How AI Agents Handle Contract Review for Legal Teams

A mid-market legal team reviewing contracts manually follows the same pattern everywhere. A contract arrives. A junior associate reads it start to finish. They compare it against the company’s template. They flag deviations. They write a summary. A senior attorney reviews the flags and makes decisions. For a team processing 50-100 contracts per month, this works. At 200+, the pipeline breaks. Associates start skimming. Deviations get missed. Turnaround times stretch from 2 days to 2 weeks.

We build agent systems that handle the first pass of contract review: intake, extraction, comparison, and risk scoring. The agents don’t negotiate. They don’t make legal judgments. They process the volume and surface the issues so lawyers spend their time on the parts that require a law degree.

The Manual Process and Where It Fails

We documented the contract review workflow at three LATAM companies before building the agent system. The companies ranged from 80 to 350 contracts per month across vendor agreements, service contracts, employment agreements, NDAs, and partnership terms. All operated in multiple LATAM jurisdictions, meaning contracts arrived in Spanish, Portuguese, and English, governed by different countries’ laws.

Step 1: Intake. Contracts arrive through email, shared drives, messaging apps, and sometimes physical paper that someone scans. There’s no single intake point. A procurement manager forwards a vendor contract. General counsel receives a partnership agreement directly. HR sends employment contracts from a different country. The first problem is that no one has a complete picture of how many contracts are in the review queue at any given time.

Step 2: Initial read. A junior associate reads the contract. For a standard 15-page vendor agreement, this takes 30-45 minutes. For a complex 40-page partnership agreement in a foreign jurisdiction, it takes 2-3 hours. The associate is looking for deviations from the company’s standard terms, unusual clauses, risk factors (unlimited liability, unfavorable governing law, broad IP assignment), and missing protections (no limitation of liability, no termination for convenience, missing data protection clauses).

Step 3: Comparison. The associate compares the contract against the relevant template. For a vendor agreement, that means checking 20-30 standard clauses against the company’s approved template. This is where the most errors occur. Associates miss deviations because they’re tired, because the contract uses different terminology for the same concept, or because a clause is structured differently from the template while having the same legal effect.

Step 4: Summary and flagging. The associate writes a memo listing deviations, risks, and recommended changes. For a standard contract, this takes 20-30 minutes. The memo goes to a senior attorney.

Step 5: Senior review. A senior attorney reviews the memo and the flagged sections of the contract, makes decisions on which deviations are acceptable, and either approves the contract or sends back revision requests.

Total time per contract: 1.5-4 hours depending on complexity. At 200 contracts per month, that’s 300-800 hours of attorney time. For a legal team of 4 (1 senior attorney, 2 associates, 1 paralegal), contract review consumes 50-70% of total capacity, leaving little time for strategic legal work, negotiations, or proactive risk management.

The error rate we measured across the three companies averaged 12% of contracts with at least one missed deviation that was later caught during negotiations or, worse, after execution. The errors weren’t random. They clustered around three patterns: fatigue-related misses in high-volume periods, cross-language comparison errors (Portuguese contract compared against Spanish template), and structural variations where a clause was present but organized differently than expected.

The Agent Pipeline

The contract review system uses 4 agents in sequence. Each agent has a defined input, a defined output, and a defined scope.

Agent 1: Intake.

The intake agent creates a single entry point for all contracts regardless of source. It monitors email inboxes, shared drives, and a web upload form. When a new document arrives, the agent:

Identifies the document type (vendor agreement, service contract, NDA, employment agreement, partnership terms, lease, other).
Detects the language (Spanish, Portuguese, English, or mixed).
Extracts metadata: parties, date, governing law jurisdiction, contract value if stated.
Assigns a priority based on contract type and value (employment agreements and high-value vendor contracts get priority over standard NDAs).
Creates a tracking record with a unique ID, timestamps, and status.

