The Fourth Factor
For three centuries a company was land, labour and capital. A fourth input has arrived, it is sold by the unit, and almost nobody has learned to allocate it yet.
Economists have a name for the basic things a business combines to make anything. They call them the factors of production, and for a long time there were only three.
Land, meaning the raw materials and the ground you build on. Labour, meaning human effort, the hours and the skill people bring. Capital, meaning the machinery, the buildings and the money you put to work. Every company that has ever traded is built out of those three. Get the mix right, add a margin, and you have a business.
Later, economists debated whether there was a fourth. The one most often suggested is enterprise, the work of organising the other three and carrying the risk, and people have put forward knowledge and technology since. The list was never fixed. That is the point. A factor of production is an input you buy, allocate, and combine with the others to produce something worth more than the parts. When a new input appears that meets that test, it earns a place on the list.
The newest input shows up on the invoice as tokens, and almost everyone has filed it in the wrong place. We reach for what we already know, so we call it technology, or digital transformation, or an IT line to be controlled. And that is understandable. It has come at us faster than almost anything before, and the only language we had for it came from software, so software is where we put it. But it is filed wrong, and that leads to the wrong decisions.
It only looks like a technology cost. It is intelligence, sold by the unit, and it behaves like a factor of production. You buy it. You allocate it. You combine it with labour and capital to produce output.
The only unusual thing about it is how general it is. Every tool you have bought until now does one defined job. Your payroll system runs payroll. Your CRM keeps your customer records and nothing else. A unit of intelligence can draft a contract, read a market, write code, answer a customer or plan a project, depending only on how you point it. It is closer to electricity than to any piece of software you have ever bought: an input you feed into almost any part of the business and get work back out.
So here is the point of this post. Intelligence is the fourth factor of production, and it is worth pricing and managing the same way you already handle the other three. You have a number and a method for land, for labour, for capital. You know the fully loaded cost of a role, what your money costs to borrow, what a machine earns over its life. For the fastest growing input on your profit and loss, you have neither. What I want to do here is give you both: a simple way to price it, and the equation behind it, in a form you could start using this week.
But pricing it is the easy part. The bigger change is what a new input like this does to everything around it. It changes how you use the others, and it changes what your people do most of all. And that comes down to one simple distinction, which is where we should start.
Your role is not your tasks
Ask most people what their job is and they will describe what they do. The analyst pulls the data, cleans it, builds the model, writes the report. The lawyer reviews the contract, marks it up, checks the clauses. The marketer drafts the campaign, builds the deck, schedules the posts. They will list activities. They believe the activities are the job.
They are not. The activities are tasks. The job is the role, and the role is the outcome you are accountable for, not the steps you take to reach it.
The analyst’s role is not pulling data. It is giving the business a reliable read on the market so it makes good decisions. Pulling the data was never the point. It was the means, the only means available, back when a person with a spreadsheet was the only way to get the read. The lawyer’s role is not marking up contracts. It is protecting the business from risk it cannot see. The marketer’s role is not building decks. It is creating demand. In every case the tasks were a path to the outcome, chosen because they were the best path on offer at the time.
This is why so many people feel threatened by AI, and why so many companies waste money on it. Both have confused the task with the role. The person looks at a machine doing their tasks and concludes their job is disappearing, because they defined the job as the tasks. The company looks at the same machine, automates the tasks, and stops there, because it too defined the job as the tasks. Both made the same mistake from opposite ends.
Separate the two and the picture changes completely. If a machine can do the tasks, the role is not under threat. The role is freed. The analyst whose data work is now automated is not redundant. The data work is done for them, so they move to the part that was always the real job: working out what the numbers say, sitting in the room where the decision is made, advising the people who carry the risk. That is a bigger job than the one they had before, not a smaller one. The reason demand for skilled people has risen rather than fallen in the fields where AI took hold is exactly this. The repetitive tasks were lifted away, and the people moved up into the part of the role that only a person can do.
So the real management question is not which jobs AI replaces. It is which tasks inside each role AI now does well, and whether you have moved your people off those tasks and into the part of the role that compounds. Get that right and you have rebuilt the job around its actual purpose. Get it wrong and you pay for the work twice.
The double spend
Here is the mistake we are seeing happen right now. It is an easy one to make, and most companies have made it.
A company hands its people AI and tells them to carry on with their existing jobs. They do. They get faster. The report that took 4 hours takes 2. Everyone reports feeling more productive in the survey.
