Before your team puts company information into an AI tool, check six things in its terms: whether the provider can train on your prompts and outputs, who owns the outputs, whether the provider owes you confidentiality, how long it keeps your data, whether it will defend you if an output infringes someone’s IP, and how much it will pay if something goes wrong. Free and consumer tiers usually give you the weakest answer on all six. Business, enterprise and API tiers from the major providers generally don’t train on your data by default and come with real contract protections, which is why they’re usually worth paying for.
Key takeaways
- The tier matters more than the brand. The same provider can offer very different consumer and business terms.
- Look for a contract promise that your inputs and outputs won’t be used for training, not just a setting each user has to find.
- Owning the output under the contract doesn’t mean copyright protects it. Purely AI-generated material generally isn’t copyrightable.
- Many business tiers include an IP indemnity, but the exclusions decide what it’s worth.
- Expect a liability cap tied to what you paid, and broad indemnities running from you to the provider.
- Customers will ask about your AI vendors in security questionnaires, so choose tools you can describe with confidence.
Why do AI tool terms matter for a startup?
Every startup now runs on AI tools for code, writing, research, support and sales. The terms of those tools decide what happens to the source code, customer data, financials and plans your team pastes in. Three risks follow. If you can’t show you protected your confidential information, you can lose trade secret protection for it. If you paste in another company’s confidential information or your customers’ data, you may breach an NDA, a customer contract or a privacy law. And if your product is built on a model API, your customers will want to know exactly where their data goes. See What Are Trade Secrets?
Can the provider train on what we put in?
On consumer tiers, often yes. Many consumer plans let the provider use conversations to improve its models unless the user opts out, or ask each user to choose. Anthropic, for example, asked consumer Claude users in 2025 to choose whether their chats could be used for training, and extended retention to five years for those who opt in. Its commercial products, including team and enterprise plans and the API, are governed by separate commercial terms that don’t work that way. OpenAI’s business terms likewise say it won’t use customer content to improve its services unless the customer agrees.
For company use, you want that promise in the contract that governs your account, not in a setting each employee has to remember to change. Also check retention. Even business tiers usually keep inputs for a period for abuse monitoring or legal reasons, and some API providers offer shorter or zero-retention options on request. Terms change often, so note the version date when you approve a tool.
Who owns the output?
Most AI terms say that, between you and the provider, you own your inputs and outputs, and some assign the provider’s rights in outputs to you. Read the rest of the clause. Some terms keep a license for the provider to use your content, and most warn that other users may receive similar outputs, which you won’t own.
Contract ownership also doesn’t create copyright. The US Copyright Office concluded in January 2025 that AI outputs are protected only where a human contributed enough creative expression, and that writing prompts alone isn’t enough. The D.C. Circuit held in 2025 that a work must have a human author, and the Supreme Court declined to review that decision in March 2026. The parts of a work that people write, select, arrange or substantially modify can be protected. This is a new area of law and it’s likely to shift, so keep records of who built what. See Who Owns Startup IP?
Does the provider owe us confidentiality?
Consumer terms often contain no confidentiality promise from the provider at all, which leaves you no contract claim if your information leaks. Business terms usually treat your inputs and outputs as your confidential information, limit who at the provider can see them and require reasonable security.
If personal data will go into the tool, you also need a data processing agreement, a list of the provider’s subprocessors and data locations, and a privacy policy that discloses the use. See What Are the Startup Best Practices for Data Privacy? and GDPR Basics.
Will the provider defend us if an output infringes?
On free and consumer tiers, generally not. On business and API tiers, several major providers now promise to defend customers against third-party claims that the service, or in some cases its outputs, infringes someone’s copyright. These promises come with exclusions. Common ones cover claims caused by your own prompts or content, outputs you modified or combined with other material, uses that break the provider’s policies, and some patent and trademark claims. Some providers put the output indemnity in separate service-specific terms, so read every document the agreement incorporates.
The indemnity also runs the other way. Most AI terms require you to defend the provider against claims arising from your content, your use of the service and the applications you build on it. That obligation is usually broader than anything the provider promises you.
What liability cap should we expect?
Low. Providers disclaim any promise that outputs are accurate and exclude indirect and consequential damages. Business terms commonly cap the provider’s total liability at the fees you paid in the prior 12 months, with exceptions for things like indemnity obligations. Free tiers may cap liability at a nominal amount. Below enterprise-level spend, these terms are rarely negotiable. At enterprise spend, ask for a higher cap for data breaches and confidentiality breaches.
How do consumer and business tiers usually compare?
| Free or consumer tier | Business, enterprise or API tier | |
| Training on your data | Often allowed unless the user opts out | Generally off by default under the contract |
| Confidentiality | Often none | Your inputs and outputs treated as confidential |
| Output ownership | Usually yours, subject to provider rights | Usually yours, often with an assignment |
| IP indemnity | Rarely | Often, with exclusions |
| Liability cap | Nominal | Typically fees paid in the prior 12 months |
| Admin controls | Individual accounts | Central billing, single sign-on, retention settings |
These are typical positions, not guarantees. Read the terms for the specific plan you’re buying.
What about AI models built into our product?
When your product sends customer data to a model API, the AI provider becomes your subprocessor. Check the API’s data processing terms, retention and data location. Then look at your own customer contracts. Many enterprise customers restrict sending their data to third parties or using it to train AI, and their security questionnaires will ask which AI vendors you use, whether they train on customer data and how long they keep it. Make sure your terms of service and privacy policy describe the AI processing accurately, and pass the provider’s usage restrictions through to your customers. See How Do I Negotiate Customer Contracts? and How to Draft Terms of Service.
Checklist: before you approve an AI tool
- Confirm which plan you’re on and which terms govern it.
- Training on your inputs and outputs is off under the contract.
- Retention period, deletion rights and any zero-retention option are known.
- The provider owes you confidentiality for inputs and outputs.
- A data processing agreement and subprocessor list are in place if personal data is involved.
- You’ve read the IP indemnity and its exclusions.
- You know the liability cap.
- Admin controls (central accounts, single sign-on, the ability to remove users) are available.
- The tool is on your AI-use policy’s approved list and your vendor list for security questionnaires.
Can I use AI to review these terms myself?
Use it to get organized. AI is good at summarizing a long set of terms, flagging the clauses above and turning your concerns into a short list of questions for your lawyer. That makes a lawyer’s review faster and cheaper. AI-drafted legal documents and redlines are a different story. They usually take a lawyer longer to review than a plain description of what you need, which can raise the cost. See Can’t Afford a Startup Lawyer Yet?
For related topics, see Copyright Infringement, US Data Privacy Laws and Startup Intellectual Property.
If you’d like an AI vendor’s terms reviewed before you sign, see our contract review services.
