Play 53 — Legal Document AI

AI-powered legal document analysis — contract review, clause extraction with layout-aware parsing, risk scoring against industry benchmarks, redlining suggestions, version comparison, jurisdiction-aware analysis, UPL-safe disclaimers on every output, and attorney-client privilege markers.

Architecture

Full architecture details: architecture.md

Aspect	Play 06 (Document Intelligence)	Play 53 (Legal Document AI)	Play 38 (Doc Understanding V2)
Domain	General document processing	Legal contracts specifically	Multi-page entity linking
Output	Extracted fields	Clause risk scores + redline suggestions	Structured extraction
Compliance	PII redaction	UPL compliance + privilege markers	Confidence thresholds
Benchmark	Schema-based	Industry-standard clause benchmarks	Entity linking accuracy
Legal Safety	N/A	”Not legal advice” on every output	N/A
Jurisdiction	N/A	State/country-aware analysis	N/A

DevKit Structure


53-legal-document-ai/
├── agent.md                              # Root orchestrator with handoffs
├── .github/
│   ├── copilot-instructions.md           # Domain knowledge (<150 lines)
│   ├── agents/
│   │   ├── builder.agent.md              # Clause extraction + risk + redline
│   │   ├── reviewer.agent.md             # UPL + privilege + PII
│   │   └── tuner.agent.md                # Clause library + benchmarks + cost
│   ├── prompts/
│   │   ├── deploy.prompt.md              # Deploy legal pipeline
│   │   ├── test.prompt.md                # Review sample contracts
│   │   ├── review.prompt.md              # Audit UPL compliance
│   │   └── evaluate.prompt.md            # Measure extraction accuracy
│   ├── skills/
│   │   ├── deploy-legal-document-ai/     # Doc Intel + clause + risk + redline
│   │   ├── evaluate-legal-document-ai/   # Clauses, risk, UPL, redline quality
│   │   └── tune-legal-document-ai/       # Clause library, benchmarks, UPL
│   └── instructions/
│       └── legal-document-ai-patterns.instructions.md
├── config/                               # TuneKit
│   ├── openai.json                       # Legal model (temp=0), redline config
│   ├── guardrails.json                   # Risk benchmarks, UPL rules
│   └── agents.json                       # Clause library, jurisdiction rules
├── infra/                                # Bicep IaC
│   ├── main.bicep
│   └── parameters.json
└── spec/                                 # SpecKit
    └── fai-manifest.json

Quick Start


# 1. Deploy legal AI pipeline
/deploy
 
# 2. Review sample contracts
/test
 
# 3. Audit UPL compliance
/review
 
# 4. Measure extraction accuracy
/evaluate

Key Metrics

Metric	Target	Description
Clause Detection	> 90%	Expected clauses found
Risk Calibration	> 80%	Scores match attorney assessment
UPL Compliance	100%	Disclaimers + privilege markers
Redline Relevance	> 85%	Suggestions address identified risk
Critical Risk Detection	> 95%	High-severity risks caught
Cost per Contract	< $3.00	20-page MSA review

Cost Estimate

Service	Dev	Prod	Enterprise
Azure OpenAI	$100	$900	$3,500
Azure AI Search	$0	$250	$1,000
Azure Document Intelligence	$0	$100	$400
Azure Blob Storage	$5	$40	$150
Cosmos DB	$5	$75	$350
Key Vault	$1	$5	$15
Application Insights	$0	$30	$100
Container Apps	$10	$80	$350
Total	$121	$1,480	$5,865

Detailed breakdown with SKUs and optimization tips: cost.json · Azure Pricing Calculator

WAF Alignment

Pillar	Implementation
Responsible AI	UPL disclaimers, privilege markers, “not legal advice” on every output
Security	PII de-identification, Key Vault for secrets, no PII in LLM context
Reliability	Deterministic scoring (temp=0), clause-by-clause processing
Cost Optimization	gpt-4o-mini for classification, batch clause extraction, redline threshold
Operational Excellence	Clause library per contract type, jurisdiction rules, version comparison
Performance Efficiency	Layout extraction for structure, 5-section batching, cached benchmarks

FAI Manifest

Field	Value
Play	`53-legal-document-ai`
Version	`1.0.0`
Knowledge	R2-RAG-Architecture, T2-Responsible-AI, O2-Agent-Coding, R3-Deterministic-AI
WAF Pillars	security, reliability, responsible-ai, cost-optimization
Groundedness	≥ 85%
Safety	0 violations max