The Transformative Potential of Vertical AI in Business

The discussion opens with a reflection on the rapid advancements in vertical AI agents, which are poised to replace entire teams and functions within enterprises. The speaker expresses amazement at this progression, noting that the competitive landscape has shifted from a single dominant player, OpenAI, to a more diverse ecosystem, fostering consumer choice and opportunities for founders.

Gary introduces the episode of 'The Light Cone' alongside co-hosts Jared Harge and Diana, highlighting their collective experience in funding startups worth hundreds of billions of dollars, particularly in their early stages.

Jared passionately discusses the potential of vertical AI, emphasizing that many startup founders, especially younger ones, may not fully grasp the magnitude of this emerging sector. He argues that vertical AI agents are not a novel concept, as they have been funded previously, but the broader market has yet to recognize their potential, predicting the emergence of companies worth over $300 billion in this category.

Jared draws an analogy between vertical AI and Software as a Service (SaaS), explaining that many startup founders view the industry through the lens of consumer products, which often overlook the prevalence of SaaS tools designed for businesses. He highlights that over 40% of venture capital funding in the last 20 years has gone to SaaS companies, resulting in more than 300 SaaS unicorns, underscoring the sector's significance.

Reflecting on the history of technology, Jared identifies the XML HTTP request as a pivotal catalyst for the SaaS boom, enabling the development of rich internet applications. This innovation allowed web applications to function similarly to desktop applications, leading to the creation of services like Google Maps and Gmail, and marking a shift from traditional software installation to web-based usage.

The conversation touches on Paul Graham's early recognition of the potential of HTTP requests, which he connected to Unix prompts, effectively creating one of the first SaaS applications in 1995. However, the initial SaaS applications struggled due to poor user experience, as they required full page reloads with each interaction, limiting their adoption until the widespread implementation of XML HTTP requests in 2005.

Jared likens the current advancements in large language models (LLMs) to the transformative impact of XML HTTP requests, suggesting that LLMs represent a new computing paradigm that enables fundamentally different applications. He notes that, similar to the questions raised during the rise of cloud and mobile technologies, there is now a pressing inquiry into how to leverage this new technology effectively and identify valuable opportunities.

The speaker categorizes billion-dollar startups into three distinct buckets. The first category includes obvious mass consumer products like documents, photos, email, calendars, and chat applications, which have predominantly been dominated by incumbents such as Google, Facebook, and Amazon. Notably, despite numerous attempts by various startups to bring Microsoft Office online, Google ultimately prevailed. The second category consists of unexpected mass consumer ideas, exemplified by companies like Uber, Instacart, DoorDash, Coinbase, and Airbnb, which emerged unexpectedly and were not initially pursued by incumbents until it was too late. The third category encompasses B2B SaaS companies, with over 300 billion-dollar firms, indicating a significant number of successful startups in this space. The speaker notes that there is no single dominant company in the SaaS sector, unlike in the consumer product space, which has led to a proliferation of diverse companies.

The speaker reflects on the early skepticism surrounding SaaS, recalling Mark Benioff's experiences at Y Combinator, where he shared that many doubted the feasibility of sophisticated enterprise applications being delivered via the cloud. This skepticism stemmed from a perception that traditional box software was the only reliable option. The speaker draws parallels between the early days of SaaS and the current hesitance regarding the use of AI and LLMs in enterprise applications, suggesting that similar doubts about their capabilities may hinder adoption despite their potential.

The discussion highlights the classic innovator's dilemma, particularly in the context of large companies like Google. The speaker argues that these incumbents often avoid pursuing risky innovations, such as Uber and Airbnb, due to the potential threat to their existing revenue streams. The fear of jeopardizing their established 'pot of gold' leads to a reluctance to clone or develop new products, even when the success of these startups becomes evident. This dynamic illustrates the challenges faced by established companies in adapting to disruptive innovations.

Travis Kalanick, during the early years of Uber, expressed significant fear of facing prison time due to the risks involved in building the company. This highlights the intense personal stakes and challenges faced by entrepreneurs in the tech industry, contrasting with the more secure positions held by employees at established companies like Google.

The discussion delves into why incumbents like Google have not ventured into B2B SaaS markets. It is suggested that the complexity and niche requirements of B2B SaaS products necessitate deep domain expertise, which large companies may lack. For instance, Gusto, a payroll service, thrives because it addresses intricate payroll regulations that larger firms may find cumbersome to navigate.

The conversation touches on the evolution of software sales from expensive, box-based installations to SaaS solutions. Salesforce's success demonstrated that SaaS could be as powerful as traditional enterprise software, paving the way for vertical SaaS solutions. This shift has allowed companies to create specialized products that significantly enhance user experience compared to the 'jack of all trades' approach of legacy systems like Oracle and SAP.

