How to Write an NDA for an AI Project

A non-disclosure agreement for an AI project needs to do everything a normal NDA does, plus name the third-party AI vendors that will touch your client's data, disclose what happens to that data once it leaves your hands, and assign responsibility if a model provider mishandles it. Generic templates from a template site or a quick search rarely mention any of this, because they were written before an API call to OpenAI or Anthropic was a normal part of doing client work. If you build software with AI models instead of writing every line yourself, your NDA needs a clause most competitors' contracts don't have.

Why a Standard NDA Template Falls Short for AI Work

Most NDA templates were built for a simple scenario: two humans exchange documents, and one promises not to repeat what they heard. A non-disclosure agreement, at its legal core, is just a promise to keep specific information secret. That model breaks down the moment your workflow includes a large language model.

When you build an AI feature for a client, their data doesn't stay between the two of you. Product specs, customer records, internal pricing logic, or sample support tickets often get pasted into a prompt, uploaded as context, or piped through an API to generate embeddings, fine-tune a model, or power a chatbot. That data now touches at least one more party: the AI vendor whose infrastructure processes it.

A generic NDA is silent on this. It talks about "confidential information" and "authorized representatives" without ever asking whether a client's data left your laptop and landed on a server run by OpenAI, Anthropic, Google, or a smaller API provider. For protecting client data as an AI freelancer, that silence is the actual risk. If something goes wrong with how a model vendor stores or handles your client's information, a standard NDA won't tell you or your client who is on the hook.

One-Way vs Mutual NDA: Which Fits Your AI Project

NDAs come in two basic shapes, and picking the wrong one either leaves you exposed or asks your client to sign something heavier than the relationship needs.

A one-way, or unilateral, NDA protects information flowing in one direction. In most freelance and small-agency AI projects, that's the shape you want. The client shares confidential information, their data, their business logic, their user records, so you can build the thing, and you agree to protect it. You aren't handing them anything comparably sensitive in return, so there's no reason to burden the contract with mutual obligations.

A mutual, or bilateral, NDA protects information flowing both directions. Use it when you're also disclosing something worth protecting: your own prompt library, an internal architecture you're licensing, or a pre-built AI tool you're customizing rather than building from scratch. It also fits partnership-style engagements where you and the client are co-developing a product and both sides are exposing IP.

A quick way to check yourself: if you'd be uncomfortable seeing your own internal notes cited back to you as "confidential information the client will protect," you need a mutual NDA template rather than a one-way one. If not, the one-way version is lighter, faster to negotiate, and just as enforceable.

The Sections Every AI Project NDA Needs

A workable NDA for AI project work does not need to be long. It needs to cover six things clearly.

• Definition of confidential information, written broadly enough to include prompts, model outputs, and any data used as context for a model, not just "documents marked confidential."

• Third-party AI processing disclosure, naming which AI providers you use and stating that client data may pass through those providers' infrastructure during the engagement.

• Data handling and subprocessor terms, covering how long data is retained, whether it is used to train any model, and what happens to it once the project ends.

• Standard exceptions, since information already public, independently developed, or already known before disclosure isn't covered. Courts expect this, and leaving it out looks amateurish.

• Term and survival, since most AI project NDAs run two to five years past the engagement, though genuine trade secrets can warrant an indefinite term.

• Remedies, a plain statement that a breach can cause harm money alone won't fix, supporting a request for a court order rather than just damages.

An Example Clause: Disclosing Third-Party AI Processing

Here's an illustrative clause addressing the gap described above. It is not legal advice and not a complete NDA on its own. It's meant to show the kind of language that closes the third-party processing gap, so you or a lawyer can adapt it to your actual engagement.

3. Third-Party AI Processing Disclosure

3.1 Recipient may process Confidential Information using third-party artificial intelligence tools and application programming interfaces ("AI Providers"), including but not limited to OpenAI, Anthropic, and Google, solely as necessary to perform the services described in the applicable statement of work.

3.2 Before processing any Confidential Information through an AI Provider, Recipient shall: (a) identify the AI Provider(s) to be used, in writing; (b) confirm that the applicable AI Provider's terms of service exclude submitted data from model training, or obtain Disclosing Party's written consent if no such exclusion is in place; and (c) configure account-level settings to the shortest data retention period the AI Provider offers.

