray to the IT gods just to print a pedia. Exactly. An AI agent right out of the box cannot navigate that. It will crash. It will fail. It does not know that client ID old is actually the field you need to use, not client ID new. So how does the service factory solve that? They perform what the source calls AI archaeology. I love that image. Indiana Jones, but for legacy SQL databases. It is a great term. They use agents to go in and document the mess. They map the old system out completely. They create the specifications from the chaos. This is boring, heavy, difficult work. It sounds painful. It is. And that is the moat. Most clients do not have the patience or the skill to do this. They just want the invoice paid. They do not want to spend six months mapping their own database. So what do you do? What do you do? What do you do? What do you do? What do you do? What do you do? What do you do? What do you do? What do you do? So the factory does the dirty work of cleaning up the past so the future can run smoothly. So barrier one is the mess. What is barrier two? Barrier two is the specification problem. We touched on this with the domain translator role. Writing instructions precise enough for a machine to follow blindly is a rare skill. It requires systems thinking. Most organizations do not operate on systems thinking. They operate on tribal knowledge. Right. Oh, just ask Susan. She knows how to handle the refunds. Or Steve knows which suppliers are reliable. And you cannot upload. You cannot upload Susan to the cloud. Not yet. Anyway. Exactly. So the service factory brings that skill. They are the ones who can articulate the process step by step. That is why the client pays them. They are not just paying for the processing power. They are paying for the structure. This leads perfectly to the strategy part. If you are listening to this and thinking, I want to build a service factory. What is the move? Do I build an admin factory for everyone? Do I try to be the Amazon of admin? The source is very clear on this point. Absolutely not. Do not try to be everything to everyone. You will fail. Why is that? Because of the edge cases. If you try to handle admin for a dentist, a trucking company, and a corporate lawyer, the amount of variation is simply too high. You will spend all your time writing new scenarios. You will need too many humans to handle the exceptions. So you have to niche down. You have to be ruthless about the niche. A factory specifically for dentist appointment booking. Or trucking company customs declarations. Or e-commerce returns for fashion brands. Why does narrowing it help so much? Because if you narrow the domain, you can cover 99% of the scenarios. The problems a dentist faces day to day are repetitive. The problems a trucking company faces are repetitive. But they are completely different from each other. If you cover 99% of the scenarios in one specific niche, you can actually turn the lights out. You can run without humans. And if you go broad? If you go broad, you have too many exceptions. You are just a consultancy with some cool new tools. You are not a factory. That distinction right there between a consultancy with tools and a factory. That feels like the core of this entire deep dive. It is. A consultancy solves problems with people aided by software. A factory solves problems with software aided by people. It sounds like a subtle semantic difference, but in terms of revenue and scale, it is a completely different universe. So let us bring this home. We are seeing a shift from managing people checking timesheets, motivating staff, dealing with turnover to orchestrating systems. That is the summary. We are moving from labor arbitrage to intelligence arbitrage. The input is the client's messy reality. The output is structured, verified data. And the fuel that runs the machine is no longer monthly salaries. It is tokens. It is a fascinating and honestly slightly intimidating future. But it makes sense. If the cost of intelligence drops to near zero, the value is not in doing the work anymore. The value is in designing the machine that does the work. That is the key takeaway. The value shifts entirely to the architecture. To the system builders and the translators. I want to leave you, the listener, with a final thought today. Something to really chew on. We have talked about how these firms are aiming for $3 million to $5 million in revenue per employee. That is a massive transfer of wealth and efficiency. It is an economic earthquake waiting to happen. So here is the question for you. Look at your own industry. Look at your own company. Are you building a factory right now? Or are you just buying hours? Because the companies that understand this, they are not just cutting costs. They are building a completely different species of business. And everyone else might just be waiting for the lights to go out for the last time. That is the real divide. Which side are you on? Thanks for diving deep with us today. We will see you on the next one.