When Intelligence Is Free, Only Habit Pays

Walk through a company today and follow the software nobody opens.

The sales team doesn't open HubSpot. They ask the AI who's likely to close this quarter, and it answers by reading HubSpot for them.

The engineers don't open Linear. They ask what's blocking the release, and it reads the tickets they stopped writing.

The analyst doesn't open Metabase. She asks for last week's numbers in a sentence, and the dashboard she would have built goes unbuilt.

Even the wiki in Notion sits unread. The page that summarizes the project is slower than asking for the summary directly.

It sounds like progress, and in a way it is. But look at what the software is doing for its money now.

Nobody updates the CRM by hand anymore; they ask the AI to do it. Nobody reads it either; they ask the AI what's inside. So one person tells the AI to write and another asks the AI to read, and the software just sits in the middle of a loop the AI could close on its own. A tollbooth on an AI talking to itself.

And the copy in the middle was never any good.

Every sales manager has had this conversation:

Multiply that by every rep, every week, and you see what a CRM actually holds. Not what happened. A late, partial, sometimes fictional account of what happened, filtered through human memory, human laziness, and a plugin that may or may not have fired.

The CRM was never a record of reality. It was a distortion layer sitting between reality and the people who needed to see it.

For thirty years we paid people to type a worse version of their own work into software, a deal stage here, a ticket there, because computers couldn't read what they had actually done. It was a bad bargain, but the only one on offer. AI is the first real alternative: it reads the work directly, so the retyping, and the software built to collect it, lose their reason to exist.

Routing that distortion through a protocol, so an agent can recover a worse version of what it could have read first-hand, isn't integration. It's ceremony.

And the bill is about to make the whole thing obvious.

Per-seat pricing was a tax on stability. You paid the same whether your team lived in the software or never touched it. That is how these companies made money from customers who barely used them.

A seat nobody sits in can't be sold. So the vendors are switching to usage pricing, charging per action instead of per head. And usage pricing shows you what the seat hid: a bill that lists every read and write, most of them an agent fetching something a human already said.

Outcome-based pricing is a fantasy pushed by people who don't understand how the economy works: most companies lose money and most buyers never hit the result they bought your tool for, so getting paid only when they win means starving. Subscriptions and usage pricing are both fine models. The catch is that the user is no longer a human in a seat, it's an agent, and that quietly breaks the assumptions under both. They survive, but every SaaS contract is about to be reopened.

The AI labs, OpenAI, Anthropic and Google, are really two businesses wearing one name.

One is the raw intelligence, the model itself. That part is heading toward the price of electricity. Every lab lands near the same level, and open models already handle most of what business software needs.

The other is the front door, ChatGPT and Claude as the place people go to get things done. That part is expensive, because it owns the habit.

MCP is what connects the two. It lets the front door reach into your software and use it, with the cheap model as the engine. The moment you connect, you have quietly answered a question most companies never thought to ask: who does the customer talk to?

The answer decides everything, and it isn't technical. There is always one party the person speaks to directly, and that party keeps the relationship.

Right now almost everyone is getting this backwards. Software companies are racing to expose themselves through MCP, proud of the integration, not noticing that they are turning into thin layers behind someone else's front door. They win the demo and lose the customer.

Slamming the door shut is also wrong. Refuse the agents and you only go dark faster, because the customer already lives in the chat and won't come back to your login to humor you.

The real move is in between. Hand the agents the data they could rebuild on their own anyway, and charge only for the part they can't.

But know what you are selling. Anything you can charge an agent for today, the model will work out for itself tomorrow, straight from the raw data you already handed over. That revenue has a clock on it. Take it while it lasts, and don't mistake it for a moat.

The front door has the same weakness in reverse. It has no lock-in either. The same prompt runs on every model, and ChatGPT's share of AI traffic has already slid from the high eighties to the high fifties in about a year. Winning the customer and keeping the customer are not the same thing here.

So the front door takes the customer. The next question is which software it takes them away from, because not every tool is equally exposed. Some of it is pure middleman and dies first. Some only looks like one.

Two questions sort almost the entire stack.

