When the Terms Change
The relationship between practitioners and their AI tools has shifted quietly over the past two years. Early on, the question was simply whether a given tool could do the job. Now the question is more nuanced: what happens when the provider changes the arrangement?
The Cursor situation made this concrete. OpenAI models are scheduled to stop working inside Cursor on November 12 — not because users did anything wrong, but because a corporate acquisition changed the terms. The models still exist. The app still exists. But the combination no longer works, and anyone who built workflows around it has a problem they did nothing to cause.
The Model Is Not the Asset
The instinct is to worry about losing access to a specific model. That worry is mostly misplaced. Frontier models are increasingly interchangeable; a good one can substitute for another on most tasks within a session or two of calibration.
The harder thing to replace is working context — the instructions, decisions, corrections, and project history that accumulate inside a provider over months. That is what makes a tool genuinely useful for a specific person’s specific work. An export button satisfies data portability in a technical sense, but a bulk archive of old conversations does almost nothing for work continuity. The context that matters is not the transcript; it is the accumulated understanding that shaped what came next.
Products become more useful as they learn about a practitioner’s work, and the same features that create value also create switching cost. A price increase is visible and can be planned for. A missing project history is only discovered when someone tries to leave.
Auditing the Dependency
A useful frame is “concentration with an exit”: use one provider deeply for each distinct job, but keep the durable parts of the work — files, decisions, project state — outside any single provider so that a switch is painful but not catastrophic.
The practical tool for this is a provider independence audit: a structured test run against a rival AI to measure actual switching cost. The exercise goes like this:
1. Pick a real project with history.
2. Write a project note outside the provider — capturing current state in plain language.
3. Hand that note to a rival model.
4. Watch where the rival fails.
5. Trace each failure back to its cause.
If the rival fails because a key decision only exists in an old chat, or because a source file only lives as a provider upload, that is a fixable dependency. The goal is not a perfect handoff — the goal is for the rival to understand current state well enough to take a useful next step.
The Seven-Question Project Note
Most practitioners who try this exercise discover the same gap: source files are present, but project decisions are missing. A folder of code or documents is not enough to continue a project. What is missing is the reasoning behind the current state — why things are the way they are.
A lightweight format that addresses this is a seven-question project note: objective, current state, sources, decisions made and why, constraints, definition of done, and next action. Kept outside any provider — in ordinary files, in version control, wherever the practitioner controls access — this note becomes the portable layer that makes any AI substitutable.
Owned Versus Rented
The underlying distinction is between material that must survive and intelligence that can be rented. Compute, model capability, and application features are all rentable — interchangeable enough that losing one provider is a manageable inconvenience. Working context, decisions, and project continuity are owned assets; losing them is a disruption that compounds over time.
As AI reaches further into daily operations, this distinction becomes more consequential. The deeper the integration, the more important it becomes to keep work understandable outside the AI — not as a hedge against any specific provider, but as a practice that makes the work itself more durable and transferable to any collaborator, human or machine.
The five-prompt audit is roughly one hour of friction now against a potentially much larger recovery cost later. That trade is worth making before the next corporate deal changes the terms again.
Text summarized and optimized using Anthropic’s models and reviewed by a human.