legal-redliner

name

legal-redliner

description

Jurisdiction-aware legal document redlining agent: ingests customer-facing legal documents (NDAs, MSAs, DPAs, T&Cs, SOWs), segments them into clauses, rewrites each clause to align with internal policy positions and regional legal requirements, produces a fully redlined document plus a structured audit pack with change-by-change rationale and citations. Supports UK, EU, US, and custom jurisdictions. Use when user says "redline this contract", "review against our playbook", "localize this agreement", "adapt for EU", "adapt for UK", "jurisdiction review", "policy-aligned redline", "clause review", "contract redline", "playbook review", "mark up this agreement", "regional adaptation", "legal playbook", "NDA review", "MSA review", "DPA review", "T&Cs review", "standard positions review", "contract markup with audit trail", or needs jurisdiction-specific, policy-aligned contract redlining with full auditability.

Legal Redliner

Redline customer-facing legal documents clause-by-clause — aligning each clause to your internal policy positions and jurisdiction-specific legal requirements — with a complete audit trail showing what changed, why, and what evidence supports each change.

When to Use

• User has a legal document (NDA, MSA, DPA, SOW, T&Cs) that needs redlining against company playbook positions
• User needs to adapt a contract for a specific jurisdiction (UK, EU, US states, etc.)
• User wants a structured audit pack showing every change with rationale and citations
• User needs to compare contract language against internal standard positions
• User wants clause-by-clause analysis with risk flags for non-standard terms

When NOT to Use

• Litigation or dispute management → use legal-review
• Vendor evaluation or spend analysis → use vendor-intel
• General contract term extraction without redlining → use vendor-intel
• Sending the redlined document to someone → use email after redlining

Quick Start

User: "Redline this MSA against our standard positions for a UK customer."

Steps:

1. Ingest and segment the document into clauses
2. Identify the document type, jurisdiction, and applicable policies
3. For each clause: retrieve relevant policy positions, check jurisdiction requirements, rewrite if needed
4. Assemble the redlined document with tracked changes
5. Generate the audit pack (change log, citations, risk flags, human review checklist)

---

How It Works

Architecture Overview

This skill operates as a clause-by-clause redlining pipeline — rather than loading all policies and legislation at once (which would overwhelm context), it performs targeted retrieval per clause.

INGEST → CLASSIFY → SEGMENT → [PER-CLAUSE LOOP] → ASSEMBLE → DELIVER
                                    |
                                    ├── Retrieve internal policy snippets
                                    ├── Retrieve jurisdiction-specific requirements
                                    ├── Rewrite with minimal changes
                                    ├── Validate against constraints
                                    └── Record evidence + rationale

Why Clause-by-Clause

• Precision: Each clause gets only the relevant policy and legal context
• Auditability: Every change maps to specific evidence
• Control: Deterministic loop ensures nothing is skipped
• Transparency: The audit pack traces each decision back to sources

---

Redlining Workflow

Step 1: Document Ingestion

1. Locate the document — check input/ for uploaded files first

2. Extract text — use the pdf skill for PDFs or docx skill for Word files to extract full text

3. Identify document metadata:

   Field	Value
   Document Type	NDA / MSA / DPA / SOW / T&Cs / Other
   Counterparty	[Name]
   Target Jurisdiction	UK / EU / Germany / US-CA / US-NY / etc.
   Governing Law	[As stated in doc]
   Date	[Document date]
   Our Role	Customer / Supplier / Both

Step 2: Classification & Context Setup

Based on the document type and jurisdiction, establish:

Policy positions to retrieve (examples by document type):

Document Type	Key Policy Areas
NDA	Definition of confidential information, exclusions, term, return/destruction, permitted disclosures, injunctive relief
MSA	Liability caps, indemnification, IP ownership, warranty, termination, change control, dispute resolution
DPA	Processing purposes, sub-processors, international transfers, breach notification, data subject rights, DPIA obligations
SOW	Scope definition, acceptance criteria, change orders, milestones, payment terms, resource commitments
T&Cs	Liability limitations, consumer rights, refund/cancellation, data collection notices, cookie consent

