← Back to HAQQ Blog

How In-House Teams Automate NDA Review

By Stephane Boghossian · · 8 min read · Guides

Automate NDA review in three layers: a self-service template, a written playbook with fallbacks, and AI first-pass review a lawyer signs off. With the 10-clause checklist and GCC points.

Why do in-house teams automate NDAs first?

NDAs come in high volume, follow a familiar structure and rarely carry the deal's main risk, so they absorb lawyer time out of proportion to their value. In a 2023 survey of 400 private-markets professionals, including in-house lawyers, about 65% said they personally spend six or more hours a week negotiating NDAs, and 17% more than ten. The survey was commissioned by Ontra, a contract-automation vendor, so treat it as a signal rather than a census.

The time per document is real too. In a 2018 study run by the vendor LawGeex, 20 US-trained lawyers took an average of 92 minutes to review five NDAs, between 51 and 156 minutes each lawyer. That study measured only how well issues were spotted, not how they were negotiated.

What should every NDA review check?

A law-firm checklist published on Legal 500 in 2024 lists ten provisions. They are the backbone of any playbook:

Step 1: Put low-risk NDAs on self-service

The fastest NDA review is the one that never reaches legal. Approve one mutual NDA template, publish it where sales, procurement and business development can reach it, and let them send it without legal review as long as nobody edits the text. Legal sees only the NDAs that arrive on the other side's paper or come back with changes.

Step 2: Write the playbook

A playbook turns a lawyer's judgement into rules someone else, or a tool, can apply. For each clause it records your standard position, the fallback you will accept without escalation, and the point at which a lawyer must decide. The positions below are examples to adapt, not recommendations for your company.

ClauseStandard positionAcceptable fallbackEscalate when
DirectionMutualOne-way if only we discloseWe receive but do not disclose
Term2 years from signatureUp to 5 yearsPerpetual, except trade secrets
Carve-outsAll four standard carve-outsDrop 'independently developed' only with record-keepingAny standard carve-out removed
RepresentativesEmployees, advisers, affiliatesNamed affiliates onlyNo advisers allowed
RemediesInjunction availableInjunction subject to court discretionLiquidated damages or penalties
Non-solicitNone12 months, named employees onlyAny non-compete
Governing lawOur home jurisdictionNeutral seat (e.g. DIFC Courts)Counterparty's home courts

Step 3: Let AI do the first pass, and keep the lawyer on the redline

Give the tool the counterparty's NDA and your playbook, and ask it to compare the two clause by clause, list each deviation ranked by risk, and propose language that brings the clause back to your standard or fallback. The lawyer then reads the ranked list rather than the whole document.

The independent evidence supports splitting the work this way. The Vals Legal AI Report (February 2025) tested tools against a lawyer baseline. On data extraction, the lawyer baseline was 71.1% and tools scored between 64.0% and 75.1%, with two of them beating the lawyers. On redlining, lawyers scored 79.7%, Harvey 65.0% and Vincent AI 53.6%, and the report concluded that legal AI tools should only be used for certain types of redlining tasks. Finding what is in a contract is where tools are strong; deciding how to change it is still the lawyer's job.

Is it safe to upload an NDA to an AI tool?

An NDA under negotiation is itself confidential, and it often names the deal it protects. Before you route NDAs through any tool, check where the text is processed, whether the provider trains models on it, who inside your company can see the history, and whether personal data leaves your region. In the United States, ABA Formal Opinion 512 (July 2024) tells lawyers to get informed client consent before putting client information into self-learning generative AI tools and to verify what those tools produce. It does not bind GCC lawyers, but it is a sensible baseline for any in-house policy.

What should GCC teams build into the playbook?

What should you measure?

Track four numbers from the first week: median days from receipt to signature, the share of NDAs signed on self-service, the share escalated to a lawyer, and which clauses trigger escalation most. The last one tells you where the playbook needs a better fallback. These are the same numbers a business case for legal AI needs, covered in our guide to the legal AI business case.

Where HAQQ fits

Disclosure: we build HAQQ. In HAQQ Legal AI, you upload an NDA as a PDF or Word file and ask for a review against your playbook. Every flagged clause comes back ranked High, Medium or Low, with why it matters in the governing jurisdiction and a suggested redline as tracked changes you can accept, edit or reject, or ask to make softer or firmer. Keep the playbook and your approved template in a Project so every review there can read them. Before a conversation reaches an external AI model, names, emails, addresses and company names are replaced with stand-ins and restored in the answer. A lawyer still approves anything outside the playbook.

Key takeaways

Sources

FAQ

How do in-house legal teams automate NDA review?

In three layers: an approved self-service NDA template the business can send without legal review; a written playbook with a standard position, a fallback and an escalation point for each clause; and AI first-pass review of counterparty NDAs against that playbook, with a lawyer approving anything outside it.

Can AI review an NDA accurately?

AI is strong at finding what is in a contract and weaker at deciding how to change it. In the Vals Legal AI Report (February 2025), tools scored 64.0% to 75.1% on data extraction against a lawyer baseline of 71.1%, but on redlining lawyers scored 79.7% and Harvey 65.0%. Use AI for the first pass and keep a lawyer on redlines.

What should an NDA playbook include?

For each clause (direction, definition, carve-outs, representatives, required disclosure, return or destruction, remedies, term, governing law, restrictive covenants), the playbook records your standard position, the fallback you accept without escalation, and when a lawyer must decide.

Is it safe to upload an NDA to an AI tool?

Only after checking where the text is processed, whether the provider trains models on it, who can see the history, and whether personal data leaves your region. ABA Formal Opinion 512 (2024) asks US lawyers to get informed client consent before using self-learning generative AI with client information and to verify the output.

How long should an NDA last?

Many NDAs set a term of two to five years, with longer protection for trade secrets. In Saudi Arabia, the Civil Transactions Law provides that termination of a contract does not end an obligation to maintain confidentiality.