Rights and contracts
NDA for sharing data samples: clauses to include
By SourceX Editorial · Reviewed by Noah Loul ·
Short answer
An NDA for sharing data samples should go beyond standard confidentiality. It should limit use to evaluation, prohibit training or fine-tuning on the sample, ban re-identification and reverse engineering, set a deletion deadline with written certification, and switch off any residuals clause. Share only prepared samples, with personal and confidential details already removed.
Key takeaways
- A standard mutual NDA protects against disclosure but rarely stops a recipient from training on a sample.
- Strike or narrow residuals clauses, which can let reviewers keep and use what they remember.
- Define derived materials, such as embeddings, statistics and synthetic data, and require their deletion too.
- An NDA is a backstop; preparing the sample properly is the first protection.
- Signing an evaluation NDA commits neither side to a license.
Why a standard NDA is not enough for data samples#
A standard NDA is not enough for data samples because it guards against disclosure, while the main risk with a sample is use. An AI developer evaluating your support tickets or code reviews could keep them, train on them or fold them into an internal benchmark without ever disclosing them to anyone.
Many NDAs also contain terms that work against a data supplier: residuals clauses, broad carve-outs for information the recipient independently developed, and no fixed deletion date. A sample-specific agreement, often called a data evaluation agreement, closes those gaps.
There is also a legal reason to formalize the terms. Under the Defend Trade Secrets Act, information is a trade secret only if its owner has taken reasonable measures to keep it secret, and DOJ guidance lists confidentiality agreements and need-to-know access among such measures. Handing over operational records with no signed use restrictions can weaken that position.
The agreement can be a short standalone document or an addendum to an NDA you already have. Either way, it should name the sample, the evaluation purpose and the people allowed to handle it.
Clause checklist for a data sample NDA#
The clause checklist below covers the terms that matter most when sharing a data sample. The wording column is an outline to adapt with counsel, not finished contract language.
The re-identification clause can carry legal weight as well as practical protection. Under the CCPA as amended by the CPRA, information counts as deidentified only if the business publicly commits not to re-identify it and contractually obligates any recipient to comply with the same requirements. If you treat a California-linked sample as deidentified, the evaluation agreement is where that recipient obligation lives.
| Clause | What it should say | Why it matters |
|---|---|---|
| Definition of sample data | Identifies the files, record types and delivery date, and covers all copies and extracts | Prevents disputes over what the obligations reach |
| Evaluation-only purpose | Use limited to assessing quality and fit for a possible license | Separates evaluation from any productive use |
| No training or tuning | No training, fine-tuning, distillation, benchmark building or retrieval index from the sample | The central protection, and the one standard NDAs omit |
| Derived materials | Embeddings, statistics, labels, summaries and synthetic data from the sample are covered and deleted | Stops value leaking through outputs rather than files |
| No re-identification | No attempt to identify people, customers or clients, and no linking with other datasets | Protects the people in the records and your client relationships |
| No reverse engineering | No attempt to reconstruct removed content or infer internal systems beyond the evaluation | Protects redactions and internal know-how |
| Access limits | Named individuals on a need-to-know basis; no subcontractors without written consent | Keeps the sample inside a small, accountable group |
| Security and tools | Stored in a controlled environment; not uploaded to outside AI tools or services | Closes a common accidental leak route |
| Deletion deadline | All copies and derived materials deleted by a stated date or on request, with written certification | Gives the obligation an end point you can check |
| No residuals | Any residuals clause does not apply to sample data | Residuals language can permit use of what reviewers remember |
| No license, no obligation | No rights granted beyond evaluation; neither party must proceed | Keeps commercial terms for the license itself |
| Remedies and survival | Injunctive relief, prompt notice of breach, obligations survive termination | Lets you act quickly if something goes wrong |
What to send, and what to hold back#
What goes into a sample is a decision that comes before the NDA: send the smallest prepared extract that answers the buyer's questions. A good sample shows structure, linkage and quality, not volume.
Check the confidentiality you already owe others before choosing records. Customer MSAs, vendor NDAs and client engagement letters often define confidential information broadly enough to cover counterparties' names, pricing, configurations and technical details that appear inside your tickets and project notes. Those details come out of the sample, or the record stays home. A sample prepared to the same standard as the eventual delivery also gives the buyer an honest view of the data, which reduces surprises in later diligence.
- Pick records that show a full workflow, such as a ticket linked to the engineering fix and the customer outcome.
- Remove names, emails, phone numbers, addresses and account identifiers, and check free-text fields by hand.
- Scan code, logs and attachments for credentials, API keys and tokens.
- Exclude client-owned material, privileged documents and anything under a third-party confidentiality duty.
- Include a field dictionary and a short note on how records were selected.
- Keep an exact copy of what was sent, with a file listing or hashes, so deletion can be checked against it.
Mutual or one-way?#
A data sample NDA is usually one-way in substance even when drafted as mutual, because the supplier carries most of the risk. Buyers may share evaluation methods or a data specification they want protected, and a mutual form handles that, but the sample-specific clauses should attach only to the supplier's data.
