1The belief comes first
Three children stand at an open nursery window in their nightclothes, and a boy they met that night tells them it is easy. There is a moment there I keep coming back to. The frightening part was never the flying; it is stepping out of the window, for something that only exists once you believe it does. The belief comes first, and then you fly.1
People say language is the new programming language. It always was. Long before we wrote in C or Python we wrote in stories, and a story is what tells a system how to act when there is no rule in front of it. Name a thing well and you have settled most of the arguments about what it should do. I have named a few companies, and none of the names was decoration. Each was a small bet about how the thing should behave, placed before the thing existed.
A generation of software founders is standing at that window now. Most of them are watching the flying, which is the technology, and the technology is the easy part. The frightening part is the sill, because everything they own is on the near side of it: the workflow, the seats, the roadmap, the moat. I know the near side well. I lived on it for fifteen years.
Everyone who builds companies carries a belief about what is outside. This is mine.
2A generation is about to discover it built the wrong thing
For twenty years the startup religion was simple, and I was a practising member for fifteen of them. Pick a workflow, build a product around it, capture the data, lock the customer into the workflow, charge per seat, expand. Do it well and you owned a system of record. Do it late and you shipped one more dashboard to people already drowning in software.
That game is ending. Software still matters. It got cheap enough that code is no longer where value hides.
Vertical SaaS was a good business because it bundled three things: the workflow, the data model and distribution. The application defined how work happened. The schema defined what reality was allowed to look like, and the sales motion decided who reached the buyer. That bundle was the moat. AI breaks the bundle, and once a moat cracks, even a little, value does not politely stay where it used to sit.
Music was the rehearsal. Before Spotify you bought the package: album, label, format, shelf space. In 1999, at the peak of the CD, the package was a twenty-two billion dollar business, and fifteen years later it was thirteen.2 Spotify pulled out the one activity people cared about, listening, and rebuilt the economics around access and use.
Spotify did not kill the record labels, and it never needed to. Their pain grew big enough that the old economic model came apart, and someone was standing there with a better one. The labels knew it. In December 2008 a filing in Luxembourg recorded that Sony BMG, Universal, Warner, EMI and the independents had taken eighteen per cent of Spotify between them, for 8,804 euros and 40 cents.3 The people being disrupted bought nearly a fifth of the disruption for the price of a used car.
Enterprise work is going through that unbundling. The buyer does not care which app created the task or which screen looked official. They care that the invoice reconciled, the contract risk was understood, the claim paid correctly and the audit evidence held.
The setup is in place again, and this time the fuse is lit. On a Friday, 30 January 2026, Anthropic released eleven open-source plugins for its office agent. By the following Tuesday some 285 billion dollars of software market value was gone, a trader at Jefferies had named it the SaaSpocalypse, and by late February the S&P index of software and services had lost a quarter of its value.4 The trigger was a folder of plugins, and no earnings report was involved.
Take Atlassian. It reported its first billion-dollar cloud quarter, with seats still growing, lost more than half its market value between January and March, and then cut a tenth of its people.5 The revenue was there and the belief was gone. Belief goes first on the way down too.
The debt has not noticed yet. By the end of 2025 private lenders had more than five hundred billion dollars out to SaaS companies, nearly a fifth of all direct lending, on the theory that a subscription is an annuity. In July the Bank for International Settlements looked at the part of that lending it can see, the listed funds, and found that the loans had not been affected yet, that no one had priced the software exposure differently, and that spreads had narrowed.6 The shares say the annuity is over, and the loan book says nothing has happened.
The systems of record remain, as sources, targets, constraints and audit anchors, and they stop being the prize. The application was never the value; it was the packaging. The value moves to the layer that turns activity into outcomes, and that layer is where the new companies are built.
3Promotion requires proof, not enthusiasm
I started vibe coding a year and a half ago.7 It began with one human and a handful of agents: high-taste prompting, fast iteration, and a willingness to throw away bad attempts before they hardened into process. It works, and it teaches its lesson quickly. Prompts carry taste and urgency, but they do not preserve truth, and they cannot coordinate workers that disagree. Authority cannot live in the prompt.