And the company now pays twice for the same output. It pays the salary it always paid, and it pays for the tokens on top. The analyst still produces 1 report. It arrives a little sooner, the freed hours fill back up with more of the same kind of task, and nothing the business produces has changed. You have added a cost and held the output flat. That is the double spend, and it is where most corporate AI budgets are going right now. Real money in, no measurable change out.
The double spend is what confusing role and task looks like on the profit and loss account. You automated the tasks but left the role defined as the tasks, so all you bought was a faster version of the same list. The gain only appears when you move the person off the automated tasks and into the rest of the role. The analyst produces 5 analyses in the time they used to spend on 1, or stops producing reports altogether and goes to sit where the decisions are made. Same tools, opposite result. The whole difference is whether you redesigned the work around the role or left it wrapped around the task.
You cannot manage that by instinct. You need numbers. So here is the equation.
The equation
The equation has three parts, and each answers a different question a company has to ask about this new input.
The unit question: what does one good outcome cost?
Effective cost per outcome = cost per run ÷ acceptance rate
The investment question: is this worth doing?
Return on intelligence = value of the outcome ÷ total cost of producing it
where value of the outcome = (cost removed + new value created) × quality
and total cost = tokens + build + oversight
The strategy question: why does the advantage last?
Compounding advantage = return × iteration speed × context depth
Three equations, stacked. The first prices a single result. The second tells you whether the result is worth having. The third tells you whether anyone can take the advantage away from you. They look abstract on their own. With real numbers in them, they tell you exactly what to do.
Part one: what one good outcome costs
Effective cost per outcome = cost per run ÷ acceptance rate
This is the foundation. Get it wrong and nothing above it holds.
The cost per run is the tokens plus the human time spent directing and checking the work. The acceptance rate is the share of outputs good enough to use without meaningful rework. That second number is the one almost nobody measures, and it is the one that decides everything, because a failed run still costs you. The tokens were spent. The human still has to redo the work.
Measure a real workflow and the shape of it comes through quickly. A piece of competitor analysis takes a good analyst the best part of half a day: say 3.5 hours of focused work at a loaded cost of around £70 an hour, so roughly £245 for a finished piece, every time, with no economies of scale. Run the same work with AI in the loop and the tokens come in at about £6. A person spends half an hour or so directing and checking, call it £35. So a run costs around £41. But not everything comes back usable first time. A bit under two thirds does, an acceptance rate of about 63%. Divide £41 by 0.63 and the real cost is roughly £65 per usable piece.
£65 against £245. The AI path costs around a quarter as much per outcome, and arrives the same day rather than the next. That is a number a finance director can work with. And notice what it does to the comparison everyone makes by reflex. The person looking at the £6 token cost is missing that the real figure is £65, because the failed runs and the human checking are where the money goes. The tokens were the smallest part of the bill.
Acceptance rate matters so much that it can flip which option is cheaper. Picture two workflows. Workflow A costs £2.20 a run but under half its outputs are usable, around 45%, so the real cost is about £4.89 per outcome. Workflow B costs £3.40 a run but 90% are usable, so the real cost is about £3.78. The one that looked far cheaper per run is the dearer one per result. A company that buys on run cost alone picks the wrong workflow and never knows it did.
Now look once more, because this is where the double spend hides. If you produce that one piece with AI and the analyst still spends their half day on it, you have paid £245 of salary plus £6 of tokens for a single report that is no better than the one they produced before. You spent more to stand still. The £65 figure is only real if the analyst now turns out four or five pieces in the time one used to take, or leaves the analysis behind and moves into the decision it was feeding. The numbers only pay off if you reorganise the work around the role. Doing the sums and changing how people work are the same job. You cannot get one without the other.
Part two: whether it is worth doing
Return on intelligence = value of the outcome ÷ total cost of producing it
Cost per outcome tells you the price of a result. It does not tell you whether the result is worth having. For that you go up a level and weigh what the result is worth against what it costs you to produce.
Take the value side first, because it has two parts and most companies count only one. There is the cost you remove, the human hours no longer spent on tasks the machine now handles. And there is the value you create, the things that did not exist at all before: the analysis you could never previously justify, the deeper research, the speed that won the deal, the capacity you added without hiring. The double spend captures only the first part and usually loses even that. Real return needs both, the cost taken out and the new value put in.
