If your favorite AI model disappeared tomorrow, would you lose a week of convenience, or would you lose your business’s memory? Below is the full script of this episode as an article.
On October twentieth, I turned off every scheduled task Claude was running for my business.
Claude had been my primary AI. It ran recurring work on a schedule. It had a folder of its own in my files, called Assets slash Claude, and in that folder lived a file called CLAUDE dot M-D. If you haven’t built anything like this, that file is a set of standing instructions the tool reads every time it starts. Over time, operating knowledge about my business ended up in there.
This month I moved off Claude entirely. I’m going to skip the reasons, because they would make this an episode about one vendor, and this is an episode about every vendor. What I want to tell you about is what the leaving showed me.
The argument
Here’s what I’ll argue, and you’re welcome to disagree with it. Your favorite AI model should be replaceable in a week, and getting there means deliberately giving up some of the convenience of going deep with one tool. Most advice runs the other way. Pick a platform, learn it well, let it learn you. I think that advice quietly turns a supplier into a member of the family.
Overfunctioning, and what it has to do with AI
Let me explain what I mean by family, because this is where my training kicks in.
I’m trained in organizational psychology and family therapy. One of the most useful ideas from family systems work comes from Murray Bowen, and it’s a pattern called overfunctioning and underfunctioning. In a lot of families, one person picks up more and more of the load. They remember the appointments and know where the paperwork is. And the rest of the family, very reasonably, lets them. Nobody planned it. It just accumulated.
The overfunctioner usually looks like the healthiest person in the house. The cost shows up the day they step away. Suddenly the family discovers it doesn’t know how its own life works. The knowledge was never held by the family. It was held by one member, and everybody else was organized around that member.
That’s what I found when I went to turn Claude off.
My recurring work had been organized around one tool. Some of what my business knew about itself was written in a file named after that tool, sitting in a folder named after that tool, in a format designed for that tool to read. The system was working. It was working the way an overfunctioning family member works, right up until you need to function without them.
Why you’re not off the hook
Now, you might hear that and think, that’s a you problem, Aaron. You built a lot of automation. Most people are just chatting with an assistant.
Here’s why I don’t think you’re off the hook.
The model you depend on is going to change underneath you whether you leave it or not. In January of twenty twenty-six, OpenAI announced it was retiring GPT-four-oh from ChatGPT, along with several other models, effective February thirteenth. They said only zero point one percent of users were still choosing it each day. Against a user base TechCrunch put at eight hundred million a week, that’s roughly eight hundred thousand people who had built some piece of their work or their life around one model’s way of responding. That retirement had already been attempted once, the previous August, when GPT-five launched. The backlash was strong enough that OpenAI brought it back for paid users. Six months later, it was gone anyway.
And retirement is the visible version. The invisible version is drift. In December of twenty twenty-five, a group of researchers published a method for detecting when an AI provider quietly changes the model behind an endpoint. They monitored one hundred eighty-nine endpoints from ten providers, every hour, for more than four months. They flagged thirty-seven suspected changes across twenty-nine of those endpoints. The name on the label stayed put while the model behind it moved.
So even if you never leave, the thing you’re attached to leaves you, a little at a time.
What I did instead
Here’s what I did, and what I’d suggest you do.
I didn’t migrate Claude’s files. I left them exactly where they were, as an archive. Moving the knowledge from one tool’s folder into another tool’s folder would just recreate the same arrangement with a new member at the center.
Instead I set a rule. The knowledge base has to be platform-agnostic. Any model should be able to walk in the door and pick up context.
In practice, that means every client and every project has two plain-language documents. A Brief and a Decisions doc. The Brief says what this is, who it’s for, what we’re trying to do, and where things stand. The Decisions doc records what we decided and why. Those two documents are the single entry point. They live in open formats, the kind any person or any program can read without a particular app.
And then the strict part. Nothing a person or another AI needs may live only inside one tool’s folder, or inside a tool-specific file. If a tool insists on its own instruction file, the way Claude wants a CLAUDE dot M-D, that file is allowed to contain one thing. A pointer back to the Brief.
That last rule is the one I’d underline for you. Tool-specific files are where knowledge goes to get trapped, because they’re convenient. You tell the assistant something once, it remembers, and you never write it anywhere else. Every one of those moments is a small bit of overfunctioning you just handed to a supplier.
Differentiation
Back to family systems for a second, because the fix there maps closely onto the fix here.
In a family where one person overfunctions, the work is for that person to step back a little and for everyone else to step up, so the family holds its own functions. Bowen’s word for the healthy end of that is differentiation. Each member stays connected, and the system keeps working when any one of them steps back. A differentiated family can lose a member to a new job across the country and grieve it without forgetting how to pay the electric bill.
A differentiated business can use an AI model heavily, even prefer one, and still swap it out in a week. The model does real work. It just doesn’t hold the memory.
And to be fair to the other side of my argument, going deep with one tool does buy you something. Platform-only features, memory that builds up over months, less setup. Spreading work across agents gives some of that up. Writing things down in a neutral place is slower than letting the assistant remember. I think that cost is the premium on an insurance policy, and I think most businesses are currently uninsured.
What you can do this month
So here’s what you can do this month.
First, run a replacement drill on paper. Pick the AI tool you lean on most. Write down what knowledge would vanish if it disappeared on Friday. Custom instructions, saved memories, project files, anything that exists only inside that product. That list is your exposure.
Second, pick your single most important client, project, or line of work, and write the two documents. A Brief and a Decisions doc. Plain language, in a format you could open in any text editor. If you can’t write the Brief without opening the AI tool to remember what’s going on, you’ve just found where your business keeps its memory.
Third, go into your AI tool’s custom instructions or project settings, and start moving anything substantive out into those documents. Leave behind a line that says, read the Brief first.
Fourth, once you have a Brief, test it. Open a different model than the one you usually use, give it the Brief and the Decisions doc, and ask it to tell you where the project stands and what it would do next. If the answer is in the neighborhood of what your usual tool would say, your knowledge is portable. If it’s lost, the gap tells you exactly what you still haven’t written down.
And one more, for anyone running automations. List your scheduled jobs and mark which ones only work because of one vendor’s specific feature. You don’t have to change them yet. Just know which ones they are.
The test
When I turned off those tasks on the twentieth, the archive stayed right where it was. I can still open it. But nothing I need lives only there anymore, and the next model I bring in will start by reading a Brief that I wrote, in words a person can read too.
That’s the test I’d hold any AI setup to. If your favorite model vanished tomorrow, would you lose a week, or would you lose your memory?
—
This show extends the book AI Empowered: The Psychology of Extraordinary Human-AI Collaboration. Read the first chapter free at https://www.aiempoweredbook.com, or get the book on Amazon.
Aaron Douglas runs Auspicious, a fractional marketing practice.
Blog: https://auspicious.llc/insights/
X: https://x.com/aaron_douglas
LinkedIn: https://www.linkedin.com/in/aaronddouglas/
