AI Will Take the SaaS Castle from the Inside
Software companies were never protected by the code they wrote. They were protected by the habits of the teams that use their products. AI companies are now walking into the place where those habits are made: the meeting room.
00The castle
On November 12, 2026, Anthropic announced that Claude can join meetings. Claude enters the video call as a participant everyone can see. It notes what is said, lists the tasks, and starts writing the document while the meeting is still going.
Anthropic had started the year differently. At the end of January it released plugins for legal, sales and marketing work. Investors feared Claude would replace the software that does those jobs; software stocks lost more than 950 billion dollars in a week. That day Anthropic stood outside, facing the software companies with a rival product. In November it came inside.
Since February everyone has been asking the same question: what is left to keep software companies standing? The debate turns around one word: moat. The thing that protects a company from its rivals.
Turkish does something interesting with that word. In English the moat is the castle's defense. The owner sits inside, the rival stays on the far bank. In Turkish the same ditch is the obstacle you have to get across. We have an idiom, "getting a camel to jump a ditch". The Turkish Language Association's dictionary explains it as getting someone to do something very hard, close to impossible. Its example sentence is from the essayist Salâh Birsel: "İnsanlara bir şeyi anlatmak deveye hendek atlatmaktan güçtür." Explaining something to people is harder than getting a camel to jump a ditch.
In English the ditch is the castle's problem. In Turkish it is the camel's.
I will keep the two words apart in this essay. Moat means what it means in the debate: what keeps a company standing. Ditch means what it means in Turkish: the obstacle to get across.
Here is my claim. Everyone looks for the moat of software in the product, in the code, in the features. I think what keeps a piece of software standing is that it has settled into the way its users work. When AI arrives, that way of working has to change. That is the hard part. AI made writing code easy; it did not make it easy to explain a new way of working to people. That job is still getting a camel to jump a ditch. Whoever gets the team across takes the seat in the team's new order.
We have another old saying: a castle is taken from the inside. The title comes from there, and it can be read three ways. The one who takes the castle from the inside can be the team that gets its own camel across. It can be the AI company that walks in through the meeting room. Or it can be a single person inside the team. On the way to an answer we will stop at a car factory, an electric motor, and a backup nobody ever tested.
01Glossary
Seven words come up often. Let's agree on them first.
- SaaS: renting software by the month, not buying it.
- Seat: the unit of rent. You pay separately for each person who uses the software.
- Usage fee: the new unit that arrives next to the seat. You pay for as much as the AI works.
- Agent: an AI that carries out a job step by step on its own.
- Moat: what keeps a company standing for a long time.
- Ditch: in this essay, the Turkish meaning: the obstacle to get across.
- Camel: the company's working order. Who decides, how work is shared, what counts as done.
02The debate
I have been following the debate on software moats for a while. Everyone has an answer. One says distribution, one says data, one says taste, one says speed. The AI companies say "we will do all of it".
Put the answers side by side and something stands out: everyone points at what they already hold. The company with hundreds of millions of users says distribution. The fund that invests in data companies says data. The organization that earns from speeding up startups says speed. It is hard to call this a debate. It is closer to an advertising war.
It gets stranger when you look at the owner of the word, Warren Buffett. Buffett stays away from fast-changing industries to begin with. In 2007 he set a measure along these lines: a moat that has to be dug again and again is no moat. Software is one of the fastest-changing industries there is; the product is rewritten every year. Today's debate looks for Buffett's word in a place where, by Buffett's own measure, it cannot be found. Then, not finding it, it panics: software has no moat left.
Yet the investors in this debate write it themselves: code never protected anyone. What kept software companies standing was how hard it is for a customer to switch, the records piled up over years, the brand.
So the sentence "AI made code cheap, software is over" is built wrong from the start. Code got cheap, true. But code was not the protection. The question to ask is this: are the things that did protect still in place?
03The walls
They are. But not every piece of software is protected by the same thing. The debate always runs on giants like Salesforce. Yet there are software companies in every industry, of every size. I sort them into three kinds.
The first is software you are stuck with. Try an experiment. Say code is free and you can write a Salesforce clone in an afternoon. When does your company leave Salesforce? Most likely not for a long time. You would have to move years of customer records, teach the sales team new screens, and rebuild every connection to other software. Moving the memory and the data is the trouble.
