How We Got Legal Document Review Down to 13% of the Usual Cost

Legal document review is one of those problems where the economics just do not make sense anymore. A junior associate at a mid-size firm bills $50 to $150 per hour. They read documents one at a time. They flag privilege issues, extract contract terms, and assess risk. For a case with 500 documents, you are looking at weeks of work and a bill north of $40,000.

We built the Legal Discovery tool to change those numbers. Three AI agents work in parallel on every document. The total cost per 100-document batch comes to roughly $6.50 in credits. That is 13% of what even the cheapest human review would cost, and the results come back in minutes, not weeks.

Three Agents, One Document, Parallel Execution

The core of the Legal Discovery tool is an multi-agent AI system. AI agent (formerly AutoGen) lets us define specialized AI agents that work together on a problem. For legal discovery, we run three agents simultaneously on every document:

The parallel execution matters. A single agent reading documents sequentially would still be fast compared to human review, but running three agents at once means each document gets analyzed from three different angles in the time it takes for one pass. The results merge into a single structured report per document.

How the Output is Structured

Every document produces a structured result with three sections. This is not free-form text that someone has to interpret. Each field is typed and consistent across all documents in a batch.

100-Document Batches

You can process documents one at a time, but the real value is in batches. Upload up to 100 documents at once, and the system queues them for parallel processing. Each document gets all three agents.

For a 100-document batch, the timeline actually looks like this: upload takes a few seconds (depending on file sizes), processing takes 5 to 15 minutes depending on document length and complexity, and then you have a complete structured report you can download as PDF.

The PDF export uses PyMuPDF's Story API. We chose it over ReportLab or WeasyPrint because it handles long, structured documents well and produces consistent formatting. Each document gets its own section in the PDF with privilege flags, extracted terms, and risk scores laid out in a clean, readable format.

Persistent Legal Context via Knowledge Bases

Something that makes the tool more useful over time. The Legal Discovery tool uses a system assistant with a dedicated Knowledge Base. As you process documents, the system builds up context about your legal matters.

The system assistant is created with is_system=True and the identifier __system__legal_discovery__. Its Knowledge Base stores chunked and embedded versions of your processed documents, so when you ask follow-up questions or process new documents related to the same matter, the system already has context.

In practice, this means: process 50 contracts from a vendor relationship, then ask "which contracts have indemnification caps below $1 million?" and get an answer in seconds, with citations to specific documents.

The Math on Cost Savings

Let us be specific about where the 13% number comes from.

A junior associate at $75/hour (the low end of the $50-$150 range for document review) takes roughly 20 minutes per document for basic privilege and term review. That is $25 per document, or $2,500 for 100 documents.

Our tool processes the same 100 documents for approximately $6.50 in credits. That is 0.26% of the human cost. Even if you add in a senior attorney spending 2 hours reviewing the AI output (which you should), you are still at roughly $325 total versus $2,500. That is 13%.

And that is the conservative comparison. Large firms using senior associates at $150/hour for initial review are looking at $5,000 for the same batch. The savings scale directly with document volume.

What This Does Not Replace

We are not claiming this replaces lawyers. That would be irresponsible and also wrong.

What the tool replaces is the grunt work phase of document review. The initial pass where a junior associate reads every document, highlights potentially privileged material, and extracts basic terms. That phase is mechanical, repetitive, and expensive.

After the AI does that first pass, a qualified attorney still needs to review the flagged items, verify the extracted terms, and make judgment calls on the risk assessments. But instead of starting from scratch with a stack of raw documents, they start with structured data, flagged issues, and scored risks. Their time goes to judgment, not data entry.

The privilege detection in particular should always be verified by a human. Missing a privileged document in discovery has real consequences. Our agent catches the obvious patterns (attorney-client headers, work product language, confidentiality markers), but novel or ambiguous privilege issues still need human eyes.

Getting Started

The Legal Discovery tool sits inside your NeuroGen dashboard under the Tools section. Upload documents (PDF, DOCX, or run them through the File Processor first), select batch processing, and the three agents start working.

Results appear in your dashboard as they complete, and you can export the full batch report as PDF at any time. The Knowledge Base builds automatically in the background, so your second batch is even more useful than your first.

Back to Blog