Then came hundreds of runs at a black-box document problem, where the visible result mattered more than any internal representation. At that scale a local green checkmark stopped meaning anything. The document had to survive contact with the real artefact, and of 3,367 test runs 2,337 failed.8 The failures forced evidence-based loops into the centre of the work.
Then came thousands of sessions, where no human can hold the thread. You treat software construction as a loop around disposable attempts: workers execute, governance decides, resources lease, oracles prove, the ledger remembers. It came to 3,136 agent sessions, 41,259 correction chains and 119 gates and oracles. The lesson at that scale was to make speed recoverable, and slowing down was never part of it: promotion requires proof, not enthusiasm.
Each jump in scale taught the same lesson from a new angle. Stop prompting one attempt at a time, and design the looped system that produces what you need.
Most people feel half of this. They have seen AI write code. They know a teenager can build in a weekend what used to take a team a quarter. But they file it under a productivity gain inside the old model, and it is nothing of the kind. When the cost of making software collapses, the reasons software companies existed collapse with it. The backlog loses its power. So do the roadmap and the workflow moat.
The hard thing was never building the thing; it is knowing what to build, why it matters, what outcome it should optimise for, what proof is required, and how the system should learn when reality disagrees.
Software is getting cheaper. Learning is getting more expensive.
4A business is a loop, not a map
A thermostat is the smallest honest example. It has a goal and it observes the room. It compares what is to what should be, acts, heat on or heat off, and does it again. There is no running once and declaring victory: the loop holds the room against the goal, continuously, and acts only on the difference.
Kubernetes is the thermostat grown up, and its manual explains itself with one.9 You declare what should be true. It looks at what is true. If a pod dies, it brings one back, and if capacity shifts, it adjusts. No manager reads a dashboard and decides the next move.
We have trained ourselves to see a business as a static picture instead: the org chart, with its boxes and reporting lines, or the pipeline, with work in on the left and revenue out on the right. Both are maps, and both are dead the moment they are drawn.
An organisation runs on the thermostat’s machinery and does not know it. Underneath the map, every function that matters is already a loop: close the books, renew the policy, patch the fleet. Today humans run those loops by hand, playing middleware between systems that should already know what they are trying to achieve.10 Your business has loops. The question is whether anyone built them on purpose.
Engineers call the pattern a reconciler: a control loop with a goal, which senses the world, compares it to the goal, acts on the difference and repeats. A reconciler keeps a reduced, biased picture of the environment, the few variables that move its goal, and treats the rest of the world as noise. That is the design, not a shortcut. A cell survives that way, closed around what keeps it alive and indifferent to everything else. The environment is the goal, and bias is the direction.
The goal can be almost anything a business can price: cash payback on a cost removed, a cleaner insurance renewal, a faster close. Point the loop at a different number and you get a procurement thermostat, or an insurance thermostat, or a finance one.
The pattern keeps reappearing because reality keeps demanding it. Biology calls it autopoiesis, systems that produce and maintain themselves. Cities and markets run on dense local signals and corrections that keep them alive without central control. Cybernetics is a field made of feedback, and computer science rebuilt it as controllers, schedulers and reconcilers. The oldest name in the genealogy is a job description: kybernetes, the steersman, the Greek word that gave us governor, cybernetics and Kubernetes.11
Static software fights this pattern. It says: define the process, freeze it into a workflow, force humans through it, extract structured data, and hire people to clean up the exceptions. That made sense when software was expensive and computers were stupid. It is madness when intelligence is cheap enough to read messy reality at runtime. The world is not a form, and work is not a sequence of buttons.
The last stack made humans simulate machines, and the next one makes machines participate in human systems.
5Money is the least fake truth
Take the smallest piece of a business anyone would pay for: reconcile this invoice, review this clause, answer this security questionnaire. Give it boundaries, inputs, evidence, a human checkpoint where one matters, a price and a result. That is a unit of work, a small, governed machine and not a workflow. It can be reused, improved, or killed if it does not work.
Every unit of work is also a sensor. It observes which inputs were needed, which document mattered, which policy changed the decision, which human correction altered the outcome, which source system was trusted, which exception repeated, which output was disputed, which model call was waste.