Then the quality multiplier, because not all outcomes are equal. This is not the same as acceptance rate. Acceptance rate asked whether an output was usable at all. Quality asks how good the usable ones are. An output can pass review and still be worthless: it tells the client nothing they did not already know. So the value of the outcome is the cost removed plus the new value created, multiplied by how good the result is. Cheap and adequate scores poorly. Costly and excellent scores well. This multiplier is why chasing the cheapest possible output usually means chasing the wrong thing.
Now the cost side. It is made of three things, and most companies worry about the wrong one. First, the tokens. Second, the build: the cost of creating and maintaining the system that runs the work. Third, the oversight: the human time spent directing and checking. For most companies today the tokens are the smallest of the three by a wide margin. Build and oversight dominate. Which means the companies negotiating hardest on token price are squeezing the cheapest input while ignoring the two expensive ones. Raise the acceptance rate so a person checks less, and the whole ratio moves. Cut the token price by 30% and almost nothing happens.
So return on intelligence is not a token question. It is an organisational one. It asks how much of your work can run at a high acceptance rate with light oversight, and how much still needs a person in the loop. That is a question about how the company is built, not about the price list.
Part three: why the advantage lasts
Compounding advantage = return × iteration speed × context depth
Return tells you a workflow is worth running. It does not tell you whether it gives you an edge over the company across the road, because they can run the same workflow tomorrow. For an advantage that holds, you go up one final level.
Token prices fall for everyone on the same day. Model capability rises for everyone on the same day. Neither of those is an advantage, because your competitor receives the identical upgrade at the identical moment, from the same suppliers. Two things are not shared. The first is how fast you improve your own systems, your iteration speed. The second is how much of your proprietary knowledge those systems can draw on, your context depth. Both compound privately, inside your business, where no competitor can buy them by signing the same contract.
This is where durable advantage comes from: the orchestration and context layers, not the model. The model is a commodity that everyone rents at the same price. The edge is in how well you direct it and what you feed it, and both of those get better only through deliberate, repeated work that cannot be copied off a price list. A company that improves its systems weekly and feeds them years of proprietary context will pull away from a competitor running identical models, because the multiplier compounds in a place the competitor cannot reach.
That is the full equation. Cost per outcome at the bottom, return in the middle, compounding advantage at the top. Three numbers, each answering a real question, and all pointing to the same thing: the value is not in the tokens, it is in how you organise the work and the company around them.
What you do with what you free up
Once a task costs less to do with intelligence than with a person, you have freed something: time, money, capacity. What you do with it is a choice, and it is the most important one. There are three answers. They are different bets, and most companies fall into one by accident rather than choosing on purpose.
Same people, more AI. You keep every person and every salary. Nothing on the cost line changes. But each person now clears far more work, because the routine tasks are handled by the machine and their time goes to the part of the role that matters. The company solves more problems, serves more customers, builds the second product it never had the hands for. A team of 10 that used to handle 100 client jobs a year now handles 160 with the same 10 people. Output rises, cost stays flat. For most companies this is the strongest option, because the thing holding them back was never money. It was capacity. You added capacity without adding cost.
More people, more AI. This is the one that surprises people, and it is the case Jensen and others have been making. If every person paired with AI produces far more value, then hiring becomes more attractive, not less, because each new person now brings more output than a new person used to. So you add people on purpose and point them at growth: new business lines, new markets, more things built at once. The company grows its top line quickly, because every hire is multiplied by the tools they are given. This is the expansion play, and it is how a company turns a cost saving into a growth engine. There is a catch. It only works if you have real demand and real outcomes to aim the extra capacity at. Add people and AI to do more low value work, and you have only multiplied the double spend.
Fewer people, more AI. This is the version everyone fears and the one most overstated. You hold output where it is and reduce headcount, banking the difference as cost reduction. It works where the work is routine and the result does not improve with more human judgement. But it is the smallest prize, because the most you can ever save is what you were already spending. You cap your own upside at your old cost base. In a growing market, the company that takes this route hands its growth to the competitor who chose one of the other two. It can be the right call in a shrinking business or a purely cost driven one. As a default, it is the least ambitious choice available.
Notice that only one of those three is about saving money, and it is the weakest. The other two are about doing more: more problems solved, more products built, more markets entered, more growth. This is the point that gets lost when AI is filed under cost reduction. Cutting cost is the smallest reason to adopt intelligence. The real prize is everything you can now build, solve and grow that you could not touch before. You are not buying a cheaper way to do today’s work. You are buying the capacity to do far more of it, and to do work that was out of reach entirely.