Jurisdiction-specific requirements (examples):

Jurisdiction	Key Considerations
UK	UK GDPR, Consumer Rights Act 2015, Unfair Contract Terms Act 1977, PECR
EU	EU GDPR, EU consumer directives, local member state implementations
Germany	BGB (German Civil Code) provisions, stricter GDPR interpretation, specific AGB rules
US — California	CCPA/CPRA, California-specific consumer protections
US — New York	NY General Obligations Law, specific warranty rules
US — Delaware	Choice of law preferences, corporate governance considerations

Step 3: Document Segmentation

Use the segmentation script to split the contract into individually addressable clauses:

python scripts/segment_contract.py <contract_text_file> --output working/segmented.json

The script automatically:

• Detects clause headings (numbered sections, articles, schedules)
• Extracts preamble/recitals as CL-000
• Classifies each clause by content type (definitions, obligations, liability, termination, data processing, IP)
• Extracts defined terms and cross-references
• Returns a confidence score for segmentation quality

Review the JSON output and verify the segmentation. If confidence is "low", manual segmentation may be needed.

The script produces a structured table:

Clause ID	Section	Clause Heading	Clause Text	Classification
CL-001	1	Definitions	[full text]	Standard / Non-standard / Missing
CL-002	2.1	Scope of Services	[full text]	Standard / Non-standard / Missing
...

Classification key:

• Standard: Matches or closely aligns with our standard positions
• Non-standard: Deviates from our standard positions — needs redlining
• Missing: A clause our playbook requires that's not in the document
• Acceptable: Non-standard but within acceptable deviation range

Step 4: Per-Clause Redlining

For each clause marked as Non-standard or Missing, apply this process:

4a. Retrieve Relevant Context

• Internal policy: Find the company's standard position for this clause type
• Jurisdiction requirements: Find any legal requirements that affect this clause in the target region

Sources (in priority order):

1. User-provided policy documents (uploaded to input/)
2. Organization's SharePoint playbook (if user directs to a location — use enterprise search to find it)
3. Official legislation — see Legislation Sourcing Rule below
4. General best practice for the clause type and jurisdiction

Legislation Sourcing Rule (MANDATORY): All jurisdiction-specific legal citations MUST be retrieved from official government or legislative sources at redlining time — never from pre-training knowledge alone. Use the deep-research agent or web search/fetch tools to retrieve the current text of any statute, regulation, or directive being cited.

• If the user has provided explicit legislation documents (uploaded to input/ or linked in their request), use those as the authoritative source. No web lookup is needed for provisions covered by the user's documents.
• For all other legislation, fetch from official sources before citing. See references/jurisdiction-sources.md for the full jurisdiction source table with official URLs. Key sources:
  • Germany: gesetze-im-internet.de (BGB, UrhG, BDSG, HGB, etc.)
  • EU: eur-lex.europa.eu (GDPR, directives, SCCs decisions)
  • UK: legislation.gov.uk (UK GDPR, DPA 2018, UCTA 1977, CRA 2015)
  • US: govinfo.gov, state legislature sites (CCPA, UCC, state statutes)
• If an official source cannot be reached or the provision cannot be verified, flag the clause as NEEDS_RESEARCH with the note "Legislation not verified from official source" rather than citing from memory.
• Record the source URL and retrieval date in the audit pack's Legal Ref column for every legislative citation.

4b. Rewrite the Clause

Non-negotiable rewriting rules:

1. Minimal changes — prefer targeted edits over full rewrites
2. Preserve defined terms exactly (capitalisation, wording)
3. Preserve structure — don't renumber or reorganise unless required
4. External law prevails — if internal policy conflicts with mandatory law, law wins (note the conflict)
5. Conservative approach — when in doubt, flag for human review rather than rewrite
6. No invented law — all legal citations must be verified against official government sources at redlining time (see Legislation Sourcing Rule in Step 4a). If the provision cannot be verified, flag NEEDS_RESEARCH rather than citing from memory

4c. Record the Change

For every change, capture:

Field	Content
Clause ID	CL-XXX
Change Type	Insert / Delete / Replace / Add Missing Clause
Original Text	[exact original wording]
Revised Text	[proposed new wording]
Rationale	[why this change was made — in plain language]
Policy Citation	[which internal policy position this aligns to]
Legal Citation	[which law or regulation requires or supports this]
Risk Level	Low / Medium / High
Needs Human Review?	Yes / No
Open Questions	[anything unresolved or uncertain]

4d. Validate

Before finalizing each clause:

• Does the revised text align with the cited policy position?
• Does it comply with the cited jurisdiction requirement?
• Are defined terms preserved?
• Is the commercial position unchanged (unless compliance requires it)?
• Is the change truly necessary, or is the original acceptable?

---

Output Deliverables

The legal redliner produces two documents:

1. Redlined Document

A Word document with tracked changes showing:

• Deleted text (strikethrough, red)
• Inserted text (underline, red)
• Comments on each change referencing the audit pack entry

Use the docx skill to create the redlined document with proper tracked changes formatting.

Include the privilege notice from assets/privilege-notice.md at the top of the document and as the header on every page.

When inserting missing clauses (e.g., force majeure, third party rights, severability), use the pre-approved standard wording from assets/standard-clauses-uk.md instead of generating from scratch. This ensures consistency with the legal team's approved language.

2. Audit Pack

A comprehensive audit record. Generate as an Excel workbook using the xlsx skill with these sheets:

Sheet 1: Change Log

Clause ID	Section	Change Type	Original Text	Revised Text	Rationale	Policy Ref	Legal Ref	Risk	Human Review?

Sheet 2: Risk Summary

Risk Level	Count	Clauses
High		[list]
Medium		[list]
Low		[list]

Sheet 3: Human Review Checklist

#	Clause ID	Item for Review	Reason	Reviewer	Status
			Insufficient policy data / Judgment call / Complex clause		Pending / Reviewed / Approved

Sheet 4: Coverage Summary

Total Clauses	Standard (no change)	Redlined	Missing (added)	Needs Research	% Coverage

All counts and percentages must be computed using code tools.

---

3. Risk Dashboard (Adaptive Card)

Before generating the file deliverables, render an inline risk dashboard using the render-ui skill so the user can review the summary at a glance. The card must include:

KPI Row (ColumnSet with 4 columns):

• Total Clauses (count)
• Redlined (count + percentage)
• Standard (count, green)
• Coverage (percentage, green if 100%)

Risk Distribution (Chart.Donut):

• High Risk (attention colour) with count
• Medium Risk (warning colour) with count
• Low Risk (good colour) with count
• Standard (neutral colour) with count

Clause Summary Table (Table):

• Columns: Clause ID | Clause Name | Key Issue | Risk (as Badge) | Human Review? (as Badge where applicable)
• Only show non-standard clauses (omit standard clauses from the table)
• Sort by risk level: High first, then Medium, then Low

Legislation Verified (FactSet):

• List each statute cited with its source (e.g., "UCTA 1977 s.2 → legislation.gov.uk")

Render the card before generating the Word and Excel files — this gives the user a chance to review the analysis before committing to file generation.

---

Supported Document Types

Type	Abbreviation	Typical Clauses Reviewed
Non-Disclosure Agreement	NDA	Confidentiality definition, exclusions, term, return, remedies
Master Service Agreement	MSA	Scope, liability, indemnity, IP, warranty, termination, governing law
Data Processing Agreement	DPA	Processing purposes, sub-processors, transfers, breach notification, DSAR
Statement of Work	SOW	Scope, deliverables, acceptance, milestones, change control, payment
Terms and Conditions	T&Cs	Liability, consumer rights, data, refunds, dispute resolution
Software License Agreement	SLA	License scope, restrictions, support, updates, audit rights
Employment Agreement	EA	Duties, compensation, restrictive covenants, IP assignment, termination

---

Built-In Skills Used

This skill leverages Copilot Cowork's built-in capabilities:

Built-in Skill	How It's Used
Word	Read input documents, create redlined output with tracked changes
PDF	Extract text from PDF contracts, create final PDF versions
Excel	Generate the audit pack workbook with change log, risk summary, and coverage
Enterprise Search	Find internal playbook documents and policy positions across SharePoint
Deep Research	Research jurisdiction-specific legal requirements when policy sources are insufficient
Email	Send the completed redline and audit pack to reviewers for approval
Communications	Draft cover notes using assets/cover-note-template.md explaining the redline for stakeholders
Adaptive Cards	Display the risk summary and coverage dashboard inline before generating files
Scheduling	Book review meetings with legal team to walk through the redline
Calendar Management	Check reviewer availability before scheduling review sessions
Meetings	Prepare briefing notes for the redline review meeting

---

Workflow Integration

End-to-End Process

1. User uploads contract → Legal Redliner segments and classifies
2. Redlining → Clause-by-clause analysis with audit trail
3. Review dashboard → Display risk summary and coverage as an Adaptive Card for quick review
4. Generate deliverables → Redlined Word document + Excel audit pack
5. Circulate for review → Email the package to the legal team with a cover note via Communications
6. Schedule review → Use Scheduling to find time for the legal team to review together
7. Post-review → Update the redline based on feedback, generate final versions

Multi-Jurisdiction Handling

When a document needs adaptation for multiple regions:

• Run the redliner once per jurisdiction
• Generate a comparison view showing regional differences
• Highlight clauses where jurisdictions require conflicting approaches
• Recommend the most conservative position that satisfies all jurisdictions, or separate regional versions

---

Quality Assurance

Pre-Delivery Checks

Before delivering the redline:

• Every change has a rationale and at least one citation
• All defined terms preserved exactly
• No commercial position changed without explicit flagging
• Risk levels assigned consistently
• Human review items clearly identified
• Coverage > 90% (all clauses addressed or explicitly marked)
• Audit pack totals match the redlined document
• Document formatted correctly with tracked changes visible

Confidence Indicators

For each clause assessment:

• High confidence: Both policy source and legal source available, clear alignment
• Medium confidence: Policy source available but no jurisdiction-specific confirmation, or vice versa
• Low confidence / Needs Research: Insufficient sources — flagged for human review

---

Guardrails

• This skill does NOT provide legal advice: All redlines are drafting suggestions based on policy alignment and identified legal requirements. A qualified lawyer must review before any document is finalized.
• Never invent law or policy: All legislative citations must be retrieved and verified from official government sources (e.g., gesetze-im-internet.de, eur-lex.europa.eu, legislation.gov.uk) at redlining time — not from pre-training knowledge. The only exception is when the user has explicitly provided the legislation as uploaded documents. If a provision cannot be verified from an official source, flag NEEDS_RESEARCH — never fabricate or cite from memory
• Minimal intervention principle: Prefer targeted edits over rewrites. The goal is compliance and policy alignment, not stylistic preference.
• Commercial neutrality: Do not change the commercial position (pricing, scope, payment terms) unless specifically required for legal compliance — flag any commercial impact prominently
• Confidentiality: Contract documents are commercially sensitive. Never reference details from one counterparty's contract in another review.
• Privilege: Mark all redlining output as "PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT" unless user instructs otherwise
• All calculations computed by code tools: Coverage percentages, clause counts, risk distributions — never estimate

Common Issues

Issue	Cause	Fix
Contract is a scanned PDF with no text	Image-only PDF	Use pdf skill with OCR extraction first
No internal playbook provided	User hasn't uploaded policy documents	Ask user to provide or point to SharePoint location; use enterprise search to find
Jurisdiction not supported	Niche jurisdiction with no available sources	Flag all jurisdiction-dependent clauses as NEEDS_RESEARCH, recommend local counsel review
Too many "Needs Research" flags	Insufficient source material	Engage the deep-research skill to investigate specific legal requirements, then re-run affected clauses
Redline changes commercial terms	Clause mixes legal and commercial language	Separate the legal compliance change from the commercial term, flag the commercial element for business review
Audit pack doesn't match redline	Manual edits after generation	Always regenerate audit pack from the final redlined document