YC
Why Vertical LLM Agents Are The New $1 Billion SaaS Opportunities
@ Y Combinator
10/04/24
Takeaways
JH
Jake Heller
10/04/24
@ Y Combinator
I saw the first version of GPT-4, and within 48 hours, we decided to shift the entire company to focus on building a new product called Co-Counsel based on that technology.
JH
Jake Heller
10/04/24
@ Y Combinator
In the early days of Casetext, I aimed to bring the best of technology to the legal space, as I found the existing legal technology frustrating and inefficient.
JH
Jake Heller
10/04/24
@ Y Combinator
We initially tried to create a user-generated content site for lawyers to annotate case law, but it failed because lawyers are too busy to contribute their time.
JH
Jake Heller
10/04/24
@ Y Combinator
When ChatGPT was released, lawyers realized it would fundamentally change their work, prompting them to seek out AI solutions like ours.
JH
Jake Heller
10/04/24
@ Y Combinator
Before launching Co-Counsel, we were already making real revenue and had positive feedback from customers, indicating we were close to product-market fit.
JH
Jake Heller
10/04/24
@ Y Combinator
The launch of Co-Counsel led to overwhelming demand, with our servers going down and an inability to hire support and sales staff quickly enough, which is a sign of real product-market fit. We built an AI Legal Assistant that acts like a new member of the law firm, capable of reading millions of documents and summarizing findings, which significantly extends the workforce of law firms.
JH
Jake Heller
10/04/24
@ Y Combinator
Before the public launch of GPT-4, we had law firms using our AI without knowing it was powered by GPT-4, and they were amazed by its capabilities, completing tasks that previously took a whole day in just minutes.
JH
Jake Heller
10/04/24
@ Y Combinator
During the months leading up to the public launch of GPT-4, our entire team was intensely focused, which allowed us to iterate quickly and stay ahead of the market.
JH
Jake Heller
10/04/24
@ Y Combinator
Convincing employees to embrace the shift towards AI was challenging, especially when the company was already performing well with a 70-80% growth rate and $15-$20 million in ARR.
JH
Jake Heller
10/04/24
@ Y Combinator
I led by example and built the first version of our AI product myself, which helped persuade the team and stakeholders about its potential.
JH
Jake Heller
10/04/24
@ Y Combinator
We brought in customers early to test our product, which helped change the minds of skeptical team members when they saw real-time reactions from users.
JH
Jake Heller
10/04/24
@ Y Combinator
The release of GPT-4 marked a significant turning point, as it performed much better than previous models, achieving results that surpassed 90% of test takers in legal assessments.
JH
Jake Heller
10/04/24
@ Y Combinator
We approached the development of our AI by breaking down the legal research process into steps, mimicking how the best attorneys would tackle a problem.
JH
Jake Heller
10/04/24
@ Y Combinator
Our development process involved creating a series of tests to ensure that the AI could accurately respond to legal queries, which was crucial for its effectiveness in practice.
JH
Jake Heller
10/04/24
@ Y Combinator
We adopted a test-driven development approach for our AI prompts, which allowed us to refine the system to achieve high accuracy in responses.
JH
Jake Heller
10/04/24
@ Y Combinator
When solving a problem for a customer, you need to add many layers to get the job done effectively, which makes you more than just a 'GPT wrapper'—you're creating a full application that integrates proprietary data and specific legal document management systems.
JH
Jake Heller
10/04/24
@ Y Combinator
The transition from 70% to 100% accuracy in AI applications is significant; users are willing to pay substantially more for reliable solutions that work consistently compared to those that only work part of the time.
JH
Jake Heller
10/04/24
@ Y Combinator
Achieving 100% accuracy in AI applications is crucial, especially in legal contexts where providing incorrect information can have serious consequences for lawyers handling important cases.
JH
Jake Heller
10/04/24
@ Y Combinator
Using a test-driven development framework helps identify patterns in AI mistakes, allowing you to refine instructions and improve accuracy over time.
JH
Jake Heller
10/04/24
@ Y Combinator
It's essential to make the first encounter with AI technology successful for users, especially in conservative fields like law, to encourage deeper investment in the technology.
JH
Jake Heller
10/04/24
@ Y Combinator
The latest AI models, like OpenAI's, demonstrate a higher degree of precision and thoroughness in understanding nuanced legal texts, which previous models struggled with.
JH
Jake Heller
10/04/24
@ Y Combinator
Injecting domain expertise into AI systems can enhance their problem-solving capabilities, making them more effective in specific fields like law.
JH
Jake Heller
10/04/24
@ Y Combinator
Many companies still lack an understanding of the advancements in AI technology, often repeating outdated beliefs about its limitations, which can hinder innovation.
JH
Jake Heller
10/04/24
@ Y Combinator
The future of AI in various fields, including law, holds great potential for efficiency and strategic thinking, reducing the time and cost associated with traditional methods.
JH
Jake Heller
10/04/24
@ Y Combinator
Despite concerns about AI hallucinations and inaccuracies, there are pathways to improve these technologies, and the jobs they create will evolve to be more interesting rather than disappear.