Legacy systems lagged that signal so badly that they saved everything: the data lake as a prayer. Store it all, structure it later, hope someone finds value in the swamp. That made sense when interpretation was expensive. It makes much less sense when a loop can ask a better question: did this signal improve an accepted outcome? If yes, promote it. If no, leave it as residue.
Graphs fall under that discipline too. A graph is valuable when it makes the next accepted unit of work cheaper or more auditable, and storing more relationships does not count. Candidate edges are not truth. Promote only what survives provenance, freshness and a comparison against the simpler baseline.
This is why activity-based pricing matters. It is the highest-order evidence layer there is, and the commercial model is almost a side effect. When a client pays because a unit of work created savings, processed a claim or proved compliance, the system receives a cleaner signal than any survey or roadmap vote could give it. Money is not the only truth, but in business it is the least fake one.12
Pay for accepted work and the loop becomes economic. The people who build units earn when they run and create value. Clients pay when work is accepted. Dead units stop earning, and useful ones compound. Traditional SaaS runs the opposite loop, where revenue can lock in for years while the product quietly gets worse.
Somebody still has to decide which number the loop holds and which evidence counts toward it, and that is a taste question. Business people misread taste as polish: nicer typography, rounded corners, a founder being precious. Taste is knowing what matters before the spreadsheet can prove it. It is the scar tissue of judgement.13 A security graph with taste knows which proof creates trust, where one without it is a pile of controls. A procurement graph with taste knows where the money leaks, where the policy is ignored and where the bargaining power lives, where one without it is a supplier database in costume.
If software can be generated around the problem, taste becomes the scarce input.
6The model does not know the client
A lot of enterprise software is not worth much. That is harsh, and it is also liberating. Much of it is routing, forms, approvals, field mapping, dashboards and reminders: people pretending that clicking buttons is knowledge work. It exists because building something more specific used to be too expensive, and changing the system underneath used to be too risky.
AI changes that. The common ninety per cent becomes automation infrastructure. The domain-specific ten per cent is where the value lives: evidence policy, exception logic, economic tuning, human judgement, taste and the loop. The learning from automation is valuable, and the screen is not.
This is why the consulting and service layer becomes more valuable, not less.14 It holds the operating reality software vendors rarely see: the playbooks, the workarounds, the exception queues, the escalation paths, the audit evidence, the human corrections. What traditional software treated as garbage is the shadow operating model of the enterprise. Thirty years of outsourcing and consulting residue can become gold.
The way there is not one training lake with every client’s data in it, which would be stupid and probably illegal, but lived operating knowledge turned into reusable units of work, late-binding ontologies, evidence policies and domain loops that prove their value through accepted outcomes. Residue becomes a unit of work, the unit becomes accepted work, accepted work becomes signal, and signal improves the next run. That is the company.15
When software stops being a destination, two things are left: intelligence and distribution. Right now almost all the money is betting on intelligence: models, agents, benchmarks, compute, data centres, energy. That bet is real. Intelligence will be valuable. It will also commoditise faster than most people want to admit. The frontier moves, open source catches up, routing abstracts the differences, and enterprises stop caring which model did the work if the outcome is trusted.
Distribution is different. In the enterprise it means trust, liability, context, access and reputation, and a download button has nothing to do with it. It is the ability to sit with a client when something breaks and say: we own this. No Fortune 500 company rebuilds core operations with a random vendor because a benchmark moved three per cent. They call the people they already trust.
Silicon Valley is betting that intelligence replaces the human layer. I am betting that intelligence makes the human layer more valuable when it is amplified correctly. Most enterprise AI is sold as a subtraction: fewer hands touching the work, fewer people the system still needs. A forest does not grow taller by removing trees.
Human in the loop is too small a phrase for this, because it makes the person a safety brake on an otherwise autonomous machine. The consultant closest to the problem can now author a unit of work. The domain expert can turn judgement into a reusable asset, and the firm can leave behind something better than a slide deck.
Taste, distribution and learning: that is the new moat. The model does not know the client. The human does.
