AI document review pipeline.
The associate stops reading 200 pages to find three problematic clauses.
Status
Sector
Legal · Corporate firm
Client
Corporate law firm · MX
Year
2026
Hook ─ Position
Premium legal work isn't reading documents — it's deciding on them. AI doesn't replace the lawyer; it gives back the time lost on mechanical review.
Elige un contrato y mira al sistema operar.
Pipeline real · salidas guionadas para demo
Theory ─ How we think
AI automation fails when it replaces professional judgment. It works when it gives the professional back the time spent on mechanical review so they can focus on what actually requires judgment.
- 01
Taxonomy defined by partners, not by the model
The 14 risk categories were defined by partners across 3 sessions. The model never decides what risk is — only finds it.
- 02
Deterministic pipeline, not open agent
Ingest → split by clause → parallel analysis → consolidation. Every step is auditable. No step is 'a conversation'.
- 03
Exact citation, always
Every flag includes the full clause, page number, exact position. The lawyer verifies in 10 seconds what used to take 30 minutes to read.
- 04
The lawyer always signs
The system doesn't deliver results to the client. It delivers a report that the lawyer reviews, edits, and signs. Responsibility never changes hands.
Visual ─ Architecture
Ingest · Split · Analysis · Synthesis.
Internal portal
upload · SSO auth
Splitter
clause by clause
AI agents
14 parallel categories
Report
signable markdown
The associate uploads the contract. The system segments it by clause and fires 14 parallel analyses (one per risk category). Results consolidate into a markdown report the lawyer can edit before signing.
How it shipped.
−72%
sustained average over 3 months
+31%
vs historical manual review
−61%
internalized by hour
─ Deliverables
- ✓Internal portal with firm SSO
- ✓AI pipeline with 14 configurable categories
- ✓Reports with exact citations and page numbers
- ✓Integration with existing document management system
- ✓Audit panel: what was processed, when, by whom
─ Stack
- Next.js 15
- Auth.js + SSO
- shadcn/ui
- Claude Sonnet 4.6
- Structured Outputs
- pgvector
- Inngest
- Postgres
- Cloudflare R2
─ Timeline
12 weeks · included calibration with 200 historical contracts
Testimonial ─ Client
“I used to spend the morning reading. Now I spend the morning deciding. That's the difference between associate and partner.”
— Senior associate · Corporate firm MX
Does your case look like this?