Three things stand behind this habit. One is the company's past: if the work was built on SAP years ago, it has to go on that way. Another is the ecosystem you live in: if your company uses Microsoft, you lean toward what works with Microsoft. The third is the economy built around the software: consultants, partners and plugin sellers make their living from it. A company that leaves the software has to leave these people too.
The second is software you chose. Notion or Typeform entered nobody's life by force. They took a job that already existed, taking notes or collecting a form, and rebuilt it with a different eye. People liked that view and made a habit of it. What protects this kind is the view it brought to the job.
The third is the tool that does a single job: cutting a video, collecting a form, tracking a task. Of the three, its habit is the shallowest.
Still, this kind should not be looked down on. I worked at Jotform, founded in 2006 and today one of the largest online form tools in the world. Jotform began with one belief: a form should not be coded from scratch every time; it should be easy to build in an online editor. The company worked on that idea patiently for years. The idea stayed the same; the conditions changed. In the rush to digital that came with the pandemic, the company found its worth. A single-job tool can become a giant this way too. What made it a giant was years of patience and the right moment arriving. Being able to write the same tool in an afternoon today replaces neither.
The numbers differ from kind to kind. In the February crash, the customers of large public software companies did not leave: a customer who paid 100 dollars last year pays 109 on average this year, and that ratio has barely moved in two years. What fell was the value investors put on them: companies that sold for about 6 times their yearly revenue came down to 3.4 times. With smaller companies the picture is different. In a measurement covering thousands of companies, the same customer pays 82 dollars where they used to pay 100.
The "AI killed SaaS" stories also shrink when you look closely. The most told one was the payments company Klarna: its chief executive said they had shut Salesforce off. A year later he corrected it. They had not replaced the software with a language model.
It went the same way in history. When open source arrived, and later when SaaS arrived, "the end of old software" was declared. The big companies lost part of their business and kept the customer. In The Freeloading Heir I wrote that the gain moves from what can be copied to what cannot. Same here: code gets copied; being stuck, the view and the habit do not.
The walls are standing; thick for the big ones, thin for the small. But a standing wall still wears down.
04The wear
It wears in three places.
First, the value put on software. A customer who sees that a piece of software can be written in an afternoon starts to question the money paid for it all these years. Even with no plan to write it, the customer sits down to negotiate with "we could build this ourselves". Customers are not leaving today. But if software gets cheaper in their eyes, the money they are willing to pay can shrink too.
Second, features. The easy answer in a software company's hands is to ship more features. When code got cheap, features got cheap; every release can add a new button to the screen. But the number of features a user can learn did not grow. A feature nobody uses produces no value; it becomes maintenance. The worth of a piece of software is set by how many of its features make it into the team's daily work.
Third, profit. The shape of the price is changing: software companies do not remove the seat, they add a usage fee next to it. You pay extra as the agent works. But not all of that fee stays with the software company. A company that adds AI to its product pays the owner of the model every time its customer uses the feature. A classic software company keeps about 76 of every 100 dollars it earns as gross profit; in AI products that number is between 45 and 53. And if AI costs go up? The software company either pays out of its own profit or passes it on to the customer.
I think SaaS is not ending, it is changing shape. The customer stays put. But the software company no longer sets its price alone. Who does?
05The gate
The AI companies do, because they hold the castle gate. What brought the stocks down in February was not a clone written by some startup either. It was Anthropic's plugins.
By gate I mean the place where a person reaches a piece of software. It has three parts: the account, the wallet and the window. The "Sign in with ChatGPT" button takes the first two: you enter another app with your ChatGPT account, and your own ChatGPT subscription pays that app's AI bill. The window is moving the same way: other companies' apps now open inside the chat.
From the user's side the question is simple. I ask it myself: when I solve everything in Claude, why would I go to another window?
The oldest way to hold a gate is to be the default. Microsoft made the agent the default in Office; without choosing anything, you start your work with the agent. This is also why software companies move to where people already are. Teamflect, where I work today, is Microsoft Teams software too: it built performance management inside Teams, because people spend most of their day there.
