Negotiating an AI contract for beginners: retention, zero data retention, training and models

Four clauses decide whether an AI contract is safe: retention, training, access and model.

You do not need a lawyer to ask the right questions. You need to know what the words mean and what a good answer sounds like.

The terms

Term	What it means	What to ask for
Data retention	How long the provider keeps your prompts and the model’s answers	A number of days, and the reason for keeping them
Zero data retention (ZDR)	The provider keeps nothing once the answer is returned	Whether it is available, on which models and features, in writing
Training on your data	Whether your prompts are used to improve their models	Off, by contract, not by a setting someone can change
Logging	What the service records about each request	Metadata only. Never the content
Access	Which people and companies can read your prompts	A named list. Ideally nobody outside your account
Location	Where your data is processed	The country or region your data rules require
Model and version	Exactly which model answers, and for how long it stays available	A named version and notice before it is retired

What real answers look like

Providers publish these terms. Read them before the sales call.

Provider	What its own pages say
OpenAI, academic program	Workspace “data is not used to train our models by default”
Amazon Bedrock	Model providers “don’t have access to Amazon Bedrock logs or to customer prompts and completions”
OpenRouter	Prompts and completions “are not logged by default.” Each model provider behind it has its own policy, and you can block providers that may train on your data
Anthropic	Its documentation notes that zero data retention applies differently by feature, so ask which features are covered

Two lessons from that table.

“By default” is not “by contract.” A default can change with a setting or a policy update. Ask for the commitment in the agreement.

A router is two contracts. When one service passes your prompt to another company’s model, both sets of terms apply. Ask about the second one.

Good answers and bad answers

Question	Good answer	Walk away if
How long do you keep our prompts?	“Zero, or a stated number of days for abuse monitoring”	“It depends” or no number
Do you train on our data?	“No, and the contract says so”	“Not currently”
Who can read our prompts?	“Nobody outside your account, and here is the exception process”	A vague reference to “authorized personnel”
Where is it processed?	A named region	“Globally”
Which model do we get?	A named version, with notice before retirement	“Our latest”
What do we get for sharing data?	Nothing, because you do not share it	A discount for letting them log your prompts

OpenRouter, for example, offers a 1% discount to users who opt in to prompt logging. For research data that is never worth taking.

How to run the negotiation

List what will go into prompts. Public text, your own code, unpublished results, licensed data, personal data. Each one has a different answer.
Match it to your data licenses. A provider’s promise does not override a vendor’s license. If the license forbids AI use, the best AI contract in the world does not help. See what you may do with an agent on licensed data.
Ask the seven questions in writing. Email, not a call.
Put the answers in the contract. If the seller will not, you have learned what the answer was worth.
Pin the model. A result you cannot rerun on the same model version is hard to defend. Ask how long a version stays available.
Review at every renewal and every major model release. Terms change as fast as the models do.

Who should do this

Whoever signs the purchase order, with one researcher in the room who knows what the data is. If your department has already signed, ask to see the answers to the seven questions. If nobody can find them, that is your first agenda item.