The intake agent handles PDF, DOCX, and scanned documents. Scanned documents go through OCR before processing. The agent flags documents where OCR confidence is below 85% for manual verification.

Processing time per document: 1-3 minutes. The agent processes documents as they arrive, so there’s no queue buildup. At the end of each business day, the legal team has a dashboard showing every contract that entered the pipeline, its type, language, priority, and current status.

Agent 2: Extraction.

The extraction agent reads the full contract and produces a structured breakdown. For each contract, it extracts:

Parties and roles. Who is obligated to do what. Entity names, jurisdiction of incorporation where stated, signing authority.
Key commercial terms. Contract value, payment terms, payment schedule, currency, price adjustment mechanisms.
Duration and termination. Start date, end date, renewal terms (auto-renewal, opt-out, opt-in), termination for cause conditions, termination for convenience terms, notice periods.
Liability and indemnification. Liability caps (per-incident, aggregate, uncapped), indemnification obligations (mutual, one-sided), exclusions from liability limitations, insurance requirements.
IP and confidentiality. IP ownership (who owns work product, pre-existing IP protections, license grants), confidentiality obligations (duration, scope, exceptions), non-compete terms.
Data protection. Data processing obligations, cross-border transfer provisions, breach notification terms, data subject rights handling, sub-processor authorization.
Governing law and disputes. Governing law jurisdiction, dispute resolution mechanism (arbitration, litigation, mediation first), venue, applicable arbitration rules.
Compliance and regulatory. Anti-corruption representations, sanctions compliance, regulatory approval conditions.

The extraction handles multi-language contracts by processing in the source language. A Brazilian vendor agreement in Portuguese gets extracted in Portuguese. The structured output includes both the original language text and an English translation for each extracted element.

For mixed-language contracts (common in LATAM when a multinational’s template is in English but local schedules and annexes are in Spanish or Portuguese), the agent processes each section in its source language and flags the language boundaries.

Extraction time per document: 5-15 minutes depending on length and complexity. A 10-page NDA takes about 5 minutes. A 45-page partnership agreement with annexes takes 12-15 minutes.

Agent 3: Risk Scoring.

The risk scoring agent compares extracted terms against the company’s approved templates and risk policies. This is where the system catches what human associates miss during high-volume periods.

For each extracted element, the agent performs three checks:

Deviation check. Is this term different from our template? The agent compares each extracted clause against the corresponding template clause. Deviations are classified as: cosmetic (different wording, same legal effect), minor (different but within acceptable range per company policy), major (outside acceptable range, requires negotiation), or critical (terms the company has defined as non-negotiable).

The deviation check handles structural variations. If a contract combines limitation of liability and indemnification into a single clause where the template has them separate, the agent still identifies and compares the relevant provisions. This is the comparison that humans most frequently get wrong during manual review because they’re looking for clause-to-clause correspondence rather than concept-to-concept matching.

Risk factor identification. The agent checks for specific risk factors independent of template comparison: unlimited liability exposure, unilateral amendment rights, automatic renewal without notice, IP assignment broader than the scope of work, governing law in an unfavorable jurisdiction, arbitration under unfamiliar rules, missing data protection provisions for contracts involving personal data, non-compete terms that may be unenforceable in certain LATAM jurisdictions.

Each risk factor has a severity score (1-5) and a category (financial, operational, legal, regulatory). The scores are configured per company based on their risk tolerance and operational context.

Jurisdiction check. For LATAM operations, the agent checks whether the contract’s terms are enforceable in the governing law jurisdiction. A non-compete clause valid under Mexican law might be unenforceable under Brazilian law. A liquidated damages provision acceptable in Chile might face limitations in Argentina. The agent flags jurisdiction-specific enforceability concerns based on a rule set maintained per country.

The risk scoring output is a structured report: overall risk score (1-10), list of deviations by severity, list of risk factors by severity, jurisdiction concerns, and a recommended action (approve as-is, approve with noted risks, negotiate specific terms, reject and propose counter).

