AI Agents for Lawyers

Use Cases

NDA Review at Scale with AI Agents: A Complete Workflow

AI NDA review at scale works when an agent applies a ten-position playbook, routes exceptions to a named lawyer and logs every concession. Prompt and metrics.

By Dr. Niklas SchmidtUpdated 13 November 202612 min read
On this page
  1. Why the NDA is where an agent earns its keep
  2. The workflow, end to end
  3. The ten-position playbook
  4. The prompt that runs it
  5. Exception routing and the sign-off rule
  6. Tools and honest accuracy, September 2026
  7. Metrics, and the confidentiality of the NDA itself
  8. What to do this week

NDA review at scale means an agent takes each incoming NDA from intake to a first-turn redline against a written playbook of ten positions with fallbacks, routes anything outside the playbook to a named lawyer, and logs every concession. As of September 2026 the model is the easy part; the playbook, the exception routing and the sign-off rule are what make the volume safe.

Why the NDA is where an agent earns its keep

NDAs are the highest-volume, lowest-variance paper most firms and legal departments handle: four to eight pages, the same dozen issues in a different order, rarely worth an hour of a senior lawyer's attention. Volume makes the setup pay back, low variance means a short playbook covers nearly every clause, and stakes that are real but modest mean the sign-off rule can be light without being absent.

The commercial pressure is no longer abstract. Crosby, a New York firm of about 30 lawyers plus agents, had reviewed 13,000 contracts by March 2026 and bills $250 to $1,000 per contract with no hourly element (Forbes, 31 Mar 2026). Deloitte's survey of 121 senior legal leaders expects the hourly share of external legal work to fall from 72% to 44% within two to three years (Deloitte, 30 Jun 2026). Citi now asks firms bidding for its work to disclose how much AI saves them (September 2026). An NDA that still costs four billable hours is the first line item a client will query, and the vendors know it: Anthropic's commercial-legal plugin leads with NDA review, LegalOn's Triage agent (February 2026) auto-approves low-risk contracts or escalates them, and Legora's aOS (May 2026) promises counterparty redlines "reviewed, flagged and responded to by the time lawyers arrive". The contract review page covers the general case; this page is the NDA-specific build.

The workflow, end to end

Seven stages, each with a gate the agent cannot open by itself.

  1. Intake. The NDA arrives by email, CLM or intake form. The agent extracts the counterparty, your side (discloser, recipient or mutual), the purpose and the deadline. Missing role or purpose sends the file back to the requester.
  2. Classification. A deterministic step, not a model: your template or theirs? If yours, a document compare produces the deltas and only the deltas go to the agent. Non-NDA content (a licence, an exclusivity, a non-compete) is flagged before review starts.
  3. Playbook run. The agent applies the ten-position table below clause by clause and returns a triage table, redline instructions as insertions and deletions, and a draft cover note.
  4. Routing. GREEN clauses go to the reviewer queue unchanged. YELLOW clauses get a redline drawn from the fallback wording. RED clauses stop the agent; it writes two sentences for the responsible lawyer and does not draft.
  5. Sign-off. A named lawyer reads the full redline, GREEN clauses included, accepts or rejects each change and records their name on the file. Nothing leaves without this stage.
  6. Send and log. The cover note goes out. Every concession below the preferred position is logged with the reason.
  7. Monthly review. A playbook-monitor report shows which positions were conceded and how often; the playbook version increments whenever a row changes.

Stages 2 and 5 are the ones firms skip, and they are the two that matter. Stage 2 turns a twenty-minute review of your own paper into a two-minute review of five deltas. Stage 5 keeps the workflow inside ABA Formal Opinion 512's supervision duty and the SRA's requirement that a solicitor understand a tool "well enough to review its output critically" (warning notice, 17 Aug 2026).

The ten-position playbook

The contract playbook is the asset; the tool is replaceable. The table uses Austrian defaults for governing law. Substitute your own and keep one playbook per governing law, because a position written for Delaware paper is often pointless under Austrian or German law.

