Software is dead.

On February 3rd this year, Thomson Reuters had the worst day in its history. The stock fell about 16 percent in a single session. LegalZoom fell almost 20. By one count, software companies lost around $285 billion in market value that day.

The trigger was a release from Anthropic: a set of plugins for its Claude Cowork agent, including one for legal work.

Here's what the legal plugin actually was.

A folder of text files. About 2,500 lines of structured instructions in plain language. There was no new model in it, and hardly any code.

A folder of text files, and one of the biggest legal information companies on earth had its worst day ever.

Six weeks later, in Letter 75, I wrote that a company in 2030 might run on text files. It sounded provocative at the time. In hindsight, the market had already said it.

 

Everybody is now saying "software is dead." I agree, and I've thought about it a lot, because almost every AI strategy I build for a client this year contains a new line: go through your software subscriptions, one by one, and ask which ones still deserve to exist.

But "software is dead" can mean three very different things, and only one of them is true. So let's be precise.

What "software is dead" actually means

Meaning one: software disappears from our lives. That's false. We'll have more software than ever. Building a small piece of software is becoming like building a slide deck or an Excel sheet. In a couple of years, a lot of people will do it in a couple of hours without thinking of themselves as programmers.

Meaning two: everyone rebuilds the software they pay for. Mostly false, and this is the story you see in viral videos. Someone "rebuilds Photoshop with Claude in a weekend," and the clip gets millions of views. The people who actually test these clones keep finding the same thing: most of them fall apart the moment you use them for real work. Rebuilding thirty years of a mature product is not the threat.

Meaning three: selling software is dead. This is the one I believe. The business of charging people, every month, for a piece of software is in deep trouble. At least in the form we've known for twenty years.

So when I say "software is dead," that's what I mean. Software will be everywhere. Paying for it is what's dying.

To see why, it helps to ask a question we rarely ask: why did we ever pay for software in the first place?

Why we paid for software

I think there were four reasons.

Building was expensive. Engineers were scarce, projects took months, and maintenance never ended.

The interface was the only door to the data. If you wanted your customer history, you went through the CRM screen.

The software came with a process. When you bought a serious tool, you were also buying a way of working: how a sales pipeline moves, how an approval flows, how the books close. Someone had written down the best practice, and the software enforced it.

Humans were the ones doing the work. So vendors could charge per person, per seat, forever.

AI attacks three of those four, and that's why this time is different. The three forces below each break one of them. The one left, the process, is where it gets interesting, and I'll come back to it.

Force one: building got cheap

Every company has always faced the same choice for any tool: build or buy.

For two decades, the answer was almost always buy. Building meant hiring engineers, waiting months, and maintaining the thing forever. Buying meant a monthly fee and someone else's problem.

That math has flipped for a whole category of tools. When I sit in build-or-buy discussions with clients now, the conversation leans toward build before anyone opens a spreadsheet, because building costs a fraction of what it used to.

It shows in the data. McKinsey's latest State of AI survey found that about a third of respondents said their organizations had already decided against buying at least one software product or feature, because it could be built in-house with AI coding agents. Retool, which surveys the people building internal tools, found the same thing and looked at what gets replaced first: workflow automations, internal admin tools, dashboards.

Look at that list. These are the simple tools: a form, a table and a few rules. The kind of thing that probably never deserved a monthly fee in the first place.

So yes, this is happening. But honestly, it's the smaller force. It only kills the first reason we paid for software: building was expensive.

(I don't understand people who want to get rich by vibe coding software. If you can do it easily, so can other people... so where's the value? People apply old intuitions to a new paradigm. Software WAS valuable because it WAS hard to build. That's not true in a world where building is easy. Unless what you're really selling is something else: an audience, data, trust, security.)

Force two: you don't need the software at all

This is the one the "SaaS apocalypse" conversation keeps missing, and it's the one the market reacted to on February 3rd.

Look honestly at most business software. I made this point in Letter 75. A CRM is customer data, notes, stages and tasks. An HR tool is records, policies and approvals. A project tool is tasks, owners and dates.

Most business software does two things: it stores information, and it helps people find and update that information. It's a database with a nice outfit on.

For years, the outfit was the product. The interface was the only practical way to get at the data and use it in a meaningful way, so you paid for the interface.

Now an agent connected to the right context can answer the question directly. "Who do I need to follow up with today?" can be answered from your customer context, with the right permissions, by an agent that can read it. No dashboard required.

