The vendor exit assessment every AI contract should be born with.
What goes in the unwind clause. Why most contracts skip it. The four lines that compound across a portfolio.
Every AI contract is signed at the moment of maximum optimism and maximum leverage — yours. The vendor wants the deal; you want the capability. The one clause almost no one negotiates at that moment is the one that matters most eighteen months later: how you leave.
The unwind clause is not a termination clause. Termination ends the contract; the unwind governs what you keep — your prompts, your fine-tuning data, your evaluation sets, the embeddings you paid to compute, and the logs that constitute your audit trail. If those live only in the vendor's tenant, you do not own your operating layer. You rent it, and the rent is your switching cost.
Four lines belong in every AI contract from birth. Data egress: your data and derived artifacts exportable in a usable format on demand. Model portability: the right to take fine-tuned weights or, failing that, the training set that produced them. Continuity: a defined window where the service runs while you migrate. And evidence: the audit logs remain yours and exportable, because a regulator's question does not pause for a vendor transition.
The reason most contracts skip these is that they read like distrust at signing. They are the opposite. They are the clauses that let you adopt aggressively, because adoption without an exit is lock-in dressed as progress. The exit assessment is what makes the buy decision reversible, and a reversible decision is one a board can approve quickly.
The cost of omission compounds. One vendor without an unwind clause is a manageable risk. Six of them, each holding a slice of your model layer, is a portfolio you cannot restructure without rebuilding it. The exit assessment is four lines per contract. The alternative is a migration project you will eventually run under duress.