Contract Terms That Make an AI Vendor Exit Cheap

"Customer may export its data" sounds sufficient and usually is not. Four attributes determine whether an export right is real: format, scope, timing and cost.

Format. Require a machine-readable, documented, non-proprietary format, and name it. JSON Lines or CSV for structured records, Parquet for large tables, and the original binary files for documents. If the vendor stores embeddings or vector indexes, ask explicitly whether those are exportable or whether you must recompute them. Recomputing an embedding index over a large corpus is a real cost, and it belongs in the contract, not in a post-termination surprise.

Scope. Define customer data broadly. AI contracts need definitions that cover prompts, queries, embeddings, fine-tuning data, retrieval-augmented generation materials, user feedback, generated outputs, metadata and logs per Tish Law. If "Customer Data" means only uploaded files, everything the system learned about your business from interactions falls outside the export right.

Timing. Specify a self-service export available at any time during the term, not only in a termination window. A mid-term export right is how you test the clause before you need it.

Cost. State that export is included in the subscription fee during the term and for a defined period after termination, and that no per-gigabyte egress or professional-services fee applies to a standard export.

A worked example shows why scope matters. Suppose a support assistant processes 40,000 conversations a month for 18 months, roughly 720,000 conversations. If the contract exports transcripts and metadata but excludes the human corrections your agents made, you lose the labeled dataset that made the system worth having. Rebuilding that label set at even a few minutes per conversation is a five-figure internal cost and a multi-month delay.

The practical test: before signature, ask for a sample export from a sandbox tenant and try to load it into your own warehouse. If the vendor cannot produce a sample, the clause is aspirational.

Notice has two directions. Your notice period determines how long you keep paying during a migration. The vendor's notice period for material changes determines how long you have to react to changes you did not ask for.

Ninety days is a common starting point for customer termination notice, and it is a defensible floor. Anything longer than 90 days is effectively a lock-in term with a friendlier name. Push for 60 days if your spend gives you leverage, and treat 30 days as a win only if the contract also guarantees you access through the end of the current billing period.

Note that a 90-day migration runway is frequently insufficient in practice. Building a complete inventory of affected systems can take two to three weeks on its own, before behavioral validation, compliance documentation and stakeholder communication per Airia. Use that when the vendor argues 90 days is generous: it is generous for a notice obligation, not for a migration.

Termination assistance is the clause most often omitted. It should obligate the vendor to provide, for a defined window after notice (30 to 60 days is reasonable), a documented export, a data dictionary, read-only API access for verification, and a named technical contact. Ask for a fixed allocation of support hours rather than "reasonable assistance", because "reasonable" is not measurable.

Model pinning locks the specific model version you tested, so your evaluation results stay valid. Without it, a routine vendor upgrade can move your accuracy, latency and cost profile without a single line of your code changing. The contract should define "Model" to include a versioning and update mechanism, and should state whether the model is dedicated to you or shared across customers per Ward and Smith.

Three sub-terms make pinning useful.

Pin duration. Ask for the pinned version to remain available for the full contract term, or at least 12 months after you adopt it.

Deprecation notice. Require at least 12 months written notice before deprecation of any pinned model, with the notice delivered to a named contact. Tie the notice to something concrete: a stated end-of-life date and a documented migration path to a successor model.

Behavioral equivalence. Ask the vendor to represent that a successor model, on your evaluation set, does not materially degrade quality, or to give you a documented procedure for demonstrating it. Vendors may resist a hard guarantee because AI behavior is probabilistic, and that objection is reasonable Stanford CodeX analysis. A test protocol is the workable compromise.

One consequence worth planning for: pinning is not free. A pinned older model may carry higher per-token pricing than its successor, and it will eventually become unsupported. Treat pinning as insurance with a premium, and review whether you still need it each renewal.

Two clauses make the rest enforceable.

Change of control. If the vendor is acquired, its roadmap, pricing and data-hosting choices can change without your consent. Require that a change of control triggers a customer termination right with a shortened notice period (30 days) and no early-termination fee. This matters more as AI vendors consolidate, and enterprise software buyers have watched acquisition-driven terms shift before.

Benchmark and audit. A right to run your own evaluation against the pinned model and to receive the vendor's published model cards, change logs and deprecation schedules. Without it, you cannot prove degradation when it happens.

Payment and credit terms. If the vendor fails an export or migration deadline, a service credit is weak compensation. Ask instead for an extension of the assistance window and, for material failures, refund of the remaining prepaid term.

The four clauses are usually obtainable. What changes the outcome is how the request arrives.

Open with risk management. Frame the ask as satisfying a governance obligation, not as preparation for leaving. Draft the language as an AI addendum that supplements the master agreement, since the MSA rarely covers what matters for AI per Gouchev Law. Vendors see hundreds of addenda and treat them as routine.

Ask in the initial draft. Terms introduced late are treated as costly; terms present in the first redline are treated as baseline. The vendors that grant export rights routinely are the ones never asked twice.

Trade on dimensions the vendor cares about. Case studies, logo rights, committed spend or a longer initial term are cheap for you and valuable to them. Do not trade pinning duration for a discount.

"Would you sign this if the roles were reversed?" is a better line than a lecture about lock-in. Most commercial teams will concede a 12-month deprecation notice because they know their own product roadmap moves.

Fall back in this order: full export right, then termination assistance, then deprecation notice, then pinning. Do not let a stalled section on pinning delay an otherwise favorable agreement.

Get a sample export before you sign. This is the single highest-leverage action available to a procurement team, and it costs one email.

If your organization is running several AI vendors at once, the same four clauses can be standardized as a house addendum across every contract, which reduces per-deal legal time and makes outcomes comparable.

Exit cost is asymmetric. The vendor bears almost none of it: your departure is a lost logo. You bear all of it, and you bear it exactly when you have the least capacity, during a migration. Contracts written under that asymmetry consistently under-protect the customer.

The practical consequence: negotiate exit terms at the moment you have maximum leverage and minimum willingness to walk away, which is before signature. After signature, the best available outcome is a mid-term amendment, and amendments cost something.

For teams that want a second opinion on the whole stack rather than a single contract, a free 30-minute AI discovery call maps your goals to a tailored recommendation with no obligation: Quanta Consulting. If the migration itself is the blocker, an ISO-style export and re-index can be run against a single OpenAI-compatible endpoint, which is what Quanta Inference provides across text, image, video, audio and embedding models.