Constitutional reversibility: three cases and a matrix

The moment of signing

No one signs an AI contract thinking they are making an irreversible decision. What gets signed, at the moment of signing, always looks like an execution decision: choosing a provider, buying a capability, solving an operational problem with the tool at hand. Irreversibility doesn't appear in the clause that gets read carefully — it appears in the one no one discussed because the deal was already practically closed, or in the absence of a clause that should have been there and wasn't, because no one with enough weight demanded it in time.

The three cases that follow are not audit files or reporting on identifiable companies. They are three composite archetypes built from contractual patterns that appear again and again in the region whenever an organization adopts AI from an external provider — patterns repeated often enough to be worth naming even though none corresponds to a single, verifiable contract. What matters is not whether one particular case happened exactly this way, but that the structure repeats often enough that any reader will recognize at least one of the three in their own organization.

Three archetypes

The first is a mid-sized insurer that in 2023 contracted a language model to automate initial claims classification. Before signing, the legal team demanded three things that at the time seemed excessive for the size of the contract: that the data and embeddings generated from the insurer's own information be documented in an exportable format, that any fine-tuning of the model be the property of the insurer and not the provider, and a minimum notice period of one hundred eighty days if the provider discontinued the service or changed its terms. The provider agreed because the contract wasn't its strategic priority that quarter. In 2026, when the insurer evaluated migrating to a cheaper model from another provider, the migration took six weeks and cost a fraction of the function's annual budget. Reversibility didn't prevent the change of provider — it made it possible without drama.

The second is a ministry in a mid-sized country that in 2022 signed, under the pressure of a political deadline — an announcement already made, an event already scheduled — a contract to digitize a citizen procedure using generative AI. The technical team wanted portability clauses; communications needed the system working in eight weeks for the announcement. The deadline won. The contract was signed with citizens' data in a proprietary format with no export specification, and with the model fine-tuned on the volume of processed transactions registered as the provider's property by contractual default — a standard clause no one struck out because no one read it with the reversibility problem in mind. There was no bad faith: there was urgency, and urgency doesn't negotiate clauses that don't seem urgent at the moment of signing. In 2026 the ministry wants to migrate the procedure to a domestic provider for data sovereignty reasons. The migration estimate runs around two years, more than it would have cost to build the system from scratch with another provider back in 2022.

The third is a regional retail chain that in 2024 contracted a dominant provider — one of the two or three global players with enough scale for the service level it needed — for personalization and demand forecasting. Here there was no negligence: the legal team did demand portability. The provider rejected it, and the chain had no real alternative: competitors capable of matching the service offered equivalent or worse terms, and not adopting the system meant falling a step behind in the same season. They signed knowing that the model trained on two years of transactional data would remain, contractually, in the provider's domain. It's the hardest of the three cases because there's no error to correct with internal discipline — there's a power asymmetry that no well-drafted clause resolves if the counterparty has no incentive to accept it.

The matrix

The three cases share the same underlying question but arrive at it by different paths: negligence in one case, urgency in the second, power asymmetry in the third. That matters because the remedy isn't the same. Negligence is corrected with more rigorous legal review. Urgency is corrected with a process that separates the communications deadline from the signing deadline. Power asymmetry isn't corrected with more internal discipline — it's corrected, if at all, by negotiating as a bloc with other organizations in the same situation, or by consciously accepting the risk and compensating for it with a more expensive but explicit exit clause, rather than leaving it absent.

The matrix that emerges from these three cases crosses two axes that are almost never evaluated together before signing. The first is data portability: if the organization has to leave tomorrow, does its own data come out in an open, usable format, or does it stay trapped in a proprietary schema that has to be rebuilt? The second is capability portability: is the model, the fine-tune, or the embedding trained on the organization's own data contractually owned by the organization, or by the provider?

Crossing these two axes produces four quadrants. Portable data and owned capability is the insurer's quadrant — real reversibility, low exit cost in any given year. Trapped data and provider-owned capability is the quadrant of the ministry and the retail chain — de facto irreversibility, disguised as an execution decision at the moment of signing. The two mixed quadrants are the most common and the most deceptive, because they give a partial sense of control without the real capacity to execute a full exit: portable data but someone else's fine-tune means recovering your own information but retraining from scratch any accumulated advantage, which is a good part of the real cost of irreversibility.

The practical use of this matrix is not to classify contracts already signed — for that it's already too late, though classifying them anyway is useful for calibrating accumulated risk. The practical use is to run the matrix before signing, as a two-line question any decision-maker should be able to answer without ambiguity: if this provider disappears, raises its prices, or simply stops suiting us in eighteen months, exactly what are we left with, and how much does it cost to rebuild what we aren't left with?

Clauses to demand

From the three cases, four concrete clauses emerge that are worth demanding in any AI contract, regardless of the organization's size or available negotiating power. None of them is exotic; all are negotiable in most contracts if raised before the deal is closed in practice, even if it isn't yet on paper.

• Data portability in an open format: the organization's own data — including data generated by use of the system, not just what was initially uploaded — must be exportable in a standard format usable by another provider, without depending on a proprietary export tool the provider can discontinue.
• Ownership of proprietary fine-tunes and embeddings: any adjustment to the model trained on the organization's data must belong contractually to the organization, not the provider, even if the training happened on the provider's infrastructure.
• Discontinuation notice period: the contract must set a minimum number of days — not "reasonable," a number — of advance notice if the provider discontinues the service, substantially changes its terms, or modifies the model in a way that alters the system's behavior in production.
• Code or model escrow: for critical dependencies, an escrow clause that releases the code, model weights, or configuration needed to operate the system if the provider goes bankrupt, is acquired, or substantially breaches the contract.

None of these four clauses guarantees that the organization will never depend on a provider — dependence is part of any reasonable execution decision, and pretending to eliminate it entirely is a fantasy of self-sufficiency that costs more than the dependence itself. What they guarantee is that the dependence remains an execution decision in 2026 and hasn't turned, without anyone explicitly deciding it, into a constitutional decision the organization no longer controls.