Types of NDA
Choose the structure for the deal. A unilateral NDA fits one-way disclosure - employer to employee, company to investor, licensor to licensee. A mutual NDA suits joint ventures, co-development, and M&A where both sides share data. A confidentiality clause embedded in a wider agreement differs from a standalone NDA: in technology licensing it should prohibit reverse engineering and confirm that the confidentiality obligation grants no IP licence; in SaaS, the supplier should sign a separate NDA before sales-process disclosures, since a clause in the later subscription agreement only protects post-signature information.
M&A NDAs are the most heavily negotiated, with an expansive evaluation-material definition, non-disclosure of the existence of discussions, a use restriction limited to evaluating the possible transaction, and pre-agreed irreparable-harm language.
Non-use and the permitted purpose
The non-use obligation restricts use to a defined permitted purpose, and it is one of the most litigated NDA provisions. A vague purpose (in connection with the parties' business relationship) is dangerous twice over: it lets the recipient exploit the information in unintended ways, and it makes proving breach nearly impossible. Define the purpose narrowly and specifically.
Use restrictions can have surprising force. In Martin Marietta Materials v Vulcan Materials Co (Del 2012), a bidder that used evaluation material obtained under an NDA - whose use was limited to evaluating a possible combination between the parties - to launch a hostile bid was held to have breached the NDA, and was enjoined for four months. A tight use clause can operate as a backdoor standstill, which bidders must negotiate around expressly.
Return, destruction, and residual knowledge
Complete electronic deletion is largely a fiction - information persists across devices, cloud services, email, and backups - so return/destruction clauses now permit standard carve-outs: retention on automatic backups (overwritten in the ordinary course), retention required by law, and one archival copy held by legal counsel, all subject to continuing confidentiality. Require written certification of destruction within a set period.
Residual knowledge clauses let the recipient use information retained in its people's unaided memory. Technology companies press for these because memories cannot be selectively erased, but a broad residuals clause can let a competitor replicate innovations from memory. Mitigate by limiting it to genuinely unaided memory (excluding deliberate memorisation), excluding patents, copyright, and trade secrets, restricting which personnel can rely on it, and giving the most sensitive information no residuals carve-out at all.
Non-circumvention
Non-circumvention clauses stop one party bypassing the other to deal directly with an introduced third party. Between commercial parties of comparable bargaining power they are generally enforceable as a matter of freedom of contract; in an employment or agency context they may be assessed under the restraint-of-trade doctrine and must go no further than reasonably necessary; and perpetual non-circumvention obligations are likely unenforceable under either framework.
Conditional wording is a trap. In Goodrich Capital v Vector Capital (SDNY 2012), a private-equity firm that used NDA information to acquire a target the adviser had introduced, bypassing the adviser, faced liability under the NDA's use and non-circumvention clauses - illustrating both the risk of narrow, conditional non-circumvention language and the value of a tight use clause alongside it.
M&A clean teams and the AI-training risk
Two modern points. In M&A between competitors, a clean team adds a tier of protection for competitively sensitive information: designated members (typically outside advisers) access a restricted data room and produce only aggregated, anonymised reports after antitrust-counsel review, preventing the assumption that exchanged information will be used strategically. And in any technology deal, the NDA should now expressly prohibit using confidential information to train, fine-tune, or improve AI or machine-learning models, prohibit inputting it into public generative-AI tools, and define derivative data (embeddings, indexes, logs, model parameters) that could encode it.
Traditional NDA language was written for human disclosure; AI training is a new channel of leakage that standard wording does not capture. Address it expressly, and exclude AI systems trained on the information from any residual-knowledge carve-out.
Use at the desk
Practical checklist
- Match the NDA type to the deal; in SaaS, sign a separate NDA before sales-process disclosures.
- Define the permitted purpose narrowly - a vague purpose makes breach hard to prove (Martin Marietta v Vulcan (Del 2012)).
- Allow practical return/destruction carve-outs (backup, legal, archival) with written certification.
- Limit residual-knowledge clauses to unaided memory and exclude patents, copyright, and trade secrets.
- Draft non-circumvention for commercial parties carefully; avoid perpetual or conditional wording (Goodrich v Vector (SDNY 2012)).
- For competitor M&A use a clean team; in tech deals expressly prohibit AI training on confidential information.
This guide is informational only and is not legal advice. It does not replace advice from licensed counsel on the facts of a specific transaction.
Product demo
Use the guide for context. Use Veqtor for the Word documents.
Watch Claude compare negotiation drafts and create a separate Word document with proposed tracked changes.
See Veqtor work with Word redlines