Watch for buyer paper that imposes identical light obligations on both sides and omits the training restriction. If a buyer insists on its own form, a short addendum with the sample clauses, stated to prevail for sample data, is often easier to agree than redlining the whole document.
Clauses buyers push back on, and workable compromises#
The clauses buyers most often push back on are deletion of derived materials, officer-level certification and named-individual access, and each has a workable middle ground. The aim is to keep the protection while removing friction that does not reduce your risk.
A buyer may need to keep its evaluation conclusions for internal approval. A reasonable compromise lets it retain a written assessment that contains no records, excerpts or derived datasets. Certification can come from a responsible manager rather than an officer, and access can be limited to a defined team or role list that the buyer updates on notice. Hold firm on the no-training clause and the residuals carve-out, which are the core of the agreement.
Standard NDA versus data evaluation agreement#
The difference between a standard NDA and a data evaluation agreement is what each controls: a standard NDA governs who learns the information, while an evaluation agreement also governs what the recipient may do with it.
| Topic | Typical standard NDA | Data evaluation agreement |
|---|---|---|
| Main obligation | Do not disclose | Do not disclose, and use only to evaluate |
| AI training | Usually silent | Expressly prohibited, including fine-tuning and benchmarks |
| Derived outputs | Usually silent | Covered and deleted with the sample |
| Residuals | Often permitted | Excluded for sample data |
| End of obligations | Return or destroy on request | Firm deletion date with certification |
| Re-identification | Not addressed | Prohibited |
Illustrative: an engineering consultancy shares an RFI sample#
Illustrative: a fictional civil engineering consultancy is asked for a sample of its RFI and submittal review records. The records live in Procore and Bluebeam, with internal review comments kept in Deltek project notes.
The managing principal sends the buyer's NDA to counsel, who finds a residuals clause and no mention of training. Counsel adds a data evaluation addendum with a no-training clause, coverage of derived materials, a deletion deadline and named reviewers. The firm sends a small extract from projects where it holds clear rights, with client names, site addresses and engineer names removed. After the evaluation window, the buyer certifies deletion and the parties move to license terms.
How SourceX handles samples#
SourceX handles samples within the Preparation and Approval steps of the SourceX five-step transaction, after Supply and Rights have been reviewed. Nothing is shared during the initial assessment; a sample is prepared only when a supplier decides to go further, and the supplier approves exactly what is released.
The SourceX Evidence Packet records what was shared, the privacy record for the sample and the release authorization, so a supplier can check any deletion confirmation against a documented file list.
Frequently asked questions
Do we need a new NDA if we already have one with the buyer?
Often an addendum is enough. Check whether the existing NDA covers data samples, prohibits training, addresses derived materials and excludes residuals. If any are missing, add a short sample-specific addendum that prevails for the sample data rather than renegotiating the whole agreement.
How large should a data sample be?
Large enough to show structure, linkage and quality, and no larger. Buyers usually want representative records across time periods and record types rather than volume. Agree the scope in writing before preparing it, and decline requests for full exports at the evaluation stage.
Can the buyer evaluate the sample in our environment instead?
Yes, and some suppliers prefer it. A controlled workspace where the buyer runs queries without downloading data reduces reliance on deletion promises. It takes more setup, so it tends to suit sensitive record types more than routine ones.
What if the buyer has already trained on the sample by mistake?
Bring in counsel, use the breach notice and remedies clauses, and ask for a written account of what was done and with which models. Deletion of derived materials and affected model checkpoints is the usual demand. Your position depends on how clearly the agreement prohibited training.
Can we use this checklist as our contract?
No. The checklist describes common clause topics, not enforceable language. Enforceability, remedies and drafting conventions differ by state and by deal, so have counsel draft or review the final wording for your situation and your buyer.
Sources
- Under 18 U.S.C. 1839(3), information qualifies as a trade secret only if the owner has taken reasonable measures to keep it secret and it derives independent economic value from not being generally known to, and not readily ascertainable through proper means by, another person who can obtain economic value from its disclosure or use. Source
- DOJ guidance states that trade secret protective measures need not be absolute but must be reasonable under the circumstances, citing examples such as advising employees of the trade secret's existence, limiting access on a need-to-know basis, requiring confidentiality agreements, and keeping documents locked. Source
- Post-CPRA, Cal. Civ. Code 1798.140(m) treats information as deidentified only if the business takes reasonable measures to ensure it cannot be associated with a consumer or household, publicly commits to maintain and use it in deidentified form and not attempt to reidentify it, and contractually obligates any recipients to comply with these requirements. Source
Related resources
- QuestionShould companies sell or license their data?
- QuestionDo AI labs buy code?
- InsightCan a distributor license its pricing and quote history?
- InsightWhat permitted uses should a code license allow: training, evaluation or RL environments?
- InsightVendor AI training vs licensing your own data: who captures the value?
- IndustryBPO & contact centers data
See if your company qualifies
A short company assessment. No data uploads are needed.