An NDA or customer contract blocks licensing of the other party's confidential information and of content you hold on its behalf. It does not usually block your own operating records. Four clauses decide where the line falls: the definition of confidential information or customer data, the permitted use, any usage-data or aggregated-data right, and any restriction on AI training. Where a contract does restrict a record set, written consent from the customer is the way through.
- Sort every record set into theirs, yours and mixed before reading clauses.
- A standard NDA lets the receiver use confidential information only for the stated purpose.
- Removing names fixes privacy. It does not always fix confidentiality.
- A "no AI training" promise can sit in the main agreement, the DPA, an order form or a security questionnaire.
Does an NDA prevent selling or licensing data?
It prevents licensing the other party's confidential information. It does not reach records that are yours and contain nothing of theirs. Most real record sets are a mix, so the work is sorting them.
A widely used standard form shows the mechanism. Under the Common Paper Mutual NDA the receiving party shall "use Confidential Information solely for the Purpose" and shall "not disclose Confidential Information to third parties" without prior written approval (Common Paper). The purpose is usually to evaluate or carry out a business relationship. Training an outside party's AI model is not that.
The definition is broad. It covers what the discloser marks as confidential and what "should be reasonably understood as confidential or proprietary due to its nature and the circumstances of its disclosure" (Common Paper). Unmarked drawings, price lists and system details usually qualify. This page is general information, not legal advice. Your own agreements decide.
Which four clauses decide it?
The definition, the permitted use, the usage-data right and any AI restriction. Read them in that order for each customer agreement, then check the data processing terms if personal data is involved.
| Clause | Question it answers | What to look for |
|---|---|---|
| Definition of confidential information or customer data | Which records are theirs? | Whether it covers only what they gave you, or also what you created about them |
| Permitted use or purpose | What may you do with their information? | Usually "only to provide the services" or "solely for the purpose" |
| Usage data, aggregated data or service data | Is there a carve-out for your own analytics? | Whether it must be aggregated, de-identified, or used only to improve the service |
| AI and machine learning restriction | Did they bar training on their data? | A sentence in the agreement, the DPA, an order form, a security addendum or an AI addendum |
Also read the survival clause. Confidentiality usually continues for years after the agreement ends, and for trade secrets often without limit.
What is their confidential information and what is your own operating record?
Theirs is what they gave you or what you hold for them. Yours is the record of how your own business ran. Mixed records are yours in form but carry their information inside, and those need redaction or consent.
| Record | Usually | Why |
|---|---|---|
| Files, drawings and data the customer sent you | Theirs | Disclosed to you under the agreement |
| Content the customer stored in your product | Theirs | Customer data held to provide the service |
| Deliverables you made for the customer | Depends on the IP clause | Many agreements assign deliverables to the client |
| Your internal procedures, templates and training material | Yours | Made by you, for you |
| Your general ledger, payroll and purchasing records | Yours | Your own business records, with personal data to remove |
| Your source code and engineering history | Yours | Unless written for a client under an assignment clause |
| Support tickets about the customer's systems | Mixed | Your record of your work, containing their system details |
| Emails and chat with the customer | Mixed | Both sides' words, and often their confidential information |
| Project files for client engagements | Mixed | Your methods and their facts |
Firms whose work product is the client's by contract should read the guide for consulting firms.
What is a no-AI-training clause?
A promise that one party will not use the other's data to train AI models, or will do so only with permission. Large vendors now publish such promises, and enterprise customers ask their own suppliers for the same.
Two published examples show the usual forms. Google Cloud's terms say: "Google will not use Customer Data to train or fine-tune any AI/ML models without Customer's prior permission or instruction" (Google Cloud). Zoom's say it "does not use any of your audio, video, chat, screen sharing, attachments or other communications-like Customer Content" to "train Zoom or third-party artificial intelligence models" (Zoom). The first is a ban with a consent route. The second names third-party models, which is the case that matters when you license records to someone else.
If your customers negotiated a clause like either of these with you, their data is out unless they agree. The clause may not be in the main agreement at all. Look in the data processing agreement, order form special terms, security addenda, AI addenda and answers your company gave in procurement questionnaires, which some contracts incorporate.
Does my customer contract allow data licensing?
For customer content, almost never without consent. For usage data, sometimes, and usually only in aggregated form. The answer is in two sentences of the agreement, and they are worth reading word by word.
A common standard form puts it this way. On content: "Provider may copy, display, modify, and use Customer Content only as needed to provide and maintain the Product and related offerings." On usage data: "Provider may only disclose Usage Data to others if the Usage Data is aggregated and does not identify Customer or Users" (Common Paper Cloud Service Agreement). Under terms like these, content cannot be licensed and usage data can go to a third party only when aggregated.
Aggregated tables are of limited use for training, because the value is in the individual cases. So for most software companies the licensable set is their own operating data and code, not customer content or raw usage events. The guides for SaaS companies and for product usage data go further.
Can you license records if customer names are removed?
Sometimes. Removing names and personal details answers the privacy question. It answers the confidentiality question only if what remains no longer contains or reveals the customer's confidential information.
A redacted support ticket that still describes one customer's unreleased product or network layout is still their confidential information. A ticket about a generic fault in your own product, with the customer removed, usually is not. European law makes a similar point about personal data: records "which could be attributed to a natural person by the use of additional information" still count as personal (GDPR, Recital 26).