The strange part is that the one opening the gate from the inside is the software company itself. This year Salesforce made its software usable from inside Claude: you can reach customer records by asking Claude, without opening a Salesforce screen. Salesforce has no other choice; it has to be wherever its customer is. But every software company that opens its data to the agent removes one of the reasons its customer had to come to its screen. When the screen goes, two things are left in the software company's hands: data and service.
For the small software maker the picture is harsher. Starting a software company used to have a threshold. People with money or stamina hired staff, had the software written, and fought to get it a foothold in the market. To find an investor you first had to make someone believe in the idea. Not everyone got over that threshold. Now the threshold is gone and starting a company looks open to all.
The numbers say something else. In the first half of 2026, almost as many new apps entered the app stores as in all of the year before. Downloads grew by 2 percent. The shelf filled up, the customers did not multiply; more people are sharing the same customer. This may be an economy that pulls the average down, not up. Producing is not the problem; holding on is. When everyone gets their work done in one window, nobody visits the tool that makes them open another.
A serious objection comes out of this: the moat is not decided inside the team, it is decided at the gate; whoever holds the gate wins. I agree with half of it. Two places decide the moat today: the gate, and the team's own order. But here is the odd thing: the one who holds the gate also has to come inside to win. Why?
06The math
Because the math demands it. By The Economist's calculation this November, for the money put into data centers to pay off, AI has to produce roughly 3 trillion dollars of revenue a year by 2030. That is close to a tenth of all wages paid to workers in the rich countries.
What the number means: this money does not come out of the software budget. What a company spends on software is small next to what it spends on wages. For the math to work, the money has to come from the wage budget. In The Itch I had reached the same place by another road: a team of fifty can become three people and an AI subscription.
Today we are far from there. An ordinary company spends around 16 dollars a month per employee on AI.
A chat window used by one person will not produce that money. Wages are paid for work done together: deciding, meeting, following up. For the math to work, AI has to get into the team's shared work: the meeting where the decision is made, the channel where the work is discussed, the board where the task is tracked. The products of the last year arrived in exactly this order. First the company chat channel, then the meeting.
Anthropic also wrote it plainly: for an agent, what is not written down and reachable does not exist. So the team has to work where the agent can read.
These rooms have an owner: the communication software. Slack, Teams and Zoom hold the place where work is discussed. Today Claude enters there as a guest: a participant in the Slack channel, a participant in the meeting. The question is whether ChatGPT or Claude will build their own chat for teams. Anthropic already has its own spreadsheet and whiteboard products. The hosts see the danger too: Salesforce, which owns Slack, last year restricted the bulk transfer of Slack conversations to other companies' AI products. The first walls to be tested are those of the communication software, because the room the agent wants to enter is theirs.
The data center's debt will be paid out of your Tuesday morning meeting.
This is why the software that comes to take notes in the meeting is not only a note-taker. But is getting inside enough? Can someone from outside teach a team a new way of working? In the 1980s a car company tried.
07Toyota
In 1983 General Motors and Toyota set up a joint company in a closed GM plant in California. GM's aim was to see Toyota's production method up close. Toyota hid nothing. Workers were sent to Japan and trained. The plant became one of the places building the best-quality cars in America.
GM could not get the same result in its own plants. The method was in the open, the books had been written, the plant could be toured. What was missing was not knowledge. Years later a radio program asked a worker who had been trained at the joint plant and gone back to GM: when you returned, did anyone ask you what you had learned? The answer was short. Nobody had asked.
Knowledge of the method reached GM. What did not reach it was the way the managers, the union and the workers at the joint plant worked together. GM could not change the order in its own plants.
What protected Toyota was that its way of working could not be learned from outside. The ditch GM could not cross was Toyota's moat. There are not two things here. Toyota looks at the same ditch from the inside and calls it protection; GM looks at it from the outside and calls it an obstacle.
The strongest objection to calling a way of working a moat comes from Buffett's measure: a moat that has to be dug again and again is no moat. A way of working is exactly that kind of thing. It is not built once and left; it is rebuilt with every new employee, every new tool. By Buffett's measure it does not count. The answer I have is Toyota itself: the joint plant kept the same way of working until it closed, for twenty-seven years. But we do not know how often a company will have to rebuild its working order in the age of AI.
