AI contract terms to negotiate before you sign

Six terms to settle in enterprise AI tool and model API agreements: commitments, price protection, training opt-outs, output IP and indemnity, seat versus usage, and a budget owner.

Before you sign an enterprise AI tool or model API agreement, negotiate six things: how you pay (committed spend or pay-as-you-go), price protection, data use and training, output ownership and indemnity, the mix of seat and usage pricing, and a named budget owner. On data, the major vendors' standard business terms already say they won't train on your content, so the harder fights are usually price changes on short notice, commitments sized for adoption that hasn't happened, and indemnities that come with conditions.

The money is real and it is running ahead of plans. Futurum Group's September 2026 survey of technology buyers found 46.9% of enterprises reported running over budget on AI, and 10% had no formal AI budget to measure against. Most of that is a contract and governance problem, not a technology one.

Committed spend or pay-as-you-go

A spend commitment buys a discount. It also turns your forecast into a liability. If adoption comes in below plan, you pay for capacity you never used.

Our view: don't commit until you have at least three months of production usage data. Until then, pay as you go and accept the higher unit rate. When you do commit, ask for:

• A ramped commitment that starts low and rises with real usage.
• Rollover of unused commitment into the next period.
• A commitment that applies across models and products, so a shift to a newer or cheaper model doesn't strand it.
• No automatic increase in the commitment at renewal.

Price protection

Read the rate change clause before the rate card. Anthropic's Commercial Terms, for example, say it may update published rates, effective the earlier of 30 days after posting or when the customer receives notice. That is a normal position for a usage-based service. It is also why you negotiate.

Ask for unit rates fixed for the term on the models you use, with the benefit of any published price cut. Ask how model retirements work: how much notice you get, and what you'll pay on the successor model. A price lock on a model that is retired in six months protects less than it looks like.

Data use and training opt-outs

This is the clause most companies worry about, and it is the one where standard terms are strongest.

• Anthropic's Commercial Terms: "Anthropic may not train models on Customer Content from Services."
• OpenAI says that by default it does not use data from its business products, "including inputs or outputs," for training or improving its models, and that organizations can opt in.
• Google Cloud's Service Specific Terms: "Google will not use Customer Data to train or fine-tune any AI/ML models without Customer's prior permission or instruction."

So the work is in the edges. Confirm you are on the business product, not a consumer plan staff signed up for with a work email. Check retention periods and subprocessors. And get the commitment into your signed order form or agreement, because terms pages are versioned and change.

Output ownership and IP indemnity

Two questions: who owns what the model produces, and who pays if it infringes someone else's rights.

On ownership, Anthropic's terms say the customer "owns its Outputs" and assign the company's rights in them to the customer. On indemnity, Anthropic will defend customers against third-party claims that authorized use of the service, or outputs generated through it, violate intellectual property rights.

Read the conditions. Microsoft's Customer Copyright Commitment defends customers against certain third-party IP claims about output content, but for Azure OpenAI and configurable services the customer must have implemented Microsoft's required mitigations, must show compliance when it tenders a claim, and has six months to implement new mitigations Microsoft publishes. That is reasonable. It is also a compliance obligation someone in your company now owns. Ask for every condition to be listed in the contract, and ask whether IP claims sit under the general liability cap or outside it.

Seat versus usage pricing

Seats are predictable and waste money on people who don't use the tool. Usage scales with value and has no natural ceiling. Futurum's read is that subscription and usage-based pricing "continues to displace seat-based and fixed-price models," so expect more hybrids.

For seats, negotiate a true-down at each anniversary and pull utilization data before every renewal. We cover why in Your AI pilot is a license, not a strategy. For usage, require admin controls that set hard spend limits per workspace, project or user, alerts before limits are hit, and usage reporting you can map to cost centers.

A named budget owner

Every AI contract needs one person who signs off on the commitment, watches the monthly number and decides the renewal. Not IT by default, and not "the business." A name. We made the case in Your AI agents need a budget owner, and the contract should support it: spend caps that person controls, invoices split by cost center, and usage logs they can read without a ticket.

Before you sign

Put the six items above on one page next to the vendor's order form and mark each one as agreed, open or missing. Anything missing gets raised before signature, when you still have the vendor's attention. If you want a second pair of eyes on the full AI spend picture, start with our AI spend work.

Sources: Futurum Group, Enterprise AI overruns hit 46.9%, Anthropic, Commercial Terms of Service, OpenAI, Business data privacy, security, and compliance, Google Cloud, Service Specific Terms, Microsoft, Customer Copyright Commitment required mitigations.