As law firms move artificial intelligence from small-scale experiments into everyday legal work, the cost of each query is becoming nearly as important as the quality of the answer.
NetDocuments says structured legal context could help firms improve both sides of that equation.
The legal document management company has published its Legal Context Engineering Benchmark, a new report measuring how structured context affects the accuracy and cost of AI-generated answers. According to the findings, the company’s Legal Context Graph reduced the cost of a correct answer by 48 per cent while delivering essentially the same answer quality.
For a 2,000-person firm making four million AI queries annually, NetDocuments estimates that efficiency could produce nearly $1 million in savings each year.
“Get the context right, and firms spend less to get better answers,” said NetDocuments chief executive officer Josh Baxter.
The benchmark held the AI model, agent, questions and source documents constant, changing only whether the system had access to structured legal context. The same agent answered 300 questions across 10 legal matters, covering 874 documents and approximately 60 million characters drawn from public regulatory filings and court dockets.
By isolating context as the variable, NetDocuments sought to test a growing challenge for legal AI: giving a system access to documents does not necessarily mean it understands how those documents relate to the work at hand.
An AI tool may be able to search a firm’s files without knowing which matters are connected, which precedent the firm relies on, which document version is authoritative or how information distributed across a matter fits together. Without that understanding, an agent may need to reconstruct those relationships each time it receives a question, consuming additional tokens and computing resources.
NetDocuments’ Legal Context Graph, announced in May, is designed to turn a firm’s documents into persistent, permission-aware context that connected AI tools can use from the outset.
The benchmark identified two potential approaches for firms. Those focused on savings could maintain answer quality while lowering costs by 48 per cent. Firms could also reinvest some of the efficiency in deeper AI reasoning: in that test, answer quality improved by seven per cent while total costs remained 18 per cent lower.
Token consumption per answer fell by 52 per cent when the Legal Context Graph was used. NetDocuments also reported that the benefit increased with more capable models, suggesting that investments in context could become more valuable as the underlying technology improves.
Scott Kelly, vice-president of product at NetDocuments, said the legal industry’s AI conversation has shifted over the past six months from broadly available capabilities to the context required for dependable outcomes. He said the benchmark is intended to give firms evidence for evaluating whether their AI investments deliver lasting value to lawyers, clients and the business.
The findings arrive as AI providers increasingly adopt consumption-based and token-based pricing. That change places greater pressure on firms to measure not only whether an AI system can complete a task, but also how efficiently it can produce a reliable result at scale.
NetDocuments is also using Model Context Protocol integrations to make its structured context available to outside AI platforms. Its connector ecosystem includes ChatGPT, Claude, Microsoft Copilot and Perplexity, as well as legal technology platforms Harvey and Legora.
Along with its results, NetDocuments published the benchmark methodology to give legal organizations a framework for tracking the quality and cost of their own AI systems over time.
The Legal Context Engineering Benchmark report is available now from NetDocuments.