← Back to blog

The five contract clauses that decide whether you own your AI system

Most AI vendor contracts are reviewed like software contracts. Five clauses determine whether you own the model, the data, and the exit.

Most AI contracts get reviewed the way software contracts get reviewed. Legal checks the liability cap, the service levels, the IP indemnity, and the auto-renewal date. All of that matters. None of it determines whether, two years from now, you own a working system or you are renting one you cannot leave.

Five clauses decide that. They are usually short, often buried in an appendix or in a linked acceptable-use page that changes without notice, and drafted in the vendor's favor by default. Every one of them is negotiable before you sign and almost none of them are negotiable after.

1. What the vendor may do with your data

The clause to find is the one granting the vendor a license to your inputs and outputs. Read it for two words: "improve" and "aggregate." A right to use your data to "improve the services" is a training right in polite language. A right to use "aggregated and anonymized" data is broad enough to cover almost anything, because nobody defines the anonymization standard.

What to ask for: no training on your data without separate written consent, no use for any customer other than you, and a named subprocessor list you get notice of before it changes. If the vendor cannot agree, they are selling you a product whose price is partly paid in your data. That can still be a fair trade. It should be a decision, not a surprise.

2. Who owns the tuned model

If the vendor fine-tunes a model on your documents, tickets, or transcripts, three assets exist: the base model (theirs), your training data (yours), and the adapted model in between (unstated, in most drafts).

Get it stated. You will rarely get the weights, because they are derived from a base model the vendor licenses from someone else. You can usually get the next best thing: a perpetual, irrevocable license to use the adapted model, a guarantee it will not be deployed for any other customer, and a right to receive the training artifacts — the curated dataset, the evaluation set, the configuration — on exit. Those artifacts are what let another vendor rebuild in weeks instead of months.

3. Getting your data back, in a usable shape

Nearly every contract promises data return. Very few define the format. A vendor who hands you 400,000 rows of unlabeled JSON with internal IDs and no schema has met their obligation and left you with nothing.

Specify the deliverable: documented schema, stable identifiers, all metadata your team created inside the product (labels, corrections, human feedback, evaluation scores), a delivery window in days, and a deletion certificate afterward. The human feedback is the part people forget. It is often the most expensive thing your team produced, and it is the thing most likely to sit outside any export.

4. Control over model changes

This is the clause almost no contract has, and the one that causes the most damage. The vendor swaps the underlying model, the outputs shift, and a workflow your team validated in March behaves differently in September. Nothing broke by the SLA's definition. Availability was fine.

Ask for change control on anything that alters behavior: advance notice of model or major prompt changes, the right to stay on a prior version for a defined window, and access to the evaluation results behind the change. If the vendor holds your evaluation set, ask to run your own against any new version before it reaches production.

5. What exit actually costs

Termination clauses tell you when you can leave. They rarely tell you what leaving involves. Add a transition assistance clause and price it now, while you still have leverage.

A workable version covers:

  • A defined transition period (90 to 180 days) at the contract rate, at your option
  • Named deliverables: data export, documentation, configuration, integration specs, prompts and retrieval logic
  • Continued service during transition, at the same service levels
  • A cap on transition fees, agreed at signature rather than quoted at exit

Before the next contract goes out

Send the vendor these five questions before procurement starts drafting. The answers tell you more than a demo will. Vendors confident in their product tend to answer directly; vendors who depend on switching costs tend to answer in paragraphs.

You are not trying to win every point. On a small pilot, conceding the training rights may be reasonable. The goal is to know which of the five you traded away and what it will cost to buy back. We work through this with clients before signature rather than after, because after is where the weight lands.

Articles · Hercules A

Insights on adapting your company to AI.

Contact us