#PositionPreferredFallbackEscalate (RED) when
1Definition of Confidential InformationAll non-public information disclosed for the Purpose, marked or notMarking required; oral disclosures confirmed in writing within 30 daysOral disclosures excluded, or the existence of the discussions not itself confidential
2ExclusionsThe four standard carve-outs, burden of proof on the RecipientAdd a combination clause: parts being public does not exclude the wholeA carve-out missing, or a residuals clause permitting use of information retained in unaided memory
3PurposeNarrow: evaluating the transaction named in the recital"and any resulting business relationship"No purpose limitation, or one wide enough to permit use in the Recipient's own products
4Permitted recipientsEmployees, advisers and affiliates with a need to know, bound by obligations no less strictWritten consent for subcontractors; affiliates namedFinancing sources or potential acquirers on a sell-side NDA, or no need-to-know limit
5Term and survivalAgreement 2 years; obligations 3 years from each disclosure; trade secrets for as long as they remain trade secrets5 years from disclosurePerpetual obligations with no trade-secret carve-out, or under 12 months where you disclose
6Return or destructionReturn or destroy on request and confirm in writing; one archival copy and routine backups retained, still confidentialOfficer's certificate within 30 daysNo archival or backup carve-out, or a duty to destroy records you must keep by law
7Compelled disclosurePermitted where required by law, court or regulator, with prompt notice where lawful and disclosure limited to what is requiredNotice "where legally permitted and practicable"A prohibition on disclosing even when compelled, or a duty to contest the order at your own cost
8Non-NDA obligationsNone: no non-compete, exclusivity, standstill, no-hire or IP assignment12-month non-solicitation of named key employees, general advertisements excludedAny non-compete, exclusivity, standstill or IP assignment, however worded
9RemediesInjunctive relief may be sought; no liquidated damages; no indemnityMutual acknowledgement of irreparable harmLiquidated damages, penalties, an indemnity covering consequential loss, or a waiver of the right to contest relief
10Governing law and forumAustrian law, courts of ViennaEnglish law and London courts; Delaware or New York for a US counterparty of comparable sizeAny other law, arbitration for a plain NDA, or a forum in which the other rows cannot be enforced

Two rows do most of the escalation work. Row 8 catches the non-compete that arrives dressed as "non-circumvention" in a five-page mutual NDA; row 2 catches the residuals clause that quietly turns a confidentiality agreement into a licence. Neither should ever be drafted around by an agent. The playbook-building guide shows how to derive rows like these from your last negotiated NDAs rather than from a template.

The prompt that runs it

The prompt follows the five-part structure now standard in legal prompt guides: role, task, sources, output format, constraints. Paste the playbook in as plain rows. This version is written for a general model such as Claude Opus 5 in a Project or a Skill; the same text works, lightly trimmed, as the instruction for Microsoft's Word Legal Agent or Spellbook.

ROLE
You are a contract review agent for [Firm], reviewing a non-disclosure agreement
on behalf of [Client] ("we", "our"). Our role under this NDA: [Discloser /
Recipient / Mutual]. Purpose of the underlying discussions: [one sentence].

TASK
Review the attached NDA clause by clause against the NDA playbook below. For each
clause decide whether it matches our preferred position, matches a fallback, or
triggers an escalation. Produce (1) a triage table, (2) redline instructions and
(3) a draft cover note to the counterparty.

SOURCES
Use only the attached NDA and the playbook. Do not cite statutes or case law. If a
clause raises a question the playbook does not answer, mark it YELLOW with the
note "no playbook position" rather than inventing a position.

PLAYBOOK
[paste the ten rows here: number | topic | preferred | fallback | escalate when]

OUTPUT FORMAT
1. Triage table with the columns: Clause no. | Topic | Status (GREEN / YELLOW /
   RED) | Operative words quoted from the draft | Playbook row matched or
   breached | Proposed action.
2. Redline instructions: for every YELLOW clause, the exact replacement text as
   insertions and deletions. Use our fallback wording first and our preferred
   wording only where the draft is silent. Keep the counterparty's defined terms
   and clause numbering. Do not touch GREEN clauses. Do not draft for RED
   clauses; describe the issue in two sentences for the reviewing lawyer.
3. Cover note: under 120 words, plain English, no citations, listing the changes
   in the order they appear in the document.

CONSTRAINTS
- Never accept a position the playbook does not list, and never propose wording
  more favourable to the counterparty than the fallback.
- Flag any clause that is not an NDA clause (non-compete, exclusivity, standstill,
  licence, IP assignment) as RED regardless of how it is worded.
- Flag every inconsistency between a defined term and its use, and every
  cross-reference that points to a missing or renumbered clause.
- If the governing law in the draft is neither our preferred nor our fallback law,
  stop after the triage table and mark the whole document RED.
- Do not summarise the NDA. Do not add general risk commentary. Do not answer
  questions the reviewing lawyer has not asked.