A workable test for mixed records: could the customer recognise its own confidential facts in the redacted record? If yes, ask for consent or leave it out. The de-identification guide covers methods.
Do you need to tell customers, and how do you ask?
Where the contract restricts the records, you need written consent, not just notice. Where it does not, there is no contractual duty to tell them, though privacy law may still require notice to individuals. Many companies ask anyway for records that describe a customer closely.
Ask plainly, for a defined set, with an easy way to say no. A request that reads like a trick costs more goodwill than a refusal. The wording below is an illustration for your own counsel to adapt, not a form to copy.
| Part of the request | Sample wording |
|---|---|
| What | We would like to include de-identified records of our support work for you from 2019 to 2024 in a dataset licensed for AI model training and evaluation. |
| What is removed | Your company name, staff names, contact details, system identifiers and any attachment you sent us are removed before delivery. |
| Who receives it | The licensee may use the records only to train and evaluate its own models. It may not resell them or try to identify you. |
| Your choice | This is optional and does not affect our services to you. If you do not reply by the date below, your records stay out. |
| Consent line | [Customer] consents to the use described above as an amendment to the confidentiality terms of the agreement dated [date]. |
Silence is treated as no in this sample. That is deliberate: consent to change a confidentiality clause should be a signature, not a missed email.
What should new contracts say?
Say what you may do with your own operating records and with de-identified data, and say what you will not do with customer content. A clause that claims too much will be struck by every careful customer, and a struck clause is an exception to track.
Sample wording, as an illustration for counsel to adapt to the agreement and the law that governs it:
- Customer content. "Provider will not use Customer Content to train artificial intelligence models, and will not license or disclose Customer Content to any third party for that purpose, without Customer's prior written consent."
- Provider records. "Provider owns its own business records relating to the services, including support, quality and engineering records. Provider may use and license such records after removing Customer Content, Customer's name and personal data."
- De-identified data. "Provider may use and disclose data derived from the services that is de-identified and does not identify Customer, its users or its confidential information."
- Opt-out. "Customer may at any time notify Provider that records relating to Customer are to be excluded from any licence to a third party."
Keep a register of every customer that changes or strikes these clauses. The register is the exceptions list a buyer's counsel will ask for.
Common mistakes with NDAs and customer contracts
The clause people look for is rarely the one that catches them.
- Searching contracts for "AI" and stopping when nothing is found. The use limitation does the work.
- Reading the master agreement and not the DPA, order forms and security addenda.
- Treating expired agreements as finished when the confidentiality clause survives.
- Assuming redaction of names cures a confidentiality clause.
- Relying on an aggregated-data right to license record-level data.
- Sending consent requests that bury the purpose.
- Forgetting NDAs with prospects who never became customers but did send files.
- Keeping no list of which customers negotiated their own terms.
What is left after the contract review is the data asset
A company that has sorted its records into theirs, yours and mixed, and can show the exceptions list, holds something most companies cannot offer: records with a known contract position. AI developers ask for exactly that before they talk about price. The review also tells you which customers to ask, and for what.
Origin works with buyers and helps a company describe the records that pass the review, find out what they could be worth and decide whether to go further. Nothing is exported at the start, and no customer is named. For the full legal review, use the general counsel checklist. Origin Data Partners is an independent data partner that helps companies get their data bought. A buyer may pay Origin when a deal closes.
Does your company hold records like these? See what your data is worth.
Common questions
Does an NDA prevent selling data to an AI company?
It prevents licensing the other party's confidential information, because a standard NDA allows use only for the stated purpose and bars disclosure to third parties. It does not usually cover your own operating records that contain none of their information.
What is a no AI training clause in a contract?
A promise not to use the other party's data to train AI models without permission. Google Cloud's terms, for example, say it will not use customer data to train or fine-tune AI models without the customer's prior permission or instruction. Customers increasingly ask suppliers for the same.
Does a customer contract allow licensing data for AI training?
Rarely for customer content: most agreements let the provider use it only to provide the service. Some allow use or disclosure of usage data that is aggregated and does not identify the customer. Read the customer data, usage data and confidentiality clauses for each customer.
Can I sell customer emails for AI training?
Not as they are. Emails with a customer hold both sides' words, personal data and often the customer's confidential information. They need de-identification, and where a confidentiality clause covers the content, the customer's written consent.
Is an AI training opt-out clause the same as a ban?
No. A ban bars training unless the customer agrees. An opt-out allows it until the customer objects. Slack's privacy principles, for example, tell customers they can opt out of helping train its global models. Which one you signed decides whether you need consent or only a way to honour objections.
What data rights clause should a new customer contract include for AI training?
One that separates three things: customer content, which you will not license without written consent; your own business records, which you may license after removing customer content and names; and de-identified data. Add an opt-out and keep a register of changes.
Sources and further reading
- Common Paper: Mutual NDA standard terms
- Common Paper: Cloud Service Agreement standard terms
- Google Cloud: Service Specific Terms (training restriction)
- Zoom: Terms of Service (customer content)
- Slack: privacy principles for search, learning and artificial intelligence
- Regulation (EU) 2016/679 (GDPR): Articles 5 and 28, Recital 26
- California Civil Code section 1798.140: definitions (sell, service provider)