Scoring time: 3-8 minutes per contract.

Agent 4: Summary.

The summary agent takes all outputs from the previous three agents and produces two documents.

First, a 1-page executive summary for the senior attorney. This contains: contract type, parties, value, duration, overall risk score, top 3 issues requiring decision, and recommended action. A senior attorney can read this in 2 minutes and decide whether the contract needs detailed review or can be processed on the agent’s recommendation.

Second, a detailed review memo that replaces the junior associate’s traditional memo. This contains every extracted term, every deviation, every risk factor, and the relevant source text from the contract (quoted in original language with translation). This document is what the senior attorney uses if they decide the contract needs detailed attention.

Summary generation: 2-4 minutes per contract.

Throughput and Accuracy

Total pipeline processing time per contract: 12-30 minutes depending on complexity. For a team processing 200 contracts per month, the agent system handles all 200 through extraction, comparison, and scoring in the time it previously took to manually review 30-40.

Of those 200, typically 40-50% are standard contracts with low risk scores that can be processed with minimal human review (senior attorney reads the 1-page summary, confirms the recommendation, approves). Another 30-35% have moderate risk scores requiring the senior attorney to review specific flagged sections. The remaining 15-25% have high risk scores requiring full attorney review and negotiation.

The result: human attorney time per month drops from 300-800 hours to 60-120 hours. That’s not 60-120 hours of reading contracts. That’s 60-120 hours of making decisions, negotiating terms, and handling the complex contracts that justify having experienced lawyers on staff.

Accuracy measurement: we compared agent output against senior attorney review for 150 contracts over a 3-month validation period. The agent system identified 94% of deviations that the senior attorney flagged as significant. Of the 6% it missed, most were contextual issues where the risk depended on business circumstances the agent didn’t have (for example, a payment term that was standard in isolation but problematic given the company’s cash flow situation in a specific quarter). The agent produced false positives at a rate of about 8%, flagging issues that the senior attorney deemed acceptable. False positives are a nuisance, not a risk. Missed deviations are a risk, and a 6% miss rate compares favorably to the 12% miss rate we measured in manual review.

The Multi-Language Factor

LATAM operations produce contracts in at least two languages. Many companies deal with three (Spanish, Portuguese, English). Some contracts are bilingual, with both languages in the same document and a clause specifying which version controls in case of discrepancy.

The agent system processes each language natively. Portuguese contracts are extracted in Portuguese. Spanish contracts in Spanish. English contracts in English. Cross-language comparison (a Brazilian subsidiary’s contract compared against the parent company’s English-language template) works at the concept level rather than the word level. The extraction agent normalizes legal concepts into a common structure regardless of source language, so “clausula de limitacion de responsabilidad,” “clausula de limitacao de responsabilidade,” and “limitation of liability clause” all map to the same structural element.

This is where agents provide the clearest improvement over manual review. A human associate comparing a Portuguese contract against a Spanish template is doing real-time translation and comparison simultaneously. Mental load is high. Errors are common. The agent doesn’t have this problem because it processes structure, not surface text.

What Still Needs a Lawyer

The agent system doesn’t draft contracts, negotiate terms, advise on legal strategy, or interpret ambiguous provisions. It doesn’t decide whether a risk is acceptable given the business context. It doesn’t manage outside counsel relationships. It doesn’t appear before regulators.

When the risk scoring agent flags a limitation of liability clause as “critical deviation,” a lawyer decides whether to accept it, negotiate it, or walk away from the deal. When the jurisdiction check identifies an enforceability concern with a non-compete clause in Brazil, a lawyer decides whether to remove the clause, restructure it, or accept the risk.

The agents handle the reading, comparison, and scoring that consumed 70% of the legal team’s time. The lawyers handle the judgment, negotiation, and decision-making that they were trained for but rarely had time to do properly.

Synaptic builds AI agent systems for legal teams across Latin America. Process the volume. Focus on the judgment calls. synaptic.so