One note on running it. Anthropic's documentation says Opus 5 "verifies its own work without being told to", so the 2025 habit of appending "add a final verification step" now produces over-verification and slower runs (Anthropic, Jul 2026); leave it out. Anthropic's open-source plugin runs the equivalent review with the /commercial-legal:review command, and the Word Legal Agent takes the same instruction through the Copilot task pane.

Exception routing and the sign-off rule

The colours are the workflow's only real logic, so define them in writing.

  • GREEN: the clause matches the preferred position or a fallback. No redline; the reviewer sees it in the triage table and moves on.
  • YELLOW: the clause deviates but a fallback exists, or the playbook is silent. The agent drafts from the fallback or flags "no playbook position". A supervising associate or in-house counsel reviews the redline.
  • RED: an escalation trigger fired. The agent stops drafting that clause and routes the file to the responsible lawyer. A RED is closed only by a person, and the closure reason is logged.

This is the human-in-the-loop design the vendors themselves now ship: Harvey's agents "flag moments requiring user confirmation", Lexis Protégé's Legal Intelligence Engine "presents an executable plan for user approval" before it runs, and Anthropic's plugins call every output "a draft for attorney review". Legora's pitch of redlines answered overnight raises the obvious counter-question, which Wolf Theiss's Chief Digital & Technology Officer put plainly:

"KI-Systeme liefern Vorschläge, aber diese müssen überprüft werden." — Leonhard Wassiq, Wolf Theiss, in Kurier (2026)

AI systems deliver suggestions, but these must be checked. The checklist for the person doing the checking:

  • A named lawyer has read every tracked change, and the GREEN clauses the agent left alone.
  • Every YELLOW change carries a comment naming its playbook row; no RED has been silently downgraded.
  • Defined terms and cross-references still hold after the agent's insertions. The failures that reach counterparties are structural, not stylistic.
  • The cover note lists every change; concessions below the preferred position are logged with the reason.
  • The tool tier is recorded on the file: business or API tier under a data-processing agreement, no consumer accounts.

ABA Formal Opinion 512 (29 July 2024) frames this as the supervision duty; the Austrian bar's four practice rules put it as "the lawyer keeps final responsibility". Neither prevents an agent from working overnight. Both prevent the agent from pressing send.

Tools and honest accuracy, September 2026

ToolWhere it runsWhat it does with an NDACost, as of September 2026
Microsoft Word Legal Agent (30 Apr 2026, Frontier preview)Word; Windows and US tenants firstClause-by-clause review against a pasted playbook; a comment per clause with issue, risk and replacement wording; tracked changes that preserve lists, tables and existing changesInside Microsoft 365 Copilot at $30/user/month; requires Anthropic enabled as a subprocessor
SpellbookWord add-in; Autonomous Contract Management in early access since 30 Jun 2026Playbook redlines in Word; agents that "work autonomously to review and redline documents" via DMS and emailAbout $99–$350 per user per month; ~4,000 customers
Claude for Legal, commercial-legal pluginCowork, Claude Code, Managed Agents, Microsoft 365MSA/NDA/SaaS review to Word tracked changes; RED/YELLOW/GREEN triage; playbook-monitor agentNo extra charge on paid Claude plans; Apache 2.0
Legora aOS (7 May 2026)Platform, Word and OutlookOvernight counterparty redlines queued for lawyer reviewFrom about $3,000/user/year, 10-seat minimum (secondary source)

The Word Legal Agent was built by the team Microsoft hired from Robin AI and writes every redline through a "purpose-built insertion algorithm" (Legal IT Insider, 1 May 2026); the tool page covers the tenant prerequisites. Spellbook grounds frontier models rather than fine-tuning them. GC AI's critique of the Anthropic plugin lists three gaps that bite at NDA volume: context resets per session, no citation verification without a connector, and single-user markdown rather than team playbooks (GC AI, 2026).

On accuracy, three numbers and one caveat. Anthropic says Claude Opus 5 (24 July 2026) scored "highest of any model we tested, nearly double Opus 4.8" on first-turn NDA redlines, without publishing the absolute score (Anthropic). GC AI's vendor-run RedlineBench rose from 47.9 to 57.0 with Claude Fable 5.1 in September 2026. In the last independent head-to-head, Vals AI's February 2025 report, lawyers scored 79.7% on redlining against Harvey's 65.0%, while the tools beat lawyers comfortably on document questions and extraction (Vals AI, 27 Feb 2025). The caveat is the phrase "first turn". Every figure describes the opening redline; none was measured on your paper against your playbook. That is why the reviewer edit rate below is the only accuracy number worth reporting to a client.

