The Great Collapse

Almost every company we meet asks us the same question. It tends to come up early, from someone senior, in a slightly embarrassed tone, as though the answer ought to be obvious.

Why can't we just use Claude for this?

The "this" changes from company to company. Sometimes it's drafting proposals. Sometimes it's reviewing contracts, chasing invoices, or researching accounts before a sales call. But the shape of the question is always the same. They pay for a piece of software that does one narrow job with AI. They also have Claude or GPT open in another tab. And they've started to suspect the second could do the work of the first.

A year ago the honest answer was usually no. The models were impressive in a demo and frustrating in practice. They forgot what you'd told them. They couldn't see your systems. They couldn't do anything on your behalf except produce text for you to copy somewhere else.

A whole generation of software companies grew up in that gap. Each one took a model, wrapped it in an interface designed for a single job, connected it to a system or two, and sold it by the seat. One tool for sales emails. Another for meeting notes. Another for contract review. Another for support tickets. If you've sat through a software budget meeting lately, you know how long the list has become.

Those companies were a reasonable response to the models of the time. We think most of them are going to disappear.

We call this the Great Collapse. We don't mean a crash. We mean a collapse in the sense of many things folding into one: the dozens of AI products a company buys today folding into the one or two general models it already pays for.

The reasoning is simple. A wrapper is worth something for exactly as long as the model underneath can't do the job by itself. The product is the distance between what the raw model can do and what the business needs done. And that distance shrinks with every release. The models learn to remember. They learn to read your files, to connect to the systems where the work lives, to carry a long task through without being supervised. Each of those used to be somebody's startup.

It's a hard place to build a company. Your supplier is also your competitor. It improves faster than you do, and each improvement shows up, at no extra cost, in a product your customer already has open.

We've watched something like this happen before. There was a time when doing anything on a computer meant installing a program for it: one for mail, one for documents, one for your calendar, one for chat. Then the browser got good enough, and one by one those programs turned into tabs. Nobody held a meeting to adopt the browser. People just noticed one day that they opened Chrome in the morning and didn't leave it until they went home.

We think you're about to live in Claude or GPT the way you now live in Chrome. You'll open it in the morning and stay there. It will know your business: your customers, your documents, your numbers, the way you like things written. Most of the work you now spread across a dozen applications, you'll do by asking.

We'd go a step further. A surprising amount of what happens in email and Slack is one person asking another for something that a system with full context could simply hand over. As the model companies build ways for teams to work together inside their products, we expect more of that conversation to move to where the work already is, and inboxes and channels to get quieter.

As of October 2026 this is not yet how most companies work. But things are moving faster than at any point we can remember, and this is where the puck is going.

Which brings us back to the question. If the answer to "why can't we use Claude for this?" is now usually "you can," why does everyone keep asking?

Because "you can" comes with conditions. You can if your team has been trained to work this way. You can if the model is connected to the right tools and the right data. And you can if people have the right mindset, where handing work to the model is the first move rather than an occasional experiment. Training, tools, and mindset. Most companies have bought the license and have none of the three.

That's the strange thing about this moment. The capability arrived faster than anyone's ability to absorb it. A company can be paying for the most capable general-purpose tool it has ever owned and using it as a slightly better search box. The gap that used to sit between what the model could do and what the business needed has moved. It now sits between what the model can do and what the people using it know to ask for.

You can't close that gap by buying more software. You close it with implementation: setting the thing up properly, wiring it into your systems, building the pieces your team will reuse, and working alongside people until it becomes the default way work gets done.

That is why we started 440. We have high conviction about where this is going, and we kept meeting companies that would get enormous leverage from Claude and GPT but hadn't gone down the rabbit hole yet, and didn't have a spare quarter to find their way by trial and error. We help them get there sooner, and we do it with the tools they already pay for.

The companies that do well over the next few years won't be the ones that bought the most AI software. They'll be the ones that learned to live in the one that mattered.