First. Is what it holds a secondhand copy, an incomplete and out-of-date account of activity that really happened somewhere else, so an agent can read the source instead? Or is it the only authoritative copy, the thing that exists nowhere but here?

Second. Is the tool's job perception, letting a human read or coordinate? Or commitment, executing a consequential action that money, audits, or other systems depend on?

A tool that holds a secondhand copy and exists for perception is a distortion layer, the kind from the sales-manager story, and AI removes the only reason it existed. A tool that holds the only authoritative copy and exists to commit consequential actions is a ledger. Most software is a blend, and its fate is decided by the mix.

Dashboards die first. A dashboard is the purest distortion layer there is. Metabase, Looker, Tableau, none of them owns a single number. They render numbers that live in a warehouse, in a shape a human can read, and they render them late. The figure on the screen is whatever someone defined months ago, refreshed whenever the pipeline last ran. The instant the analyst asks the agent for last week's revenue instead, the dashboard is a window nobody looks through. The chart dies. The agreed definition of the metric survives, because the agent needs it to avoid inventing the figure.

Issue trackers are next. A ticket is a hand-typed summary of work that already exists in the pull requests, the commits and the threads around it. It's stale the moment it's written, because the work moved on and the ticket didn't. Linear holds almost no data a team couldn't regenerate, and a ticket rarely commits anything with consequences on its own. It's coordination, and coordination is the first thing an agent absorbs.

The CRM is propped up by a law, not a moat. The whole theater, a rep asking the AI to log a call, a manager asking the AI what happened on an account, exists for one boring reason. In many countries an employer is not allowed to let an AI read its employees' email and chat directly. So the work gets laundered through a CRM the company is allowed to read.

Regulation is a real barrier. It is also a temporary one. The honest question is not whether we need HubSpot. It is whether we need anything more than a checkbox, where the employee ticks "my employer may read my work communication" and the AI reads the source with nothing in between. I'd bet on the checkbox. What's left of the CRM after that is narrow: the deal as the number finance books, and the engine that actually sends the outreach. The reporting around it goes.

Notion is the repository nobody reads. Be honest: half of it was written to look organized and never opened again. The small irony of the agent era is that something will finally read your Notion, an AI told to go dig one fact out of it. And once an AI is doing the reading, why keep Notion at all? A wiki that writes itself, updating from whatever the team marks "official," beats one people have to maintain by hand and never revisit. The written memo survives because it is the work. The status page does not.

The ledgers survive by going invisible. An ERP, a payroll system, a payments processor holds the only authoritative copy of something with money and regulators attached, and its job is to commit, not describe. You can't reconstruct the general ledger from activity, because the activity is downstream of it, and you can't route around the thing the auditor signs. The warehouse beneath them, Snowflake or Databricks, survives in reverse: it's the source the dashboards were only borrowing from.

The sources were never middlemen at all. Slack, Gmail, GitHub, the calendar, and the raw call on Zoom or Meet. These are where the work happens, the original the middlemen kept photocopying. Note that a tool like Gong, which summarizes the call, is itself a copy of the recording, so it gets squeezed too. The source is the recording, not the summary of it. Whoever holds the source holds a far better position than the middlemen losing theirs.

AI won't kill software. It kills the dashboards and trackers, which were the part anyone actually paid for. The invisible plumbing underneath is fine.

The only question worth asking about any tool is which half of it the customer was really paying for.

When the intelligence is free, where can a company still make money? In two places, and only two.

You can own something you supply. A hard-to-copy asset or capability you bring to the table: Nvidia's chips, a bank's license to move money, a dataset only you hold, a factory that cost a billion dollars to build. This is the supply side.

Or you can own demand. The customer's attention and habit, the fact that they come to you first and don't shop around. This is the demand side.

They are not equal, and the gap between them is the point of this piece.

On the supply side, three things hold their value.

Notice what is not on that list: a company's own pile of customer data. Everyone called that their moat. It was never a moat. That data rebuilds itself from the activity underneath it. Read the calls, the emails, the calendar and the code, and you get the database back without the company that used to charge for it.

One caution, because the strong version of this is wrong. Reading the activity rebuilds maybe the first eighty percent of what a CRM or a tracker held, almost for free. The hard part is the last twenty percent: the exceptions, the approval someone has to sign, the compliance rule nobody wrote down, the edge case a veteran learned the hard way. That twenty percent is what separates a slick demo from something a company can actually run on. It buys the incumbents a few years, not a few quarters. But a few years of breathing room is not a moat. It's a runway, and runways run out.

Now the other side, the one most people underrate. The demand side is just the limited human brain: attention, memory, and the unwillingness to compare more than two or three options before settling on one.

Owning that is worth more than owning any asset, because a default lets you charge more than you are worth. Meta sells your attention by the impression. Apple takes thirty percent of everything sold through the phone, as rent on a habit.

Google isn't paid for the best answers. It's paid for being the box people type into without thinking, and the default always out-earns the better product.

Here is how the two connect, because it's the heart of the argument.

The supply-side assets used to be how companies bought habit on the cheap. Lock the customer in, make leaving painful, and the default forms by itself. Habit was a side effect of friction.

AI removes the friction. The same prompt runs everywhere and the data exports on request, so nothing traps anyone now. Habit has to be earned instead of trapped.

So the three supply-side assets get a new job. They don't make the moat directly anymore. They build trust, and trust is what earns the habit.

Determinism is the clearest case: it generates habit. A tool that gets the answer right, over and over, is a tool people stop double-checking. And a tool people stop checking is one they keep using without thinking about it. That is a quieter and stronger hold than any login screen, and it has to be earned every day.

Picture an ordinary Tuesday eight years out. A product lead at a robotics company opens no software all day.

The robots his company sells are not clever machines in the old sense. They are cameras and arms with a model behind them. Give a model eyes and let it reason about what it sees, and a machine can do the dull, physical, half-skilled work that used to need a person: pick the order, inspect the weld, restock the shelf, move the pallet.

By 2034 this is no niche. Machines that see and act have moved into warehouses, building sites, hospitals and farms, and they run on the same cheap models that emptied the office software. The intelligence that cleared the dashboards is the intelligence that gave the warehouse hands.

So his own day looks like everyone else's now.

He talks to one agent, and it has already read everything worth reading: the fleet logs, the support threads, the code, the calendar. There is no dashboard for robot uptime, no ticket board for the field team. When he asks why the night shift slowed down, the agent reads the raw telemetry and answers in a sentence, then throws the report away. The next question will need a different shape.

Underneath that sentence, something less casual is happening. The things that carry consequences still land somewhere exact. The robot that dropped a part files a safety record an auditor will read in two years. The delivery that completed bills a customer. The firmware pushed to ten thousand machines is signed and logged. These commit to a few quiet ledgers that nobody opens and everybody trusts.

That is what's left of the software industry on this Tuesday. Not the dashboards, which are gone, but the quiet machinery beneath them, billed by the action instead of the seat, and invisible because it works.

The split is the same one that ran through the office. A dashboard only ever described the world, so it died. A robot changes the world, so it pays.

He stopped checking the agent's work a while ago. He couldn't tell you the week he started trusting it, which is the point. There is no contract holding him and no login he would miss. There is the trust, earned slowly, and the fact that going elsewhere would mean building that trust again from scratch.

At seven he closes the laptop, and the same logic runs his evening. His personal agent has already booked the flights, moved the dentist, and ordered the part his bike needed. A delivery robot, a cousin of the ones he builds, leaves it at the door before dinner. He visited no website to arrange any of it. The brands never met him. They met his agent, arriving as structured offers it could read in a millisecond, and they won or lost on price and fit in a moment he never saw.

Step back and the economy has narrowed to a single scarce thing. Not intelligence, which is everywhere and nearly free now, in the chat and in the machines alike. Not data, which leaks and reconstructs. The scarce thing is his attention, and more exactly his unwillingness to spend it twice.

Whoever owns the agent he talks to owns what he buys. Whoever owns the trust he no longer questions owns how he works. Everyone else keeps the ledgers nobody opens, builds the machines that do the work, or sells intelligence at cost.