Say Toyota wrote its method down today as an instruction file and handed it to GM's AI agent. The agent reads the file, summarizes it, draws up a plan for each plant. Does GM cross the ditch? I think not. GM had the knowledge in the 1980s too. If the file is not enough, what is left is being inside: where the work is done, every day. This is why it matters that AI companies are entering the meeting.
What cannot be copied is not the method; it is the way people work together with that method. So what is working together, exactly?
08Work
First we have to ask what work is. In my notes I wrote this: we play games inside social structures, and at the end of the game we produce something useful. Work is one of these games. It has rules: who decides, who approves, what counts as done, who pays whom. Software is the board of this game; the interfaces, the processes and the flow of money are laid out on it. The rules sit between the players. Copying the board is not enough to move the players to a new board.
Engineers like tasks with clear edges that end when delivered. But work is mostly not like that. Work is communication, planning, deciding, arguing; it is finding out together what to do.
Communication has a simple definition too: you say something, you get something back. Until the other person shows they understood, what you said hangs in the air.
I lived this every day in my years as a manager. I managed large teams. Making the decision was the easy part. The hard part was explaining the decision to each person in the way that person would understand, and then going back now and then to check. Some people get it from a clean definition. Some from an analogy. Some you have to show once. That is where a manager's hours go.
Marx explains this with an orchestra: a violinist playing alone directs himself; an orchestra needs a conductor. He says something more important too. When people work together, more comes out than the sum of their separate powers. The employer pays a hundred people one by one; he does not pay for the power that comes from the hundred working together. That power appears on no payroll.
Where does AI stand in this picture? It produces a good document, but it does not build the agreement. What happens today is this: someone has their Claude write a document and sends it to a colleague, and the colleague's Claude does the work with that document. The answer comes back the same way. We have become couriers carrying documents between agents that cannot talk to each other. The document gets across. The "did you get it, yes I got it" part does not.
Will this change if agents start talking to each other? Maybe. But today a language model works, at the end of the day, with one person. Even if ten people type at once, it moves forward on one person's word.
Working together gives birth to a power. AI is not inside that power yet; it stands next to people one at a time. But is it not enough to speed people up one at a time?
09The bottleneck
It is not.
That one person gets faster is certain. In experiments, a worker who uses AI finishes the job clearly sooner. I see the same in my own work.
But the shared work is treading water. Gallup asked employees who use AI often: 83 percent said the time they spend in meetings had not changed. In software teams writing code got faster; having someone else read and approve that code got longer. Looked at across the whole company, the gain is almost invisible.
Factory management named this long ago: the speed of a system is set by its slowest link. According to the management writer Eliyahu Goldratt, an hour lost at the bottleneck is lost to the whole system, and an hour gained anywhere else is worth nothing. If you give everyone an agent and leave the meetings as they are, the company speeds up as much as the meetings allow.
This is not the first time. The electric motor entered factories in the 1880s. For forty years productivity did not rise as expected, because the motor was placed inside the old layout. There was one power source in the middle of the factory, a long shaft along the ceiling, and belts running down from the shaft to the machines. The electric motor replaced the steam engine and everything else stayed the same. The gain came when factories were redrawn and every machine got its own motor. This example does not show that we will wait forty years. It shows the order: the tool comes first, the gain comes when the layout changes.
Today AI is a new motor fitted into the old layout: everyone has a chat window in hand, and the team works through the old chain of approvals. In The Itch I said AI would not change the nature of work. I still think so: the tool alone does not change it; what will change it is the way the team works.
I should write the strongest objection to this view too. In one experiment a single person using AI produced as much as a two-person team without it. Maybe the team will not speed up; it will shrink. It could. But a smaller team is still a team. The number of people who have to agree goes down; the need to agree does not end.
There is a harsher objection. What if the team does not shrink but disappears? If nobody works in the jobs done at a computer, who will use the software, and why? I do not think the team will go to zero. As I described in The Freeloading Heir, companies will keep people for recognition even when they hand the work to the model: a manager needs someone to see what they do. But recognition does not take ten people; two will do. Employment shrinks either way. If the team shrinks to almost nothing, everything I have argued so far loses its meaning: there is no habit left to protect, no camel to get across, no meeting to enter. Software stops being a tool people use and becomes a service that agents call in the background. The question changes too: in an economy that pays no wages, who pays for the software? The people who earned the wages were also the economy's biggest buyers. The answer to that question is in The Freeloading Heir as well: on that day the issue stops being software and becomes ownership. AI was made from the accumulated labor of humanity; whoever takes the inheritance takes the debt too. That debt is not settled with a small share; it is settled with a say over the infrastructure. What I describe here holds for the years until that day: the years when people still work together. I do not know how long those years will last.
