What to Negotiate Now So You Can Actually Take Your Data With You if You Switch AI Agent Vendors Later

Quick answer

Negotiate four specific things before you sign, not after you decide to leave: a defined export format for full conversation history and configuration (not just a summary export), a named timeline for how quickly the vendor must produce that export on request, an explicit statement that configuration assets like prompts, fine-tuning data, and eval sets you contributed are exportable and not vendor-proprietary, and a fee cap or waiver for the export itself. Vendors have limited incentive to make switching easy, so portability terms negotiated during a competitive sales process, when you have leverage, look very different from the same request made after you are already dependent on the vendor and trying to leave.

Negotiate this while you have leverage, not after you need it

The best time to negotiate data portability terms is during initial contract negotiation or renewal, when the vendor is competing for your business, not after a decision to migrate has already been made and the vendor knows you need the export to actually happen. This is the same leverage dynamic behind negotiating SLAs and support terms in an AI vendor contract: terms you ask for as a condition of signing get real attention, and the same terms requested later as a favor from an unhappy customer often do not.

Specify the export format, not just the right to an export

A contract clause that says the vendor "will provide your data upon request" is close to meaningless without specifying the format, because a vendor can technically comply by handing over a format that is unusable without months of engineering work to parse. Require a specified, documented format for conversation history (structured, not a PDF or a proprietary binary dump) and configuration assets, ideally the same structured format you would need to feed into the migration process described in migrating an AI agent between LLM providers, so the export is actually usable the day you receive it rather than a starting point for another six months of parsing work.

Name your configuration assets explicitly, not just your conversation data

Conversation history is the obvious thing to ask for, but configuration, system prompts, prompt version history, fine-tuning datasets you contributed, and eval sets you built are often the more valuable and more contested assets in a portability dispute, because some vendors treat prompt engineering done on their platform as proprietary to the platform rather than to you. Get explicit contract language stating that any prompts, configuration, and evaluation data your team authored belong to you and are exportable in a usable format, distinct from the underlying model or platform itself which remains the vendor's.

Put a timeline and a cost cap on the export itself

A vendor with no contractual deadline can delay an export indefinitely once they know you are leaving, and a vendor with no fee cap can price an export high enough to functionally discourage it. Negotiate both a maximum response time (measured in business days, not left open-ended) and either a fee waiver or a capped fee for a standard export, treating this the same way you would treat any other exit-cost term in what happens if your AI agent vendor gets acquired or shuts down, since an unreasonably slow or expensive export is functionally the same lock-in risk as an outright data hostage situation.

Distinguish this from the vendor's own training-use rights

Data portability terms are about your ability to take data out; they are a separate question from whether the vendor is allowed to use your customer conversation data to train its own shared or foundation model, covered in what to demand in your AI vendor contract when the vendor announces training on customer conversation data. A vendor can concede full portability rights to you while still retaining broad rights to use the same data for their own training, so review both clauses independently rather than assuming one implies the other.

Test the export clause before you actually need it

A portability clause that has never been exercised is unverified. If your contract allows it, request a sample or test export on a defined cadence, even with no plan to switch vendors, to confirm the format and timeline actually work as written rather than discovering gaps in the clause during an actual, time-pressured migration.

FAQ

Is data portability typically included by default in AI vendor contracts? No, it is frequently absent or vague unless specifically negotiated, since most standard vendor contract templates are written from the vendor's perspective and favor retention over portability.

Should smaller companies with less negotiating leverage still ask for this? Yes. Even vendors unwilling to change core commercial terms will often agree to add a portability clause, since it costs the vendor little to promise and materially reduces your risk.

What happens if the vendor refuses to name a specific export format? Treat that refusal itself as a signal. A vendor confident in its own platform typically has no reason to avoid committing to a specific, testable export format.

Does this matter less if we are only using a single foundation model provider directly, without a separate agent platform vendor? It still matters, just at a different layer: the same portability questions apply to conversation logs and fine-tuning data held by the foundation model provider itself, not only to a separate agent-platform vendor.