Wilson Sonsini achieves 92% accuracy on complex contract review with Dioptra AI

Large commercial law firms handling complex, high-stakes agreements face a critical gap between legal AI's marketed capabilities and its real-world performance. Many AI document review tools demonstrate high 'demo accuracy' on simple clause detection but fail on what Wilson Sonsini's Head of Innovation David Wang calls 'dimensionality' — the interconnected nature of sophisticated contracts where a single language change can cascade through 20 or more interdependent provisions. With over 40% of in-house counsel at technology companies reporting that contract-related tasks consume more than half their working day, the cost of unreliable AI output is compounded: partners either avoid AI tools for sensitive matters entirely, or spend so much time correcting outputs that efficiency gains disappear.

Wilson Sonsini deployed Dioptra, a stealth-mode legal AI startup operating on large language models and generative AI, through its Neuron innovation group — an internal team focused on digital solutions for startup clients. Rather than ripping and replacing existing contract language wholesale, Dioptra's approach mirrors how a skilled human editor works: it parses intent, preserves what is already fit for purpose, and makes targeted, contextually-aware adjustments to reach the correct legal position. This minimizes unnecessary redlining and reduces the downstream review burden. The initial deployment focused on cloud services agreements, a high-frequency contract type for Wilson Sonsini's tech-sector client base. The service is structured as a fixed-fee offering, making it commercially predictable for both the firm and its clients.

The implementation reaches 92% consistent accuracy on third-party documents — a benchmark the firm defines not by clause detection rates but by the standard Wilson Sonsini calls the 'partner assessment' test: output that a partner would accept and stand behind, subject only to minor human edits. This level of performance makes fixed-fee pricing viable by bounding the human review effort required per engagement.

• 92% consistent accuracy on third-party cloud services agreements
• Output quality meets internal partner sign-off standards with only minor human edits required
• Fixed-fee pricing model enabled by predictable, reliable AI accuracy
• Clients gain faster time-to-revenue by accelerating contract turnaround on agreements that unlock commercial relationships

• Accuracy is the prerequisite for efficiency — AI tools that require extensive partner correction eliminate the time savings they were meant to create; adoption at the partner level only follows when output quality is demonstrably trustworthy.
• Dimensionality is the real benchmark — evaluating legal AI on simple clause detection understates the challenge; the meaningful test is handling contracts where changes to one provision ripple through many others.
• Fixed-fee pricing is a downstream indicator of AI maturity — it only becomes sustainable when accuracy is high enough to make human review time predictable and bounded.
• Piloting on high-frequency, well-defined contract types (such as cloud services agreements) creates a controlled environment to validate accuracy before expanding to broader agreement categories.