92%
retrieval accuracy on the firm's eval set
6h→40m
typical contract review time
30k+
documents indexed
5 months
from kickoff to delivery
The Client
A mid-size law firm with a strong corporate and commercial practice — dozens of attorneys handling M&A support, commercial agreements and ongoing counsel for business clients. Two decades of work product had accumulated as more than 30,000 contracts and related documents spread across shared drives, email archives and a document management system used inconsistently.
The Challenge
Institutional knowledge lived in senior attorneys' memories. Questions like "what indemnification caps have we agreed for this client before?" or "which of our agreements contain change-of-control triggers?" meant hours of manual digging — or simply going with whatever the attorney remembered. Due-diligence projects were staffed with associates reading contracts one by one, and obligations buried in executed agreements (renewal windows, notice deadlines) were tracked in ad hoc spreadsheets, when they were tracked at all.
The partners were explicit about the bar for any AI system: answers had to cite the exact clause and document they came from, hallucinated text was disqualifying, and client confidentiality demanded the whole system run inside the firm's own cloud environment. A generic chatbot bolted onto a folder of PDFs was not going to pass.
What We Did
Dillo delivered the platform as an outsourced AI/RAG project — a compact team of AI engineers and a full-stack engineer, led by a Dillo delivery lead, with two of the firm's attorneys embedded as domain reviewers from week one.
- Built an ingestion pipeline that normalized the document sprawl — Word files, scanned PDFs with OCR, email attachments — deduplicated versions, and indexed everything into pgvector with clause-aware chunking that respects contract structure.
- Implemented retrieval-augmented search on the Claude API: attorneys ask questions in plain language and get answers grounded in retrieved passages, with every claim linked to the exact clause and document it came from. No citation, no answer.
- Added clause extraction across key categories — indemnification, limitation of liability, termination, change of control, governing law — enabling portfolio-wide queries and side-by-side clause comparison.
- Generated obligation calendars from executed agreements: renewal windows, notice deadlines and payment milestones extracted into a review queue, confirmed by an attorney before anything hits a calendar.
- Built an evaluation suite with the firm's attorneys — hundreds of real question-answer pairs with known correct sources — run on every change to retrieval, chunking or prompts. Nothing shipped unless the eval score held.
The application layer is a Next.js interface with matter-level access controls mirroring the firm's ethical walls. The entire system runs in the firm's own cloud tenancy; no client documents leave their environment for training or any other purpose.
The Results
- 92% retrieval accuracy on the firm's own eval set — the metric partners tracked from pilot through sign-off.
- Typical contract review dropped from 6 hours to about 40 minutes: attorneys start from extracted clauses and cited answers instead of a blank document.
- 30,000+ documents indexed and searchable, with new executed agreements flowing in automatically.
- Obligation calendars surfaced dozens of upcoming renewal and notice deadlines in the first months — several of which no one had been tracking.
- Delivered in 5 months from kickoff, including the eval suite the firm now uses to gate every future change.
Figures reflect outcomes reported for this engagement; they are project results, not audited benchmarks.
Tech Stack
Services Used
Software Outsourcing — an AI/RAG delivery team building the contract-intelligence platform end to end, on transparent pricing.