3.3 Recipient remains fully responsible for an AI Provider's handling of Confidential Information as if Recipient had handled it directly. A breach of this Section by an AI Provider is treated as a breach by Recipient.

3.4 Recipient shall not submit personally identifiable information, financial account data, or health information belonging to Disclosing Party's customers to any AI Provider without Disclosing Party's prior written approval.

Protecting Client Data as an AI Freelancer, Beyond the Signature

The NDA sets the rule. Your actual workflow decides whether you keep it. A few habits matter more than the contract language itself.

• Turn on the strictest data retention or zero-retention setting your AI vendor offers for business or API accounts, rather than relying on defaults built for consumer chat use. OpenAI's own policy states that API data is excluded from model training by default and describes the retention windows involved.

• Keep client work inside a business or API account, not a personal ChatGPT or Claude login, since the terms differ between consumer and commercial tiers. Anthropic publishes similar default-no-training terms for its commercial API and Enterprise customers.

• Strip or mask identifying details, names, account numbers, exact figures, before they go into a prompt, whenever a placeholder would do the same job.

• Maintain a short, current list of which AI providers touch which client's data, so the disclosure clause in the NDA stays accurate instead of becoming fiction three months into the project.

This overlaps with what you'd put in a privacy policy for an AI app if you were shipping a consumer product. The difference is you're writing the internal version for a client relationship instead of the public-facing one. It's also worth reading up on whether it's safe to give AI access to your data before you commit to specific data-handling promises, since what's actually safe depends on which vendor and tier you're using.

Fitting the NDA Into Your Broader Contract Stack

An NDA is not the whole agreement. It's the piece that protects information; a separate statement of work should cover deliverables, payment, and IP ownership. If you're also selecting the AI vendor whose API will power the build, that's a distinct conversation worth having on purpose, and negotiating a contract with an AI vendor is worth doing before you lock in a provider for a client engagement.

Contracts like this are unglamorous compared to landing the client or shipping the feature, but they're part of the same business. If you're thinking through how to run an AI-based freelance practice, it's worth treating solid paperwork as part of your monetization strategy rather than an afterthought. Clients notice when a freelancer already has an answer for "what happens to our data," and that answer is often the difference between winning a contract with a cautious enterprise client and losing it to someone who just sounds more careful on paper.

One honest caveat before the FAQ: everything above, including the example clause, is a starting point for a conversation, not a finished legal document. NDA enforceability rules differ by state and country, some jurisdictions limit how broadly you can define confidential information or how long a restriction can run, and a lawyer who has actually read your specific engagement will catch problems a blog post can't. If real money or genuinely sensitive data is involved, it's worth the one-time cost of having someone licensed review the document before you send it to a client.

Frequently asked questions

Do I need an NDA for an AI project?

You need one whenever a client will share information you wouldn't want a competitor or the public to see, which is most AI builds, since they usually involve real user data, internal business logic, or unreleased product plans. Small, low-stakes projects sometimes skip it, but AI work specifically warrants one because of how data moves through third-party model providers during the build.

Is a one-way or mutual NDA better for freelance AI work?

A one-way NDA covers most freelance AI engagements, since the client is usually the only party disclosing sensitive information. Switch to a mutual NDA template when you're also sharing something worth protecting, your own tooling, prompt library, or proprietary architecture, or when the project is closer to a partnership than a straightforward client build.

What should an AI project NDA say about third-party AI tools?

It should name the AI providers you plan to use, state that client data may be processed by those providers' infrastructure, describe the retention settings you'll configure, and make clear that you remain responsible for how those providers handle the data, not just for your own handling of it.

How long should an AI project NDA last?

Most run two to five years past the end of the engagement, long enough to matter without being unenforceable in jurisdictions that frown on indefinite terms. Genuine trade secrets, like a proprietary algorithm or model architecture, can reasonably be protected for as long as they stay secret.

Can an NDA stop a client's data from being used to train an AI model?

An NDA can require you to configure vendor settings and choose AI providers whose terms exclude API data from training, and can make you liable if you don't, but it can't unilaterally change a third-party vendor's own policies. That's why the disclosure clause matters: it forces you to check an AI provider's actual training policy before you use it on client data, rather than assuming.