7Do not play the board someone else set up
On 21 October 1805 the rules of naval war said to form a line parallel to the enemy’s and trade broadsides until somebody gave way. Nelson sailed at the French and Spanish line in two columns, at right angles, and broke it.16 He broke the line instead of sailing it.
Charlie Munger’s version was invert, always invert, a rule he took from the mathematician Jacobi: do not only ask how to win the game, ask which game you should not be playing at all.
The obvious game today is to build another vertical SaaS company with AI features. Pick a category, raise money, build workflows, hire sales, overspend on acquisition, bolt on a copilot, and pretend the data model is a moat while the workflow is generated around the problem. That is chess on someone else’s board.
Fuck chess.
Invert it. What if the most valuable software company of this era does not own the workflow, but owns the ability to create, price, govern, deploy and learn from units of work across many workflows? What if the vertical app becomes an implementation detail, the static schema a temporary binding, the operational residue the training signal, and the consultant distribution to amplify, not a labour cost to remove?
Pull the old bundle apart and the value redistributes into three layers, each with its own economics and its own kind of company. On top are the new startups: small domain teams building learning systems that plug into any system of work, optimise it for a purpose, and pull better evidence into the loop. In the middle is the runtime, the execution layer where a unit of work is the first-class artefact and cost can be priced against the work removed. Around them is the reach: the firms the enterprise already trusts, with delivery at scale, the liability wrapper, and the economies of reusing units across an installed base.
The company that lives in the top layer is a small team with deep taste for one domain, building a learning system that runs on someone else’s rails and prices itself on the cost it removes. It is neither a platform nor a consultancy.
Take security. The hidden cost of security is administrative: the labour of proving that controls are in place. A company built on the loop watches for one change, a CVE landing, maps it to the obligations it touches, runs the work and ledgers the proof while it is fresh. Fresh, provable evidence lowers the insurance premium and answers the regulator’s clock, and both return as cleaner signal. It is a thermostat whose setpoint is cash. It has no model of the whole enterprise and needs none: it watches the one variable that matters, whether a control is provably in place, and lets everything else stay noise.
In year one the partner sells and integrates, and earns the majority. As the work converts to software the startup runs, the asset carries the value, not the integration. The drafting models are swappable. The per-client corpus of decisions and corrections is not, and under outcome pricing every human correction that feeds the loop is margin.
Two founders and one domain are enough, with a learning loop that gets cheaper every time it runs. Do not build the palace; build the rails the next thousand palaces run on.
8Nobody drew the org chart
Put a photograph of lichen next to a field of Truchet tiles seeded at random, and the eye has trouble telling which one somebody designed. The lichen radiates from a fixed point, cell by cell, on a local rule: grow outward, branch when there is enough to spend, stop when crowded. The tiles are computed, arc by arc, from the same origin, with two orientations and one honest coin-flip per square.17 At scale both read as one meandering line, the shape you would trace across a river delta or a root system, though no individual tile was ever asked to look like one. It is one underlying process: a simple, local, repeated rule, never art-directed from above.
Let me take it one step further. Conway’s law says organisations end up shaping the software they build: bureaucratic company, bureaucratic software.18 Invert it on purpose and shape the organisation like software. Write down how one small piece behaves. Let enough of them run. See what shape falls out.
A pod is the growing tip: three to seven people, forward-deployed inside a client’s delivery, never asked to see the whole company, only to be right about the one client in front of them. A pod does not scale by getting bigger, any more than lichen scales a branch by making it longer: when it has enough to spend, it branches into a new pod, cloned from what worked.
A spoke is the mycelium, not a department. It curates a domain, procurement or finance or payroll, across every client at once, and moves what one pod learned to the pod that is about to need it, before that pod relearns it the slow way. What a client needs is never the spoke’s decision; it stays with the pod, the way the tree decides what is worth growing toward, not the network underneath it. A pod knows the client and a spoke knows the network, and neither one runs the other.
