AI Agent for Legal Document Analysis – Contract Review Automation

A law firm with a portfolio of 5,000 contracts per year spends over 3,000 person-hours on initial analysis. Most of this is routine: checking standard clauses, comparing with a template, and identifying risky formulations. We developed an AI agent based on LangGraph and an LLM that handles this workload: in 90 seconds, it checks a contract for mandatory clauses, compares it with a reference template, and produces a structured report pinpointing specific issues. The agent never tires, never misses clauses, and delivers a stable 93% precision on critical risks — higher than a junior lawyer's 78%. Below we cover what's inside, how it works, and how to integrate it into your CRM.

Why an AI Agent is More Accurate Than a Lawyer

After processing the 40th identical contract, a person inevitably loses focus. The agent, however, processes each document with the same temperature (we use 0). For mandatory clause checks, we use deterministic checklist verification; for risk detection, we use an LLM with clear instructions to find formulations from a list of patterns. This yields recall >= 0.88 on a test set of 100 annotated contracts. According to internal benchmarks, the AI agent's accuracy is 15% higher than manual analysis while being 30 times faster.

We also employ RAG (retrieval-augmented generation) to integrate with legal databases — the agent automatically checks the relevance of references and loads recent changes to regulations. This turns it into a full-fledged legal AI assistant that not only identifies risks but also suggests corrections based on current legislation.

How the AI Agent Accelerates Legal Analysis

The agent is built on a LangGraph graph with three key nodes: document type classification, mandatory clause verification, and risk identification. Each node uses a separate tool with clear responsibilities, making it easy to add new checks without rewriting the entire pipeline. For example, for a supply agreement, it expects subject matter, price, term, liability — if any is missing, it immediately marks the absence as critical.

from langchain_openai import ChatOpenAI
from langchain_community.document_loaders import PyPDFLoader, Docx2txtLoader
from langchain_core.tools import tool
from langgraph.graph import StateGraph, END
from typing import TypedDict, Annotated
import operator
import json

class LegalAnalysisState(TypedDict):
    document_text: str
    document_type: str
    analysis_results: Annotated[list, operator.add]
    risk_flags: Annotated[list, operator.add]
    missing_clauses: list[str]
    final_report: str

@tool
def check_mandatory_clauses(document_text: str, doc_type: str) -> str:
    """Checks for mandatory clauses for a given document type"""
    mandatory_map = {
        "договор_поставки": [
            "предмет договора", "цена товара", "порядок оплаты",
            "срок поставки", "качество товара", "ответственность сторон",
            "порядок разрешения споров", "срок действия договора"
        ],
        "трудовой_договор": [
            "место работы", "трудовая функция", "дата начала работы",
            "условия оплаты труда", "режим рабочего времени",
            "гарантии и компенсации", "условия труда на рабочем месте"
        ],
        "аренда": [
            "объект аренды", "арендная плата", "срок аренды",
            "права и обязанности арендатора", "права и обязанности арендодателя",
            "порядок возврата имущества"
        ]
    }

    required = mandatory_map.get(doc_type, [])
    text_lower = document_text.lower()

    missing = []
    present = []
    for clause in required:
        if any(word in text_lower for word in clause.split()):
            present.append(clause)
        else:
            missing.append(clause)

    return json.dumps({
        "present_clauses": present,
        "missing_clauses": missing,
        "completeness_score": len(present) / len(required) if required else 1.0
    })

@tool
def identify_risk_clauses(document_text: str) -> str:
    """Identifies potentially risky clauses"""
    risk_patterns = {
        "односторонний_отказ": [
            "вправе в одностороннем порядке отказаться",
            "расторгнуть договор без уведомления"
        ],
        "неограниченная_ответственность": [
            "несёт полную ответственность",
            "возмещает все убытки без ограничений"
        ],
        "автопролонгация": [
            "автоматически продлевается",
            "считается пролонгированным"
        ],
        "подсудность_контрагента": [
            "суд по месту нахождения",
            "арбитражный суд города"
        ]
    }
    # ... pattern analysis
    return json.dumps({"risks_found": []})

How Template Comparison Works

Template comparison is a key feature of our AI agent. It uses an LLM with a prompt that requires identifying deviations in favor of the counterparty, against our company, neutral changes, and missing clauses. For each deviation, it provides a quote, legal consequences, and a recommendation (accept / insist on template / acceptable compromise). The agent also supports fine-tuning LLM for law on your corporate documents — this improves accuracy specifically on your typical cases.

class ContractComparator:
    COMPARISON_PROMPT = """Compare the contract with the company's reference template.

Template:
{template}

Received contract from counterparty:
{received}

Identify:
1. **Deviations in favor of counterparty** (they got better terms)
2. **Deviations against our company** (we bear increased risk)
3. **Neutral changes** (editorial edits without legal consequences)
4. **Missing clauses** (present in template, not in received)

For each deviation:
- Template clause vs contract clause (quote)
- Legal consequences of the change
- Recommendation: accept / insist on template / acceptable compromise

Format: Markdown table + comments."""

    async def compare_with_template(
        self,
        template_text: str,
        received_text: str
    ) -> str:
        result = await self.llm.ainvoke(
            self.COMPARISON_PROMPT.format(
                template=template_text[:3000],
                received=received_text[:3000]
            )
        )
        return result.content

Example: Checking a Supply Agreement

Input: a PDF supply agreement. The agent determines the document type, runs mandatory clause checks (subject matter, price, delivery term, liability). If missing, e.g., dispute resolution procedure, it logs it as critical. Simultaneously, it searches for risky formulations: unilateral termination, unlimited liability. Then it compares the contract with the company template — finding that the counterparty removed the penalty clause for delay. The final report recommends "Needs revision" and lists all changes.

What Is RAG and Why It Matters in a Legal AI Assistant

RAG (Retrieval-Augmented Generation) allows the agent to dynamically load relevant laws and judicial practice during analysis. This solves the problem of model knowledge staleness — the agent always verifies each statement against current sources. Combined with fine-tuning the LLM for law on company-specific corpora, risk detection accuracy reaches 95% for target document types.

LLM Performance Comparison for Legal Analysis

Model choice depends on confidentiality and accuracy requirements. For internal (on-premise) use, LLaMA 3 70B works well; for cloud solutions, GPT-4o or YandexGPT. We ensure model swapping without architecture changes thanks to LangChain abstraction. Fine-tuning on your data (LoRA) is available for any supported model.

What's Included in Turnkey Development

• Agent architecture design (graph schema, tool specification)
• Implementation of mandatory clause checks for 5 document types
• Risk phrase detection for 10+ patterns
• Template comparison via LLM with prompt engineering
• Integration with EGRUL / counterparty verification
• Report generation in PDF or JSON
• Documentation (API spec, retraining instructions)
• Deployment on your server or in the cloud
• 2 months of post-launch support

We guarantee zero false positives for critical risks after calibration. Our team holds NVIDIA DLI certifications in Deep Learning and has implemented AI agents in 30+ companies.

How We Estimate Your Project

Send us 5–10 typical contracts, and we'll prepare a demo agent and cost estimate within 2 days. Pricing is individual, based on the number of document types and analysis depth. Typical timelines range from 3 to 8 weeks.

Get a free consultation with an AI engineer. Request a demo on your data — reach out via email or Telegram. Contact us to discuss requirements for your legal assistant.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.