![\"\"](\"https:\/\/www.artificiallawyer.com\/wp-content\/uploads\/2025\/06\/Screenshot-2025-06-17-at-10.50.49-678x381.png\" "\\"Screenshot")
\n<\/figure>\n
In a very creative approach to using genAI for contract review, Syntracts \u2013 co-founded by Uber and Latham & Watkins alums \u2013 doesn\u2019t need prompts, and instead uses orchestrated small language models (SLMs) trained on synthetic data, each focused on different aspects of a contract to provide lawyers with very high accuracy \u2013 and without hallucinations.<\/p>\n
Syntracts is also different in another way: it operates on-prem and thus doesn\u2019t need to get into the cloud at all, thus providing law firms and inhouse teams with, in theory, much better security.<\/p>\n
In short, this is a real departure from how many genAI tools are deployed, and especially for contract review.<\/p>\n
As the US-based company explained: \u2018Unlike legal AI tools that rely on third-party APIs, cloud-hosted LLMs, or brittle prompt engineering, Syntracts takes a fundamentally different approach.<\/p>\n
\u2018Its proprietary, open-source models are trained using synthetic legal data<\/strong>, which is generated from each customer\u2019s own legal documents. This results in highly accurate, structured outputs built for real-world legal workflows, all while keeping sensitive information entirely on-prem.\u2019<\/p>\n Doug Bemis<\/strong>, co-founder of Syntracts and former CTO at Uber AI Labs, told Artificial Lawyer<\/strong>: \u2018You can\u2019t turn a generalist LLM into a legal expert just by prompting. LLMs should be perfect for law, but they haven\u2019t delivered. Legal work demands clarity, consistency, and control \u2013 none of which prompting can guarantee. Syntracts is for firms that need trustworthy performance from day one, with models trained on firm-specific synthetic data and deployed fully on-prem to produce structured, database-ready outputs.\u2019<\/p>\n He added: \u2018LLMs are trained to sound right, not be right. People try and prompt LLMs into submission<\/strong>. But, we use a different approach.<\/p>\n \u2018We create a foundational data layer that can handle legal language. We divide contracts into small pieces, then use synthetic data to fine tune small language models<\/strong>. We use multiple SLMs and then orchestrate \u2013 but we don\u2019t prompt and don\u2019t need lawyers to handle prompts.\u2019<\/p>\n Meanwhile, fellow co-founder Christopher Martin<\/strong>, who was a senior lawyer at Latham & Watkins and served as Manager of Emerging Technology, told this site: \u2018I have been disappointed in LLMs\u2019 ability to handle real legal workflows. You need a data foundation to do interesting things and in Big Law even a small mistake can have a big impact.\u2019<\/p>\n Martin explained that Bemis was an old friend and they had wanted to work together in the past, but now the time was right and together they have developed an approach to achieve what they wanted.<\/p>\n What they have built can deliver:<\/p>\n As noted, this is a long way from just tapping one of OpenAI\u2019s models to give a contract a once-over.<\/p>\n Martin added: \u2018Prompt-based systems don\u2019t scale \u2013 you shouldn\u2019t need more people to make AI work. We built Syntracts to meet the strictest data and compliance standards while delivering real results: structured, high-accuracy outputs based on how lawyers actually read contracts.\u2019<\/p>\n The company also noted they\u2019ve just signed a deal with a Top 25 AmLaw firm and are also in A&O Shearman\u2019s Fuse Incubator.<\/p>\n You can find more about Syntracts here<\/a>.<\/p>\n Subscribe to get the latest posts sent to your email.<\/p>\n<\/div>\n<\/div><\/div>\n\n
![\"\"](\"https:\/\/www.artificiallawyer.com\/wp-content\/uploads\/2025\/06\/Syntracts-Diagram-1024x576.png\")
<\/figure>\n
\nDiscover more from Artificial Lawyer<\/h3>\n