Legal Agents: Contract Review and Discovery Support
Legal work is document-heavy, deadline-driven, and full of pattern matching across clause libraries and large corpora. Agents help with review acceleration and packet assembly. They do not replace licensed legal judgment, client relationships, or court filings.
Use this page as a scoping cheatsheet for contract review and discovery-support agents.
How to Use This List
- Start with the quick map, then pick a pattern table for contracts or discovery.
- Treat every "execute" cell as human-owned unless your firm policy explicitly says otherwise.
- Pair with Regulated-Industry Agent Checklist: Compliance, Audit, and Human Sign-Off before production.
- Revisit privilege and confidentiality controls when data sources expand.
Quick Map
| Job to be done | Agent fit | Typical output | Human owns |
|---|---|---|---|
| First-pass clause flagging | Strong | Issues list + clause cites | Negotiation position |
| Playbook deviation check | Strong | Redlines suggestions | Accept/reject edits |
| Large corpus responsiveness review | Strong assist | Ranked docs + rationales | Privilege & production calls |
| Privilege log drafting assist | Medium | Draft entries | Final log accuracy |
| Legal advice to client | Poor alone | N/A | Counsel |
| Autonomous court filing | Premature / high risk | N/A | Counsel + process |
Contract Review Cheatsheet
| Capability | What good looks like | Failure mode |
|---|---|---|
| Clause extraction | Span-level cites into the source PDF/DOCX | Paraphrase without location |
| Playbook compare | Explicit rule IDs from firm playbook | Generic "looks risky" vibes |
| Risk ranking | Consistent severity rubric | Everything is "critical" |
| Missing clause detect | Checklist of required clauses | Silent omission |
| Counterparty paper intake | Structured metadata + anomalies | Assumes standard forms |
| Draft markup | Track-change style suggestions | Overwrites without review |
| Obligation extract | Parties, dates, triggers table | Missed renewals/notice windows |
| Cross-ref consistency | Flags defined-term conflicts | Ignores definitions section |
Contract agent tool surface (typical)
| Allow | Avoid by default |
|---|---|
| Read document store | Send email to counterparties unsupervised |
| Search playbook/clause library | Change DMS permissions broadly |
| Create review checklist tasks | File signed contracts as final without counsel |
| Draft redline package for lawyer | Provide client-facing "legal advice" bot answers |
| Cite page/section anchors | Train public models on confidential deal terms |
Contract review workflow skeleton
- Ingest and classify agreement type.
- Run playbook + risk extractors.
- Produce issues list with severity and cites.
- Lawyer prioritizes and negotiates.
- Agent assists on next-turn redlines only as requested.
- Final execution stays in human-controlled signature process.
Discovery Support Cheatsheet
| Capability | What good looks like | Failure mode |
|---|---|---|
| Early case assessment | Theme clusters + key custodian hints | Over-confident case theory |
| Responsiveness ranking | Calibrated scores with examples | Random high scores |
| Issue coding assist | Consistent labels + rationales | Label drift across batches |
| Privilege candidate flag | High recall on privilege signals | Privilege miss (severe) |
| Dedup/near-dedup assist | Transparent similarity basis | Hidden drops of unique docs |
| Chronology build | Event timeline with doc IDs | Invented events |
| Deposition prep packet | Linked exhibits + questions draft | Missing adverse docs |
| Production QC assist | Spot-check anomalies | Rubber-stamp QC |
Discovery non-negotiables
- Privilege and confidentiality controls before broad retrieval.
- Matter-scoped access - no cross-matter leakage in shared indexes.
- Defensibility - log prompts, models, versions, and sampling methodology.
- Human final calls on privilege, confidentiality designations, and production sets.
- Legal hold integrity - agents must not delete or alter held data.
Decision Cheatsheet: Build, Buy, or Wait
| Situation | Lean |
|---|---|
| High volume NDA/vendor paper with stable playbook | Build or configure specialized review agent |
| Complex bet-the-company M&A | Human-led; agent as secondary pass only |
| Multi-terabyte discovery with existing review platform | Integrate assist models into platform, do not shadow-IT a free agent |
| No playbook, no gold labels, no review protocol | Wait - you cannot eval quality |
| Client forbids cloud processing of matter data | Private deployment or do not use LLM agents on that matter |
Risk and Ethics Cheatsheet
| Risk | Guardrail |
|---|---|
| Unauthorized practice framing | Position as lawyer assist, not public legal advice product (as applicable) |
| Confidentiality breach | Matter isolation, DLP, approved processors only |
| Hallucinated case law | Force retrieval from approved research tools; verify cites |
| Privilege waiver | Strict access; careful outbound drafts |
| Bias in responsiveness models | Sample-based QC, diverse seed sets |
| Over-reliance by juniors | Training + mandatory senior review thresholds |
Autonomy Ladder for Legal Agents
| Level | Agent may | Example |
|---|---|---|
| L0 | Search only | Find clause examples |
| L1 | Draft internal memo for lawyer | Issues list |
| L2 | Suggest redlines in review UI | Markup proposals |
| L3 | Batch-code with sampled QC | Discovery first pass |
| L4 | External send / file | Generally blocked without firm workflow + counsel action |
Most teams should live in L1-L3 with sampling and sign-off.
FAQs
What is the best first legal agent use case?
Playbook-based contract issue spotting on a high-volume, lower-severity agreement type with span-level citations and lawyer review.
Can agents replace junior associate review entirely?
No. They can compress first-pass time. Privilege, strategy, and client advice remain human responsibilities.
How do I reduce hallucinated legal citations?
Do not let the model free-generate authorities. Require retrieval from approved research systems and display the retrieved cite text for verification.
What makes discovery assist defensible?
Documented methodology, versioned models/prompts, seed/control sets, sampling QC, and human decisions on privilege and production.
Should contract and discovery share one agent?
Usually no. Different corpora, risk, and tool permissions. A shared platform can host separate matter-scoped agents.
How should redlines be presented?
As suggestions tied to playbook rules and source spans, in a review UI that requires accept/reject - not silent file overwrite.
What about client-facing "ask our AI lawyer" bots?
High risk for incorrect advice and confidentiality. Prefer internal assist tools unless product, ethics, and insurance review explicitly approve a consumer offering.
Do we need gold-labeled contracts to start?
For serious quality, yes - at least a seed set of lawyer-labeled issues. Without labels you cannot measure false negatives on missing clauses.
How do multi-agent setups help legal work?
Split extractors (dates, parties, liability, IP) with a merger that de-dupes issues. Keep a single privilege-sensitive corpus boundary.
What logs are sensitive in legal agents?
Prompts may contain confidential or privileged text. Apply matter-level retention, access control, and approved storage - not casual SaaS chat history.
When is a legal agent premature?
When there is no playbook owner, no review sampling plan, cloud processing is forbidden but only public models are available, or the goal is unsupervised legal advice.
How does this differ from general research agents?
Legal agents prioritize citation fidelity, privilege, matter isolation, and professional responsibility over broad web synthesis.
Related
- Why Industry Context Changes an Agent's Risk Profile - stakes and duties
- Financial Services Agents: Fraud Triage, Reporting, and Trading Support - regulated assist parallels
- Regulated-Industry Agent Checklist: Compliance, Audit, and Human Sign-Off - launch checklist
- Industry-Specific Use Cases Best Practices - ten practices
- Research Agents: Multi-Source Investigation and Synthesis - general research contrast
- Industry-Specific Use Cases Basics - cross-industry intro
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.