Using AI to Review an NDA: A Solo Attorney's Walkthrough

Step 1: Get the structural facts on the table

Before assessing whether any individual clause is favorable, establish what kind of agreement you're looking at. AI is efficient here because these are extractable facts, not judgment calls:

• One-way or mutual? Does the confidentiality obligation run in one direction (usually protecting the disclosing party) or both? This single fact changes who the agreement is really protecting.
• Duration. How long do confidentiality obligations last after disclosure, and separately, how long does the agreement itself run? These are sometimes different clauses and worth confirming they match your expectation.
• Governing law and forum. Which state's law governs, and where would a dispute be litigated? Routine to check, easy to skip on a fast read.

Ask the tool to extract and summarize these three facts first. It's a fast pass that orients the rest of the review, and it's exactly the kind of bounded extraction task a document analysis tool handles well.

Step 2: Scrutinize the definition of "confidential information" — and its carve-outs

This is the clause that does the most work in the entire agreement, and it's the one most likely to get skimmed. Ask specifically:

• What does the agreement define as confidential? A narrow definition (specific documents marked "confidential") protects less than a broad one (anything disclosed in connection with the discussion, marked or not).
• What carve-outs exist? Standard exclusions cover information that's already public, independently developed, or already known to the receiving party before disclosure. Confirm these are present and not unusually narrow — an NDA missing standard carve-outs is worth a second look, not necessarily a red flag, but a deliberate choice to confirm was intentional.
• Is there a residuals clause? A residuals clause lets the receiving party retain and use information that stays in employees' unaided memory. If present, it can significantly narrow the practical protection the rest of the agreement appears to offer — this is the clause most likely to surprise a client who assumes the NDA fully protects their information.

Have the tool pull the full confidentiality definition and every carve-out verbatim, side by side, so you're comparing exact language rather than trusting a paraphrase on a clause this consequential.

Step 3: Surface remedies and enforcement mechanics

Next, extract what happens if the agreement is breached:

• Injunctive relief language. Many NDAs state that monetary damages are inadequate and the disclosing party is entitled to seek injunctive relief without posting a bond. Confirm whether this language is present, since it affects how quickly a breach can practically be stopped.
• Attorney's fees. Does the prevailing party recover fees, or does each side bear its own regardless of outcome? A meaningful economic term that's easy to miss on a skim.
• Survival. Do the confidentiality and remedy provisions survive termination of the broader relationship (if the NDA is part of a larger deal), or do they lapse with it?

These are, again, extractable — the tool can pull each provision and summarize it plainly, freeing your attention for step four.

Step 4: Apply judgment — this is the part AI doesn't do

Everything above gets you a clear, organized picture of what the NDA says. It does not tell you whether those terms are right for this client, in this deal. That judgment depends on facts no document contains: your client's negotiating leverage, how much they actually need the counterparty relationship, whether this is a one-off exchange or an ongoing arrangement, and their tolerance for asymmetric protection.

A one-way NDA that heavily favors the counterparty might be completely appropriate if your client is the smaller party receiving access to the other side's proprietary information and has no confidential information of their own at stake. The same terms might be a real problem if your client is disclosing sensitive information and getting asymmetric protection in return. AI can tell you the agreement is one-way. Only you, with the client relationship and deal context, can tell whether that's acceptable.

This is a deliberate design point, not a gap to be filled later: a document analysis tool is built to extract, organize, and summarize what's on the page, flagging what's worth your attention — not to render a legal judgment about whether terms are advisable. Treat its output as a fast, thorough first pass that gets you to the judgment call faster, with less risk of missing something on a rushed read, not as the judgment itself.

A short worked example

Say a client sends over a one-way NDA from a prospective vendor before a discovery call. A fast AI-assisted pass might surface: the agreement is one-way, protecting only the vendor's disclosures; confidentiality obligations last three years from disclosure; the confidential-information definition is broad ("any information disclosed, whether marked or not"); standard carve-outs are present; there's no residuals clause; and the agreement provides for injunctive relief plus fees to the prevailing party.

That summary alone answers most of the "what does this say" question in under a minute of your time. The judgment call is still yours: is a three-year one-way obligation reasonable for a first discovery call, or is that duration disproportionate to the actual sensitivity of what's likely to be discussed? Does your client have any information of their own at stake that a one-way agreement fails to protect? Those questions depend on the deal, not the document — which is exactly why the tool surfaces the facts and stops there.

Common mistakes when using AI for NDA review

A few patterns are worth avoiding regardless of which tool you use:

• Treating the summary as the review. A fast, accurate extraction of terms is the input to your review, not a substitute for reading the clauses that matter most yourself, especially the confidentiality definition and any residuals clause.
• Skipping the carve-outs because the summary looked complete. Carve-outs are exactly the kind of detail that's easy to omit from a good-looking summary if you don't specifically ask for them. Always confirm the standard exclusions are present and check the exact language.
• Assuming "mutual" means "symmetrical in practice." A mutual NDA can still be lopsided if one party is realistically only ever going to be the discloser. Structure and practical effect aren't always the same thing.
• Skipping the human pass on a "simple" NDA. The simplest-looking agreements are often the ones drafted by a party who's used the same boilerplate across many deals without tailoring it — worth a deliberate check, not an assumption of harmlessness.

Using this as a template for other contract types

The four-step structure here — extract structural facts, scrutinize the definitional clause that does the most work, surface remedies and enforcement, apply judgment last — generalizes past NDAs. The specific clauses change (a lease has rent escalation and maintenance obligations instead of confidentiality definitions and carve-outs), but the shape of the review doesn't: let the tool do the fast, bounded extraction across the whole document, then spend your judgment where it actually belongs, on the terms that need a lawyer's read of the client's specific situation.

Privileged's contract-review workflow is built around this shape — matter-based, on-device, structured to extract and flag the terms worth your attention on a document like an NDA, without ever claiming to render the judgment call itself.