In a recent Platforum9 session, Horace Wu, CEO and founder of Syntheia.io, shared his valuable perspectives on how lawyers and technologists can collaborate to build more effective legal tech solutions. Drawing from his decade of experience as an M&A lawyer before transitioning to legal tech, Wu offered candid insights into the current state of AI in legal practice and where opportunities for meaningful innovation truly exist.
The “Jaggedy Frontier” of AI in Legal Practice
Wu began by challenging some common assumptions about AI’s role in legal drafting. “People both overestimate and underestimate the impact of generative AI,” he observed, describing what he calls the “jaggedy frontier” of AI capabilities.
While significant investment has focused on using generative AI to draft new contracts or markup documents, Wu questioned this approach: “You speak to any lawyer, nine times out of ten, the reaction is ‘Why would I use AI to draft something new when I’ve got a precedent that I know has worked?'”
This highlights a fundamental tension in legal tech development—the difference between what technology could do versus what lawyers actually need it to do. Wu suggests that technologists and lawyers must collaborate to identify where AI truly adds value rather than simply applying technology because it’s available.
The Copy-Paste Reality of Legal Practice
Wu challenged the assumption that lawyers routinely need to create documents from scratch. “It is exceedingly rare when someone needs to reinvent something from scratch in a legal context,” he explained, comparing legal drafting to home construction: “Only one in 10,000 houses would you need a creative architect to do something that hasn’t been done before. The other 9,999, you are using the same well-known technologies and the same well-known techniques.”
From a productivity perspective, Wu argued that reliable copy-paste from proven precedents is far more efficient than generating new language that then requires extensive review. This insight underscores why some generative AI applications have struggled to gain traction among practicing lawyers despite impressive capabilities.
The True Value: Data Analysis and Pattern Recognition
While Wu expressed skepticism about AI-generated contract drafting, he was enthusiastic about AI’s potential for data analysis and pattern recognition. When asked about extracting insights from existing documents, he noted that “using the model for synthesis of information and discovery of information” represents a “great use case.”
For tasks like due diligence and discovery, Wu explained that “generative models have the ability to synthesize a lot more information and do it on repeat that humans just can’t do.” This capability, combined with cost advantages compared to human review, makes AI particularly valuable for analyzing large document collections to identify patterns and anomalies.
However, Wu added an important caveat: current models rely on their “parametric memory” or training data from the internet, which means they may not match the expertise of specialists who practice in specific legal fields every day. This underscores the continued importance of domain expertise in interpreting AI-generated insights.
Understanding AI: Training vs. Fine-Tuning vs. Prompt Tuning
Wu provided clarity on frequently confused terms in AI development, distinguishing between three different processes:
- Training the foundation model – An extraordinarily expensive process involving training a neural network on vast amounts of internet text. Wu stated unequivocally that “literally nobody in legal is doing that” with the possible exception of Mike Bommarito’s team at 273 Ventures.
- Fine-tuning – Either adjusting the foundation model with new data or using reinforcement learning to teach the model to prefer certain responses.
- Prompt tuning – Simply providing examples and context within prompts to guide the model’s outputs.
This technical distinction carries practical implications for lawyers. Wu suggested that most legal professionals don’t need deep technical knowledge about AI: “As the average lawyer, you don’t need to know anything about this… As long as you have vague ideas of what the technology can and can’t do, you don’t need to become a technologist.”
The “Prompt Engineer” Debate
Wu took a provocative stance on the emerging role of “prompt engineers,” calling it a “nonsense term.” He argued, “You’re not an engineer. You’re a prompter. People who prompt can be very good at it, just like an artist is very good at art. You are not an art engineer.”
Despite this terminology critique, Wu acknowledged the importance of clear communication in prompting, comparing it to senior lawyers giving instructions to juniors: “Whether it’s a junior lawyer or an AI tool, it should be the same clarity.” He noted that this skill is particularly valuable since lawyers “are not very clear about what their expected outcomes” often are.
The Missing Axis: Context and Personalisation
Wu highlighted a fundamental limitation in current legal tech approaches. While foundation models are continuously improving on a “horizontal axis” of general capabilities (from GPT-3 to GPT-4 and beyond), they lack progress on the “vertical axis” of firm-specific and lawyer-specific context.
“The smartest models in the world are still not going to have the knowledge and the internal information of a lawyer, of a law firm, to be able to give contextually relevant responses,” Wu explained. “That missing y-axis, where as you go up the axis, you get more personalised and more relevant context, that is missing in 99% of the tools out there.”
This insight forms the core of Syntheia’s approach. Rather than focusing solely on foundation model capabilities, they provide “point solutions” that solve specific problems while building a knowledge management system in the background. This creates a feedback loop where solving immediate problems enhances the system’s ability to provide more personalised responses over time.
Building Through Collaboration
Wu revealed that Syntheia’s development model centers on collaboration with large firms: “Predominantly all of our tools have been created because we’ve collaborated with large firms.” This approach ensures their solutions address genuine practitioner needs rather than theoretical use cases.
While initially focused on AM Law 100 firms, Wu announced that Syntheia is now expanding access through self-service offerings like SuperComparer (supercomparer.com), which enables document comparison even when documents are structurally different. This tool addresses specific pain points experienced by transactional lawyers—identifying differences between documents and understanding market positions across multiple deals.
The Future of Legal Practice with AI
Throughout the discussion, Wu emphasised that AI tools will augment rather than replace legal expertise. He stressed that the most successful lawyers will combine technical legal knowledge with the ability to provide personalized, contextually relevant advice—something AI alone cannot deliver.
Wu concluded with a call for lawyers to develop a realistic understanding of AI capabilities: “If you overestimate the capability of these models, then you’re like, ‘Oh my gosh, my job is at risk.’ And if you underestimate it, then you [say] ‘I’m never going to adopt it.’ And both sides are wrong.”
For legal tech developers and law firms alike, the path forward lies in collaboration that respects both technological possibilities and practical realities of legal practice. By focusing on specific problems where AI genuinely adds value and integrating firm-specific context, legal tech can enhance rather than disrupt the lawyer-client relationship.