Metrics, and the confidentiality of the NDA itself

Axiom's survey of 528 in-house leaders found that 83% could not show whether last year's AI spend paid off and that two-thirds run general-purpose tools in their default configuration (Axiom, 9 Jul 2026). UBS's legal team went the other way: 80 candidate use cases cut to three, with a 140% ROI hurdle (Harvey FORUM, Mar 2026). Six numbers, counted from the first NDA, settle the question either way.

MetricHow to count itWhat it tells you
Receipt-to-redline timeIntake timestamp to approved redline sentThe number clients feel, and the one Citi now asks bidders about
GREEN shareNDAs with no RED and no reviewer edits, over all NDAsWhether the playbook fits the paper you actually receive
RED rateNDAs with at least one escalationAbove half means the playbook is too narrow or the counterparties unusual
Reviewer edit rateAgent changes altered or rejected, over changes proposedYour own accuracy figure, on your paper; the only one that beats a vendor benchmark
Playbook coverageClauses with a matching row, over clauses in the draftRises as the "no playbook position" list is worked through
Concession driftPositions conceded below fallback per monthA rising count means the playbook is out of date or being bypassed

One irony deserves its own paragraph. The document under review is an agreement about confidential information, and it usually contains the counterparty's. Consumer tiers of ChatGPT, Claude and Gemini train on inputs unless the user opts out; Team, Enterprise and API tiers do not, and OpenAI retains API content for 30 days for abuse monitoring even after opt-out. The Upper Tribunal has said that pasting client material into public AI tools puts it "in the public domain" (UK v SSHD [2026] UKUT 81 (IAC)); the German federal bar's December 2024 guidance treats tools that store or forward client data for training as a breach of § 43a Abs. 2 BRAO; the Austrian bar requires client data to stay on EU servers and never be used for training. The Word Legal Agent adds a question of its own, because it requires Anthropic to be enabled as a subprocessor in the tenant. The confidentiality checklist and the vendor due-diligence page cover the settings and the contract terms; the short version is that the tier goes on the file next to the reviewer's name.

What to do this week

  • Run the ten-row table against three NDAs you negotiated this year. Where the position you actually conceded sits below the fallback, the row is wrong, not the negotiation.
  • Time the last ten NDAs from receipt to send, by hand, before any agent touches one. Without that baseline the receipt-to-redline number is a vendor's number, not yours.
  • Write the sign-off rule in one sentence and put a name against it. "No NDA redline leaves without [role] having read it in full" is enough.
  • Pick the tier before the tool: business or API, under a data-processing agreement, training off. Then choose between Word, a platform and the open-source plugin according to where your reviewers already work.

Module 4 of the AI Lab for Lawyers course covers custom systems, agents and automation with n8n, Make and Zapier, which is where the intake-to-sign-off loop on this page gets wired together. The comparison of agents and CLM software is the next read if the open question is whether this belongs inside a contract-management system or beside it.

Frequently asked questions

Can AI review an NDA without a lawyer?

It can produce the first-turn redline without a lawyer, and several tools now do so overnight. It cannot sign off. ABA Formal Opinion 512, the SRA's August 2026 warning notice and the Austrian bar's practice rules all keep final responsibility with the lawyer, so every outgoing NDA redline should carry a named reviewer who has read it in full.

How many positions does an NDA playbook need?

Ten is enough for a standard mutual NDA: definition and exclusions, purpose, permitted recipients, term and survival, return or destruction, compelled disclosure, non-NDA clauses such as non-competes, remedies, and governing law. Each row needs a preferred position, a fallback and an escalation trigger written in plain sentences, because the agent applies the row exactly as written.

Which AI tools redline NDAs against a playbook in 2026?

As of September 2026: Microsoft's Word Legal Agent, Spellbook, Anthropic's open-source Claude for Legal commercial-legal plugin, Legora's aOS, LegalOn's Triage and Playbook agents, Ironclad's Redlining Agent and Harvey's playbook review. They differ in where they run, whether they write genuine tracked changes and what they cost per seat, from about $30 to over $1,000 a month.

How accurate is AI NDA redlining?

Good enough for a first draft of the redline, not good enough to send unread. Anthropic reports Claude Opus 5 nearly doubled Opus 4.8 on first-turn NDA redlines without publishing an absolute score; GC AI's RedlineBench reached 57.0 in September 2026; and in Vals AI's February 2025 test lawyers scored 79.7% on redlining against 65.0% for Harvey.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion