How I Built an AI Agent System for a Corporate Law Firm — During My Internship

M0rtal13 — Tue, 04 Aug 2026 09:22:03 +0000

How I Built an AI Agent System for a Corporate Law Firm — During My Internship

TL;DR: I built an open-source template that lets corporate law firms deploy AI agents using Obsidian + CLI tools (Claude Code, Qwen Code, Cursor). It handles contract drafting, legal research, due diligence, compliance checks, client intake, and wiki maintenance — all organized with PARA methodology and built-in ethical guardrails.

GitHub repo: M0rtal13/Agentic-AI-for-Enterprise

The Backstory

Two weeks ago, I started a corporate law firm internship. On day one, my supervisors gave the interns a challenge:

"Think critically about how AI tools might support a modern corporate law practice... On the last day, share 3–5 minutes of reflections and offer one or two concrete ideas on how AI could help improve workflow."

Most interns planned to make PowerPoint slides with bullet points.

I decided to build the actual system instead of just talking about it.

Over the next two weeks, while working on real legal research and document review tasks, I built an open-source template that any law firm can deploy. Not a demo — a working system with skills, guardrails, and a phased rollout plan.

This is what I built and how it works.

The Problem: Law Firms Waste Time on Repetitive Work

During my internship, I noticed the same patterns:

Contract drafting — Associates spend hours customizing NDAs, SPAs, and service agreements from precedent templates
Legal research — Searching through databases, summarizing cases, cross-referencing statutes
Due diligence — Reviewing hundreds of documents in data rooms, flagging risks manually
Compliance checklists — Generating matter-specific checklists for different entity types and jurisdictions
Knowledge management — Precedents scattered across folders, wiki pages outdated, lessons learned never captured

These tasks are high-volume, pattern-based, and time-consuming — perfect for AI assistance.

But law firms have unique constraints:

Confidentiality — Client data can't leave the firm's secure environment
Accuracy — Hallucinations are unacceptable in legal work
Audit trails — Every action must be logged
Human oversight — AI drafts, humans decide

The Solution: PARA + AI Agents + Obsidian

The system combines three things:

1. PARA Knowledge Management

All firm knowledge is organized into five folders:

00_inbox/          ← Unprocessed captures
01_projects/       ← Active matters (matter-2026-001-acquisition/)
02_areas/          ← Ongoing responsibilities (corporate-advisory/, m-a/, compliance/)
03_resources/      ← Precedents, clause library, legislation, case law
04_archive/        ← Completed matters
wiki/              ← AI-powered knowledge base

Every matter, document, and research memo has a defined location and naming convention. The AI agent operates on this structure.

2. CLI AI Agents

The system works with any CLI-based AI agent:

Claude Code (Anthropic)
Qwen Code (Alibaba)
Cursor
Codex (OpenAI)
Windsurf
Gemini CLI (Google)

These agents read a CLAUDE.md / AGENTS.md file that defines:

Vault structure and file organization
Skill routing (user intent → specific skill)
Writing standards (British English, OSCOLA citations)
Core principles (draft from precedents, cite everything, human-in-the-loop)

3. Six Legal Skills

The repo includes six fully-functional skills:

Skill	What it does
Contract Drafting	Identifies document type, locates precedent, gathers transaction details, drafts from template, quality checks
Legal Research	Frames question, identifies sources (priority order), researches, structures memo (Issue/Short Answer/Analysis/Authorities/Limitations), verifies citations
Due Diligence	Understands scope, triages documents, extracts parties/obligations/restrictions/termination/liability, classifies risk (High/Medium/Low), populates report
Compliance Check	Profiles entity, identifies obligations across categories (corporate/financial/listing/sector-specific/employment), generates table, flags urgent items
Client Intake	Collects client info, runs conflict check, creates matter folder, prepares engagement letter
Wiki Maintenance	Identifies stale pages (90-day review cycle), creates/updates pages using standard template, ensures tag consistency, logs changes

Each skill is a markdown file with step-by-step procedures, rules, and quality checks. The agent follows these procedures when you invoke the skill.

How It Works in Practice

Example: Drafting an NDA

You: "Draft a bilateral NDA for a tech company acquiring a startup in Hong Kong"

Agent:

Identifies document type → Bilateral NDA
Locates precedent → 03_resources/precedents/standard-nda-bilateral.md
Gathers transaction details (parties, jurisdiction, duration, purpose)
Drafts from template, customizing for HK jurisdiction and M&A context
Runs quality check (defined terms consistency, cross-border considerations, listed company provisions)
Saves to 01_projects/matter-2026-001-acquisition/drafts/nda-draft-v1.md

You: Review the draft, request changes, approve.

The agent never sends anything to the client without your explicit approval. It drafts, you decide.

Example: Legal Research

You: "Research whether a listed company in Hong Kong needs shareholder approval for a change in auditors"

Agent:

Frames question → HKEX listing requirements for auditor changes
Identifies sources (priority order): HKEX Main Board Listing Rules → SFC codes → case law → academic commentary
Researches using available databases
Structures memo: Issue, Short Answer, Analysis, Authorities, Limitations
Cites specific rule sections (e.g., HKEX Main Board Listing Rule 17.05)
Flags any propositions that require verification
Saves to 01_projects/matter-2026-001-acquisition/research/2026-08-04 - Auditor Change Approval - ClientA.md

You: Review the memo, verify citations, rely on it for client advice.

The agent cites everything. If it can't find authority, it says so instead of making it up.

Ethical Guardrails

The system enforces non-negotiable guardrails:

Confidentiality & Privilege

All AI-processed data stays within the firm's secure environment
Client-privileged information never transmitted to external APIs without explicit consent
Wiki pages never contain client-privileged information

Hallucination Prevention

Zero-hallucination policy: every citation verified or flagged [REQUIRES VERIFICATION]
Mandatory human review before relying on AI output
Agent must state when it cannot find authority, not assert without citation

Audit Trail

Every action logged with timestamps, user identity, matter reference, output version
Draft versions preserved — agent never overwrites, only creates new versions
Session memory provides context continuity but must not contain privileged information

Scope Limitation

Task	Agent can do autonomously	Requires human sign-off
Draft contracts	Yes (saved to drafts/)	Before sending to client
Legal research	Yes (saved to research/)	Before relying in advice
Due diligence review	Yes (saved to research/)	Before including in DD report
Compliance checklists	Yes	Before sending to client
Client communications	Draft only	Always requires partner approval
Conflict checking	Search and flag	Partner must review results
Filing submissions	Draft only	Always requires partner approval

OneDrive Integration

The repo includes three options for connecting to your firm's file system:

Option A — Obsidian Vault on OneDrive (Simplest)

Point Obsidian at a OneDrive-synced directory. The agent operates on local files. OneDrive handles sync.

Best for: Solo practitioners.

Option B — Microsoft Graph API (Recommended for Firms)

Agent uses Microsoft Graph API to read/write files in OneDrive or SharePoint. OAuth 2.0 provides authenticated, audited access. Local PARA structure acts as working memory.

Best for: Mid-size to large firms.

Option C — SharePoint + Custom UI

Build a web interface backed by SharePoint. Lawyers access via browser.

Best for: Large firms needing multi-office access or client portals.

Full setup instructions in the Getting Started guide.

Phased Rollout Plan

The repo includes a 13-week implementation roadmap:

Phase	Timeline	Deliverables
1 — Discovery	Weeks 1–2	Task audit, stakeholder interviews, priority matrix
2 — Pilot Design	Weeks 3–4	Select 2–3 high-priority tasks, define success metrics, choose LLM provider
3 — Build & Test	Weeks 5–8	Vault setup, agent configuration, precedent ingestion, integration testing
4 — Controlled Rollout	Weeks 9–10	Deploy to single practice group with partner supervision
5 — Iteration	Weeks 11–12	Refine skills, expand clause library, address edge cases
6 — Firm-Wide	Week 13+	Roll out to all practice groups, establish governance

Success metrics:

30–40% reduction in document review time
>95% accuracy on clause identification and risk flagging
>80% user adoption among targeted associates
No degradation in client satisfaction
Net positive ROI within 6 months

Benefits

For Associates

Save 30–40% time on document review, research, and drafting
Focus on high-value work — analysis, strategy, client relationships
Learn faster — wiki captures institutional knowledge, checklists ensure consistency

For Partners

Reduce risk — built-in guardrails, audit trails, human oversight
Scale expertise — skills encode best practices, junior associates produce senior-level output
Improve client service — faster turnaround, more consistent quality

For the Firm

Competitive advantage — modern, efficient, tech-forward
Knowledge retention — wiki preserves institutional knowledge even as staff turnover
Cost recovery — licence cost vs. hours recovered = net positive ROI

What Makes This Different

This isn't another "AI will change law" think piece. It's a working system you can deploy today.

Open source — fork it, customize it, make it yours
Vendor agnostic — works with Claude, GPT, Qwen, Gemini, Llama
Privacy-first — self-hosted options, data never leaves your control
Ethically grounded — guardrails built in, not bolted on
Battle-tested structure — PARA methodology, proven in real legal work

Try It Yourself

GitHub repo: M0rtal13/Agentic-AI-for-Enterprise

Quick start:

git clone https://github.com/M0rtal13/Agentic-AI-for-Enterprise.git
cd Agentic-AI-for-Enterprise
bash setup.sh --vault ~/LawFirmVault

Then open the vault in Obsidian, configure your CLI agent (Claude Code, Qwen Code, Cursor, etc.), and start using the skills.

Full documentation in the docs/ folder:

Getting Started — LLM providers, CLI agents, API keys, OneDrive integration
Task Identification — 6 task domains with sub-tasks and priorities
Infrastructure — OneDrive options, PARA structure, AI wiki
Compliance — Ethical guardrails, audit trails, scope limitations
Roadmap — 13-week phased rollout plan

Final Thoughts

I'm a secondary 4 student in an IB bridging programme. I'm not a lawyer, I'm not a senior developer. But I noticed a problem during my internship — law firms waste time on repetitive work that AI can handle — so I built a solution.

The legal profession is conservative about technology. That's understandable — the stakes are high, mistakes are costly, and confidentiality is sacred. But that doesn't mean we can't move forward. It means we need to move forward carefully, with guardrails, audit trails, and human oversight.

This system is my attempt to show that it's possible. AI can assist lawyers without replacing them. It can handle the repetitive work while humans focus on judgment, strategy, and client relationships.

If you're a law firm exploring AI, or a student interested in legal tech, fork the repo, try it out, and let me know what you think.

GitHub: M0rtal13/Agentic-AI-for-Enterprise

Tags: #LegalTech #AIAgents #Obsidian #PARA #CorporateLaw #OpenSource

DEV Community: M0rtal13

How I Built an AI Agent System for a Corporate Law Firm — During My Internship

How I Built an AI Agent System for a Corporate Law Firm — During My Internship

The Backstory

The Problem: Law Firms Waste Time on Repetitive Work

The Solution: PARA + AI Agents + Obsidian

1. PARA Knowledge Management

2. CLI AI Agents

3. Six Legal Skills

How It Works in Practice

Example: Drafting an NDA

Example: Legal Research

Ethical Guardrails

Confidentiality & Privilege

Hallucination Prevention

Audit Trail

Scope Limitation

OneDrive Integration

Option A — Obsidian Vault on OneDrive (Simplest)

Option B — Microsoft Graph API (Recommended for Firms)

Option C — SharePoint + Custom UI

Phased Rollout Plan

Benefits

For Associates

For Partners

For the Firm

What Makes This Different

Try It Yourself

Final Thoughts