Someone decided to move to a cheaper model after a price change. They tried it for a few days. The model was fine. The outputs were comparable and the interface was familiar enough.
Three days later they were back, and not because the model was worse. They came back because they got tired of re-explaining the same four things.
That is what switching cost looks like in practice. Not the subscription, not the interface, not the habit. The background.
Nobody sits down and decides to build up two years of working context inside a chat product. It arrives one conversation at a time.
You explain a client constraint once. You correct the tone twice. You spell out what your industry means by a term the model kept getting wrong. You paste in the house style guide, then paste in the exceptions to it a month later. Each of these takes forty seconds and none of them feels like it is creating an asset.
By conversation twenty, the tool is noticeably better than it was at conversation one, and the improvement is not in the model. It is in what the model has been told.
That difference lives inside one account, in one vendor's format, in a form you have never seen and cannot export as a document. It became load-bearing without ever becoming visible.
This takes about ten minutes and produces a written answer.
Open the AI tool you use most. Assume it becomes unavailable tomorrow. Pick whichever version of unavailable you find most plausible: shut down, priced out of reach, capped below what your work needs, deprecated in favour of a successor, or your account closed with no appeal.
Now write down what you would have to reconstruct before you could work at the same quality somewhere else.
Not the chat history. The background. Be specific:
Write actual items, not categories. "Client constraints" is not an item. "Halcyon cannot make efficacy claims without Legal sign-off" is an item.
Go through the list once and mark each item with one of two labels.
Written down somewhere. It exists in a document, a note, a wiki, a file you could open right now on a machine you control.
Only the model knows. It exists because you said it in a chat window at some point and the system retained some version of it.
Most people find the second list is longer, and are surprised by that, because none of it ever felt like something to file.
The second list is your actual dependency. Everything on it is held in a place you do not control, in a form you cannot inspect, at a level of fidelity you cannot verify. You do not know whether it survived the last summarization pass, and you will not find out until an answer comes back as though the constraint had never been mentioned.
Take the "only the model knows" list and mark each item cheap or expensive to rebuild.
Cheap is a fact you can state in one sentence and would remember unprompted. Your role, your company, the two projects you are on.
Expensive is anything you arrived at through iteration. The style rule that took four rounds to land. The distinction between two terms that the model kept collapsing. The thing you had to correct three times before it stuck. You cannot reconstruct those from memory because you do not remember them as rules. You remember them as the absence of a problem you stopped having.
The expensive column is the real number. It is what a provider change actually costs you, and it is why the person in the opening went back after three days.
This audit is worth doing regardless of which tools you use or whether you change anything afterward. It is not an argument for switching and it is not an argument against any particular product.
It is a way of making an invisible dependency legible, which is the only condition under which you can make a decision about it. A dependency you can see is a choice. A dependency you cannot see is just a fact about your situation that you will discover at an inconvenient moment.
Most people who run this find one or two items that genuinely alarm them and a longer list that turns out not to matter. Both results are useful. The second one is arguably more useful, because it tells you the exposure is smaller than it felt.
If the expensive column is short, you are fine. Note the items somewhere and get on with your work.
If it is long, the fix is not a different provider. Moving accumulated context from one company's account to another company's account changes whose terms govern it and nothing else.
The fix is that the context exists as a record you hold, at the item level, in a form you can read. Then a provider change is a paste rather than a rebuild, and the audit next quarter takes two minutes instead of ten because the answer is already written down.
Briefed exists to be that record. It runs on your computer, makes zero network calls, and stores the vault encrypted with AES-256-GCM. But the audit works without it, and the ten minutes are worth spending either way.
The Prompt Pass is a free monthly pack of context prompts, built as judgment frameworks rather than tricks. It drops straight into Briefed, or into whatever you already use.