A story went around this spring that most people read the wrong way.
A company opened for business one morning and found that every one of its accounts at one of the major AI providers had been shut off overnight. No warning. No degraded service first. An automated system had flagged something, a review team had agreed with it, and the email said so in two paragraphs that named no specifics. The only way to appeal was a web form.
The company was legitimate. The work was legitimate. Dozens of people sat down at their desks and nothing worked.
Most of the commentary treated it as a story about that vendor. It is not. Every major provider has an automated abuse system that can misfire, a usage policy that can change without notice, and a business that can pivot while you sleep. The story is not about who did it. The story is about what the company had left when it happened.
The three ways one provider takes you down
The ban is the dramatic version, and it is the rarest. Two quieter versions happen to somebody every month.
The pivot. The price doubles at renewal. The capability your process depends on moves to a higher tier. The feature you built around gets retired with ninety days' notice, which is not enough. Nobody banned you. You were priced or deprecated out, and the effect on your Tuesday morning is the same.
The outage. Every provider has had them and every provider will have more. When all of your AI runs through one pipe, an outage does not slow you down. It stops you, and it stops everyone at once, including the people who were in the middle of something for a customer.
Organizations with more than one path treat all three as an inconvenience. Organizations with one path treat them as a crisis, and they find out which kind they are on the morning it happens.
What actually disappeared
Here is the part worth sitting with, because it is where the real cost lives.
When those accounts went dark, the model did not disappear. Models are interchangeable. There were four other providers open for business that morning and any of them could have answered the same questions.
What disappeared was everything the company had put into the tool. Six months of chat histories. The prompts that worked, refined over dozens of attempts and saved nowhere else. The documents people had uploaded to get answers. The projects, the custom instructions, the shared threads a team had been using as its working memory.
None of that was the provider's fault. The company chose to keep its AI knowledge in a rented room, and the landlord changed the locks. Most organizations we walk into are doing exactly the same thing, and they do not know it, because nobody has ever asked where the work goes after the answer arrives.
The Tuesday test
Ask this question of your own organization and answer it honestly: if your primary provider disappeared on a Tuesday, how long until you are running on a different one?
The answer depends on three things, and you either own them or you do not.
Your people's skills. If your team was trained on one interface, they know one interface. If they were trained to write a prompt that gets usable output, to recognize the tasks a model will quietly get wrong, and to check something before it reaches a customer, they can sit down in front of a different model that afternoon. Skills are model-neutral or they are product familiarity, and you find out which on the day you switch.
Your workflows. A workflow that has been described end to end, with its inputs, decision points, exceptions, and handoffs, can be rebuilt on anything. A workflow that exists as a chain of prompts inside one vendor's product is not a workflow. It is a dependency, and it goes when the vendor goes.
Your knowledge. The answers your organization has already produced, the documents that matter, the decisions and the reasons behind them. If they live in one governed place you control, structured so any model can read them, then a new provider is useful the hour it is connected. If they live in chat histories, you start from zero.
What to do this quarter
None of this requires switching providers. It requires knowing what you would take with you.
Find out where the AI-created knowledge lives. Ask each team what they have built inside the tool over the last six months: prompts, projects, uploaded documents, saved threads. The answer is usually more than anyone expected and it is almost entirely in personal accounts.
Get the prompts out. Pick the teams whose work depends on AI most and have them move the prompts that work into a shared document with a note on what each one is for. This is the first time that knowledge will exist outside a vendor's product.
Decide what knowledge matters, and give it a home you own. Not everything. The policies, the contracts, the answers people ask for every week, the decisions that keep getting relitigated because nobody can find the last one. Those belong in one place with an owner, a review cycle, and a structure that works for retrieval, so that whichever model you point at it, it can answer.
Run the Tuesday test once a year. Actually try it. Point a team at a different provider for a day and see what breaks. What breaks is what you do not own.
What this does not fix
None of this makes a switch instant. A well-prepared organization loses a week to a provider change, not a quarter, and the week is mostly logistics. That is the honest goal.
It also does not make you model-neutral by itself. It makes you portable, which is the prerequisite. The organization that owns its skills, its workflow definitions, and its knowledge can decide to run on one provider, or three, or whichever is best this year. The organization that does not own them has already decided, and did not notice.
Where to go next: if the answer to "where does our knowledge live" was "in chat histories and eleven shared drives," a Source of Truth engagement consolidates it into one governed knowledge base, structured so both people and any AI model can use it, in six to ten weeks.
The platform view: our product team wrote about the same incident from the architecture side in One Email. All Gone.