The hub is what foresters call a hub tree and not the trunk of anything: the most connected node in the network, the one every other tree ends up drawing through, because of where it sits and not because it is the biggest.19 Cap it at a hundred and fifty people, for the reason a hub tree never tries to become the forest. Past a certain size, adding more of yourself stops helping and starts being the bureaucracy the company exists to replace.
Underneath runs the part no client ever sees, the nervous system and not the face. It turns the work of middle management, intake, routing, ownership, approvals, escalation and quality gates, into product. The client gets a pod that already knew and work that arrived faster than it should have, the way nobody sees mycelium unless they dig, or the dispatch layer behind the car that shows up. Invisible is the design.
The old model rewarded careful feature planning because software was expensive. Every feature was maintenance debt, and saying no was usually right. When the cost of trying collapses, reflexive restraint becomes the wrong instinct. Build quickly, ship agentically, get close to the client, watch the ass-kicking honestly, keep what works and throw away what does not.20 The company should behave more like Kubernetes than a corporate hierarchy: small units with clear purpose, bounded autonomy, health checks, fast replacement, continuous reconciliation. If a pod works, scale the pattern. If it does not, kill it and learn.21 The centre should become a scaling layer and never a permission machine.
This is asymmetric warfare against entrenched systems: small units, local knowledge, intimate terrain, and a platform that lets the pattern travel. It is more Che Guevara than NATO procurement. Reward the people who move with economics, not with vague partner points. Create accepted outcomes and you earn. Build reusable units and you compound. When the unit dies, the economics stop.
Nobody drew this shape; it is what the rule looks like once enough pods have run it.
9After an ending, belief is what carries the purpose
It would be a mistake to turn this into a roadmap. Roadmaps feel safe because they pretend the future is a sequence of planned features. What is happening is the correction of an imbalance. For decades enterprise software accumulated power because building was expensive, integration was hard and structured data created lock-in. AI breaks enough of that for the system to rebalance. Markets correct.
Capital is flooding into intelligence, and enterprises still need trust, distribution, liability and judgement. One side is a stock: by my count roughly eighteen and a half trillion dollars of capitalised belief that intelligence wins, resting on 7.6 trillion dollars of promised capital expenditure by 2031.22 The other side is a flow: about 3.7 trillion dollars a year of enterprise budgets, for software, services and process labour, that renew whether or not the market keeps believing. If five per cent of the flow crosses over, that is more than a hundred and eighty billion dollars a year finding a new home. At ten per cent it is a category. At twenty-five it is the redistribution of enterprise work. Nobody has to be wrong about intelligence for this to work. I am not betting against the crowded side; I am standing where the budgets land when they move.
The bet is that the future of software startups is learning, not code, and that the future of enterprise AI is outcomes, not agents. Work itself ends up as governed systems that understand intent, sense reality, act through small units, prove outcomes and improve continuously, which is neither humans clicking buttons faster nor robots clicking them instead. It is what I am building at Beyond Work, it is under construction, and I claim no result for it here.
The old software company asked what it could build. The new one asks what it can learn that makes every future unit of work better. That is a better question, and it is also a much bigger company.
Fight Club opens with a man furnishing his flat out of IKEA and asking himself “what kind of dining set defines me as a person”.23 He is no fool. He owns things, the things are fine, and he would like the world to hold still so that he can go on owning them. Then the flat blows up.
A great deal of the software industry is that man this year. It is clutching its IKEA, the seats and the renewals and the roadmap, asking which dashboard defines it as a company, and hoping the world does not change. The world has changed. The shares went first. The rest will go quietly at first, and then all at once, like buildings on a skyline tumbling to the ground.
The film ends at a window too. The narrator has let go of the furniture, and then of the man he invented to replace it, and he stands on a high floor holding a woman’s hand while the towers that keep the debt record come down. It is an ending, and neither of them looks away. People who have something to believe in can watch an ending, and people who only have the furniture cannot.
Peter Pan is the belief before a beginning, and Fight Club is the belief after an ending. A founder needs both, because a company is a purpose, and a purpose has to be carried across the gap where the old thing has stopped working and the new thing is not yet proved. Nothing carries it across except belief.
So believe in something new, and then go and build the proof. Belief is what carries the purpose. The rest is furniture.24