Until that day, the ones who will change the order are people. Are they ready?
10The camel
Not everyone is in the same place. On the same team, with the same tool, one person carries a long job from start to finish with the agent; the person next to them is still looking for what to say to it. Both have done the same job well for years.
The difference is not in knowing how to code, and not in talent. Working with an agent is a new habit: splitting the job into parts, giving the agent the context, carrying a long job through to the end. Nobody learned this at school or at work; those who started early found it by trying. The measurements also show that the difference comes from experience: newcomers finish the job successfully half as often as experienced users.
"Then let those who know write it down, and those who don't read it," you might say. It is not enough. Say an experienced person put their way of working in writing: instructions, checklists, all in one place. One line in the list says: before going live, try restoring from the backup. The team that receives the file is building an appointment site. The agent skips that line, nobody notices, and the site goes live with a backup that was never tested. The line was there.
What is written is the step. What is not written is why that step cannot be skipped. People know more than they can say, and what we call mastery sits in the part that cannot be said. That is what protected Toyota from its rival. But the same thing is an obstacle inside the company: what your rival cannot copy, you cannot easily teach your own new hire either. The ditch runs through the inside of the castle too.
This ditch used to be crossed by apprenticeship. The master told the apprentice what he knew, but the apprentice learned the real part while working next to the master: standing by, watching, doing a small piece of the job. Work done with an agent cannot be watched. Say your manager comes over: "Open your Claude Code, let's see what you did." What is on the screen? A chat window. Which path you dropped and why, where you stopped the agent: none of it shows. And your manager cannot say "you are using it wrong here, it would be better this way". The work stays between you and the AI, and it is only as good as your own knowledge and understanding.
The solution may look like recording everything. It is not. People are uneasy when their clumsy exchanges are watched. A person who is watched does the job by the rule and hides the shortcuts that work. In Turkey the Personal Data Protection Board drew the line this year: constant and covert monitoring of employees is against the law.
On one side is the AI company saying "let everything be in the open so the agent can read it". On the other side is the Board saying "you cannot record everything". The question is not technical: for whom will the work be visible, and who will keep the right to show it?
11The one person
Today that right is with the employee, and the employee uses it. In surveys roughly a third of employees say they hide their AI use from their employer.
Among those who hide there is a special one. Every company has one or two: they use the agent better than anyone and do the work of several people alone. One day they go to their manager and say: don't pay this much for that software, I will do the same job with Claude. The subscription is cancelled. The company saves money, and the person raises their own worth.
This is another way the castle is taken from the inside. It is the third reading of the title: the one who takes the castle is neither the team nor the AI company, it is one person inside the team.
This person should not be judged lightly. Most of them see where things are going: many people will lose their jobs. They tell themselves: let me at least be the last one out. Someone who feels a duty to society can only console themselves with that sentence.
The line is thin. When their work starts to enter other people's territory, colleagues see them as a threat. If a company's moat is the way the team works together, the one person breaks exactly that. The company stops depending on a piece of software and starts depending on a person. And the team starts hiding knowledge from one another.
Hiding is rational for that person. Their bargaining power comes from what only they know. The day they write it down, that power passes to whoever owns the place where the writing sits. So a company that asks an employee to write down what they know has to give something in return. What will it give?
12The share
To answer, we first have to look at where the money goes. When the agent enters the company, the price of work changes too: payment is tied to the outcome. One customer service product charges 99 cents for every issue it resolves. Say this agent now does the work of a twenty-person support team. The invoice says how many cents it cost to resolve the issue. It does not say the wages of the twenty people who used to resolve it. Who keeps the difference?
I wrote a harsh sentence in my notes: the wage of the employee who is let go will not go to the AI company; the boss will take that share. Today's data supports half of this sentence. Very little of the gain from AI shows up in wages. But the gain to be shared is still small. So I say it with a condition: if the gain grows, under today's order most of it stays with the employer.