Most of the day was never the job
There is a second gain, slower than the first and larger than most people expect.
Look at how an ordinary working day is spent. A large part of it goes not to the work itself but to the friction around it. Searching for a file someone saved somewhere. Copying numbers from one system into another. Reconciling two versions of the same spreadsheet. Chasing an approval. Reformatting a document so another tool will accept it. Sitting in a meeting whose only purpose is to line up what everyone is already doing. None of this is the role. All of it costs money. And almost none of it appears as its own line in the accounts, because it is hidden inside every salary in the building.
Companies learned to live with this waste because there was never an alternative. Every tool bought to fix one part of the problem added another system to wrangle and another set of data to reconcile, so the cost of running the business grew with every tool meant to reduce it. That now starts to reverse. As models get better at connecting systems and moving information between them, the data gets pulled together without a person copying it by hand, the matching up of one version against another happens on its own, and asking a question gives you the answer itself, not a list of places to look for it. The company needs fewer tools bolted together and far fewer hours spent making them talk to each other.
It does not happen in a quarter. The systems are tangled for real reasons and they will not untangle quickly. But the direction holds and it compounds. The first gain is doing each task more cheaply. The second is needing far less of the work that was never a real task at all. Together they free more capacity than any single workflow suggests, because you are not only speeding up the work, you are removing whole categories of work that should never have existed.
What this changes inside the company
Put the numbers and the three choices together and they point to something bigger than a budgeting exercise. They mean rethinking how the business itself works. Start with the job description, which in most companies is a list of tasks. It needs to become a description of the role, of what that person is there to achieve, because the tasks are increasingly the machine’s and the outcome is what is left. That one change rewrites the conversation in every appraisal, every hire, every restructure. You stop measuring people by what they do and start measuring them by what they are there to deliver, and by what they do with the hours the machine has handed back.
The rulebook gets rewritten next, and this is the part most companies avoid. Your roles were designed when people were the only option. Your approval chains were designed when documents moved by email. Your reporting cycles were designed when analysis took weeks rather than minutes. None of these are laws of nature. They are old decisions, sensible when they were made, almost none of them revisited since. Every one of them is now a place where you are either getting real value back or paying the double spend.
And where leaders spend their time changes for good. When the routine work runs at a known cost per outcome, the thing in short supply is no longer labour. It is judgement, direction, and the private knowledge that makes your systems sharper than the off-the-shelf ones any competitor can buy. That is where leaders should spend their time: not running the work, but deciding which work matters, feeding the systems what only your business knows, and reorganising faster than the competition can. The companies that understand this are rebuilding around it now. The ones that do not are buying a faster pencil and asking why the numbers have not moved.
Use it as a manual
You do not need a transformation programme to begin. You need 1 workflow and 4 numbers.
Pick something a team does over and over. Measure what it costs in human hours today. Run it with AI in the loop and record the token cost and the human oversight time. Track the acceptance rate, including every run that failed. Divide the cost per run by the acceptance rate, and you have your effective cost per outcome on a single line. Then ask the return question: what does this free your people to do that is worth more than the work it replaced, and are you moving them to it, or paying twice over. Then ask the compounding question: is this edge something a competitor could buy tomorrow, or does it get better every week because of how we direct the system and what we feed it.
1 workflow, then a second, then a portfolio of them, ranked by return, reviewed every month the way you review any other use of capital, because that is exactly what it is. You already know how to do this. You do it for land, for labour, for capital. Intelligence is the fourth on the list, and the strangest, and the one with the most room left in it to multiply what you produce. The only thing missing was the number. Now you have the equation that produces it.
None of this is settled, and none of it is obvious yet. We are still early. The shift is hard to hold in your head, because everything about the last few decades trained us to see this as technology, as a system to install and a programme to run. The real work is less about the tools than about how you think about it: moving from something you buy to something you allocate. Once that clicks, the numbers are the easy part.
The companies that pull ahead over the next few years will not be the ones with the biggest token bills or the newest models. Those are available to everyone at the same price on the same day. They will be the ones that learned to price intelligence the way they already price capital: deliberately, against outcomes they can measure, moving people off the tasks and into the parts of the role that only people can do. The equation is how you tell the two apart. The work is deciding what to do once you can.
If this gave you a sharper way to look at it, subscribe. The next anchor piece takes this same equation and points it at a single role, all the way down, to show what happens to one job when you stop automating its tasks and start rebuilding it around its purpose.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.