Five questions to ask before you sign an AI vendor contract

Standard AI vendor contracts have a structural gap: they specify what you're buying but stay quiet on what it costs to stop buying it. Data export format, model versioning rights, API continuity after cancellation—none of these are defaults you can assume. You have to ask, in writing, before you sign. Here's the checklist.

Question one: Can you actually export your data, and in what format?

Ask the vendor directly what format exports arrive in and who owns the data once it's extracted. Some platforms export conversation logs as standard JSON. Others use proprietary binary formats or strip the metadata that made the data useful in the first place. If the export format is proprietary, you're already paying a switching cost before you've decided to leave.

Push on the operational detail. Does the export include your prompts, your fine-tuning data, and your usage metadata? A vendor might let you pull raw text but withhold the context tags, taxonomy labels, or conversation threading that gave the data its structure. Raw text without that scaffolding is far less useful than what you built—and rebuilding it on a new platform is the hidden cost nobody quotes you.

Test it before you commit. Request a sample export from their sandbox environment. If they won't provide one, that's your answer. You're not seeing what you'd actually be migrating, which means you have no basis for estimating what migration would cost.

Questions two through four: Termination, versioning, and the API rewrite problem

Read the termination clause with an operational eye. What happens to your API integrations on day one after cancellation? Does the vendor give you thirty days of read-only access to pull your last exports, or do your endpoints go dark immediately? The contract language determines how long you have to migrate without production downtime. A vendor who goes dark on cancellation is forcing you to either stay current or scramble. That's leverage, and it's deliberate.

Ask about model versioning and stability. If you build your workflow on a specific model version today, can you keep running it? Or does the vendor sunset older versions on their own schedule, forcing you to retrain your prompts and workflows whether you're ready or not? Lock-in isn't just about leaving—it's about being forced to upgrade on someone else's timeline. A model change that breaks your error handlers or changes how the model interprets your taxonomy can cost your team a week of remediation work per affected workflow.

Model the API rewrite cost before you sign. Your team will build integrations, error handlers, retry logic, and monitoring around one vendor's API shape. Switching vendors means rewriting that code. Ask whether the vendor's API matches industry-standard patterns—OpenAI-compatible endpoints are the closest thing to a portability standard right now—or whether you're learning a bespoke schema. The more idiosyncratic the API, the higher your switching cost. This isn't hypothetical; it's the actual engineering estimate you need before you can evaluate total cost of ownership.

Question five: What does migration actually cost?

Don't ask the sales team. Ask the support team, or look at their onboarding documentation. Do they publish migration guides or case studies from customers who switched between models or platforms? If they don't, that gap is intentional. Complex migrations that vendors don't document publicly are usually more painful than they'd admit in a sales call.

Budget for prompt and workflow retraining separately from data export. Most switching costs aren't in the export—they're in translating your prompts from one model's behavior to another's. Claude and GPT handle ambiguity differently. Their safety training diverges in ways that affect output tone and refusal behavior. Your prompts won't port cleanly. A reasonable minimum estimate for retraining and validating a single non-trivial workflow is forty to sixty engineering hours. Multiply that by the number of workflows you've built, and the number gets real fast.

Run the break-even math before you sign. If your total switching costs—data export, API rewrite, prompt retraining, and validation—land between $15K and $30K for a mid-market SMB, and your current vendor contract saves you $5K per year over the alternative, you're economically locked in for three to six years whether the contract says so or not. That's not a worst-case scenario. That's the median outcome for a team that's integrated one vendor's tooling into three or four internal workflows over six months.

Before you run a vendor demo, share these five questions with your team and your lawyer. Make the vendor answer all of them in writing. A vendor who deflects, equivocates, or won't show you a sample export before you sign is signaling something about how they'll behave after you sign. That's not a sales problem—it's a governance signal worth taking seriously.

We help DFW SMBs structure AI adoption so you can switch vendors or models without rewiring your entire operation. If you're currently evaluating multiple vendors and want someone to pressure-test their answers against your actual stack, schedule an intro call with us. Bring the contracts.