And I'm not theorizing here. Since I learned to work properly with an agent on top of a folder of text files (that's my AI OS), I've cancelled Webflow, Pipedrive, Zapier, n8n, Gamma and Adobe. I haven't missed any of them.

These days, I only pay for software that gives me something a folder of text and an agent can't.

Now, the part that really breaks the business model. In the old world, a new feature was a roadmap item, a sprint and a release three months later. In the new world, a new feature is a skill: a paragraph describing what you want, saved in a file. That legal plugin was exactly this. Contract review, NDA triage, compliance checks, each one a text file describing how the work should be done.

You can't sell a feature to someone who can write it in ten minutes. That breaks the second reason we paid: the interface is no longer the only door to the data.

So the real threat to a software company is a customer who simply stops needing it.

Force three: the seat is dying

There's a third force, quieter, and it hits even the software that survives.

Most software is priced per seat. Per human who logs in.

Agents don't log in. When one person with three agents does the work that used to need ten people in the tool, the vendor loses nine seats, and the customer gets more done.

That's why investors panicked even about companies nobody expects to disappear. The product might survive. The way it gets paid probably won't. For example, if my team used Salesforce only through agents (we don't), we could in principle connect all our agents through one account instead of five.

That's the last reason we paid, and it's breaking... in slow motion.

So what survives?

Not everything goes. Some software keeps its value, and it's worth being precise about which, because that's where your money and attention should go.

The system of record. Where the truth lives: payments, ledgers, invoices, inventory, audit trails. You do not want "vibes" managing those. They need to give the same answer every single time, and they need to be auditable. (My own research on AI consistency is a long argument for why a model reading a text file can't be your ledger.)

What can't be copied. Data nobody else has. Networks where users make the product more valuable for each other. Licenses and regulation. Anything with a physical side.

Trust. Someone accountable when it breaks, at 3am, with a contract that says so.

Look at that list: data businesses, trust businesses and infrastructure businesses that happen to ship software. Whoever holds the context and the record keeps the customer. The interface on top is becoming free.

(It's the same pattern as Letter 103. The visible layer turns into a commodity, and the value moves underneath.)

Then there's the third reason we paid: the process.

This is the one I find most interesting, because it cuts both ways. A skill file can encode your process, written your way, for your business. That's more powerful than any vendor's generic best practice. But it only works if you can actually write your process down.

Most companies can't. As I wrote in Letter 78, they run on organizational folklore: unwritten rules in the heads of four people. Those companies will keep paying vendors for a long time, because the vendor's built-in process is the only discipline they have.

If your company can't write down how it works, it will keep paying someone who already did.

The part nobody wants to say about personal software

Now the other side, because I see this every week.

When people build software for themselves, it's not production-grade. They don't understand the code. There are security holes nobody looked for. The agent that wrote the tool often has access to the data and credentials it touches. And when the person who built it leaves, nobody knows how it works.

Inside a company, that's shadow IT on steroids. This is exactly why IT and security teams are cautious about it, and they're right to be. Building got cheap. Owning, for now, did not.

Honestly, I think we'll mostly end up trusting the system, the way we trust a spreadsheet formula without reading it, and the AI will get good enough that this stops being a daily problem. Most of the software we use will be built for one job, used, and thrown away, like the slides after Monday's meeting.

We're not there yet. Today, the moment someone else depends on your little tool, it's production software, whether you planned it or not.

What I'd do on Monday

Audit your subscriptions with one question. For each tool you pay for, ask: does it hold something I can't move, like the record, the data, the compliance, the network? Or is it an interface on information I could give to an agent? The first kind is worth paying for. The second kind deserves a hard look at renewal time.

Before you build anything, write it down. Describe the tool you think you need in a page of plain text: what it should know, what it should do, what good looks like. Half the time, you'll realize you don't need software at all. You need that file, your context, and an agent. That's the whole idea behind the AI OS, and it's what I teach in the AI OS Course.

If something you built gets used by others, treat it like production. Give it an owner, get it a security review, and bring IT into the loop early. Their job is to make it survive.

Back to February 3rd

A folder of text files wiped hundreds of billions off software companies in one day.

The files didn't do anything magical. They described, in plain language, how a type of work gets done. And the market suddenly understood that a well-written description plus an agent can replace a lot of what companies were paying for.

That's the shift. Software is turning into something you describe instead of something you buy.

And if the program is a text file, then anyone who can describe their work precisely is now, in a sense, a programmer. That's going to matter a lot more than knowing how to code.

Whoever holds the context holds the customer.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)