There are signs that the employer will keep the savings. Last year the chief executive of Salesforce said they would hire no new engineers that year.
Two calculations collide here. The AI company wants a tenth of the wage budget. The employer wants to keep what it saves with the agent. The two are pulling the same money from two ends. In The Freeloading Heir I wrote that the owners of AI say different things on every subject and fall silent together on ownership. There is a shared silence here too. Everyone talks about "value" and "outcomes". Nobody says who keeps the wage that is no longer paid.
What Marx said still holds. The power of working together was never paid for. AI did not change that; it only exposed it: the work the agent does has an invoice, and the power the team creates together still has none. A company's moat comes from that power, and that power still appears on no payroll.
For Turkey this question is not abstract. We export about 5 billion dollars of information technology a year, most of it software. What we sell is largely skilled labor. When the agent does part of that labor, where will the gain stay?
13The close
Let's go back to the meeting at the start. There is a new participant at the table, taking notes: who said what, what was decided, who took which task. When the meeting ends the notes are in front of everyone. But the note-taker does not know whether the people at the table really understood the decision.
A camel does not jump on its own; someone gets it across. The company that entered the meeting knows this, and it wants the job. The castle will be taken from the inside; by whose hand is not settled. Either the team gets its own camel across and holds the castle. Or the AI company that has moved in gets it across and sends the invoice. Or the one person inside the team takes the castle and keeps what they know to themselves. Technology will not decide which. It will be decided by who holds the notes, who keeps the right to show the work, and who keeps the wage that is no longer paid.
Salâh Birsel was right: explaining something to people is harder than getting a camel to jump a ditch. AI writes the document in a second. Explaining the document to a person is still a person's job.
To founders and investors: stop looking for the moat in the product. Code never protected you. What protected you was the way your customer and your team work together. That is exactly where the AI companies are entering today.
To employees and teams: you are the ones who provide the company's moat. Talk about what you get for it.
··Sources
- Anthropic's Claude Meetings announcement (November 12, 2026) and its essay on teams of humans and agents (June 2026).
- The fall in software stocks: Reuters (February 2026).
- The idiom and its example sentence: Turkish Language Association (TDK), Güncel Türkçe Sözlük.
- Warren Buffett, 2007 letter to shareholders.
- Customer retention and company valuations: Jamin Ball, Clouded Judgement (2024–2026).
- Gross profit in AI products: ICONIQ (July 2026). Gross profit in classic software: Jamin Ball, Clouded Judgement.
- Retention at smaller companies: ChartMogul, "The AI churn wave" (December 2025).
- Klarna: statements by the company's chief executive in 2024 and 2025.
- Salesforce inside Claude: Dreamforce 2026 announcement (secondary source). Slack data terms: 2025 (through search summaries only).
- "Sign in with ChatGPT": OpenAI developer documentation.
- App store data: Sensor Tower (July 2026).
- AI revenue calculation: The Economist (November 12, 2026; through a summary).
- Company spending on AI: Ramp AI Index (November 2026).
- The Toyota and General Motors joint plant (NUMMI): This American Life, episode 403 (2010).
- Karl Marx, Capital, volume one, the chapter on co-operation.
- Meeting time: Gallup (2026, third quarter). Code review time: Faros AI (2025). Company-level productivity: Yotzov, Barrero, Bloom, Davis and colleagues (2026).
- Eliyahu Goldratt, The Goal (1984).
- The electric motor and the factory: Paul David, "The Dynamo and the Computer" (1990).
- The one person versus two-person team experiment: Dell'Acqua and colleagues (2025).
- The gap between new and experienced users: Anthropic's research on Claude Code sessions (June 2026).
- Employees hiding AI use: Deloitte (2026) and Employment Hero (2026) surveys.
- Behavior of watched workers: Ethan Bernstein (2012).
- Turkey's Personal Data Protection Board, principle decision no. 2026/2035 (Official Gazette, October 8, 2026).
- Gains reaching wages: Humlum and Vestergaard (2025). The Salesforce chief executive's remark: the company's February 2025 earnings call.
- Turkey's information technology exports: TÜBİSAD (2026).