The speaker outlines three primary pricing tiers in software: $5 per seat for consumer products, $500 per seat for SMBs, and $55,000 per seat for enterprise solutions. This pricing structure reflects the varying needs and budgets of different market segments, with enterprise software often being criticized for poor user experience due to the disconnect between decision-makers and end-users.

There is curiosity about how the advent of large language models (LLMs) will transform the software landscape. Historically, as revenue scales, the workforce tends to grow, leading to companies with substantial revenue but large employee counts. The speaker notes that many unicorns today, including those in Y Combinator's portfolio, reach $100 to $200 million in revenue while employing 500 to 2,000 people, raising questions about efficiency and future trends in workforce management.

The speaker reflects on a shift in hiring strategies over the past couple of years, suggesting that instead of solely seeking out top talent from other departments like customer success or sales, there is now a growing need for skilled software engineers, particularly those knowledgeable in large language models (LLMs). This change is seen as a response to the need for automation to address bottlenecks in growth, indicating a potential revolution in how startups operate post-product-market fit. The speaker envisions a future where companies could thrive with significantly fewer employees, even as few as ten, by leveraging LLM systems to reduce costs and streamline operations.

The discussion highlights a personal experience from the speaker's time at Triplebyte, where traditional marketing strategies were challenged. Instead of hiring a conventional marketing executive after raising a Series B, the speaker recalls collaborating with an MIT engineer, Mike, who had developed a smart frying pan. Mike's engineering mindset proved invaluable in navigating paid advertising and Google Ads, leading to a marketing budget that peaked at one million dollars per month. The speaker emphasizes that this approach resulted in high-quality marketing campaigns that belied the company's size, demonstrating how a skilled engineer can create significant leverage in marketing efforts.

The speaker proposes the concept of 300 vertical AI agent unicorns, suggesting that every Software as a Service (SaaS) unicorn could have a corresponding vertical AI equivalent. This idea stems from the observation that many SaaS companies previously disrupted traditional software models, and now, a similar disruption could occur where vertical AI combines software with human expertise into a single product. The speaker notes that enterprises are currently uncertain about the specific AI agents they require, but experienced founders, like Brett Taylor, are exploring ways to help businesses deploy custom AI agents tailored to their needs.

The discussion begins with a reference to Vector Shift, a company founded about a year ago by two Harvard computer scientists. They are developing a platform that enables enterprises to create their own internal LLM-powered agents using no-code tools or SDKs. However, enterprises often struggle to identify specific use cases for these agents. The conversation draws a parallel to the early days of software as a service (SaaS), where initial vendors focused on convincing users to adopt software, leading to a more sophisticated landscape with specialized vertical solutions over time. The question arises whether the same evolution will occur with LLMs, where initial general-purpose solutions will eventually give way to more specialized vertical agents.

Reflecting on the history of SaaS, it is noted that consumer applications, such as email and chat, gained traction from 2005 to 2010, which facilitated the adoption of SaaS tools in enterprises. Employees, who are also consumers, became accustomed to using these tools, making it easier to sell SaaS solutions to companies. The speaker suggests that LLMs do not need to revert to a few general-purpose platforms, as enterprises have already recognized the value of point solutions and vertical solutions. This established belief may lead enterprises to embrace startups offering specialized AI agents, resulting in faster traction for these solutions.

The conversation emphasizes that the software industry typically starts with vertical solutions before evolving into broader applications. As companies grow, they often begin with specific point solutions, but eventually need to expand horizontally to capture a larger market share. The speaker highlights that the potential for vertical AI agents could surpass that of traditional SaaS, as these agents may not only replace existing software but also reduce payroll costs significantly. Companies currently spend a substantial amount on employee salaries compared to software, suggesting that more efficient vertical solutions could lead to a reduced need for human intervention in tasks like data entry and approvals.

The discussion concludes with the assertion that vertical AI solutions could potentially be ten times larger than the SaaS companies they aim to disrupt. There is a possibility that these vertical point solutions could be sufficiently large on their own, negating the need for broader applications. The speaker suggests that examples of successful vertical AI agent companies could illustrate this trend, hinting at positive developments in the sector.

The speaker discusses a Y Combinator company called Outset, which focuses on integrating large language models (LLMs) into the surveys and Qualtrics space. They highlight that Qualtrics is unlikely to develop a top-tier LLM with reasoning capabilities, emphasizing that surveys are fundamentally about language and understanding customer needs.

The conversation shifts to the challenges of selling enterprise and SMB software, particularly the necessity of addressing concerns from key decision-makers within organizations. The speaker notes that selling to teams at risk of being replaced by AI can lead to sabotage, necessitating a top-down approach where even CEOs must approve the adoption of such technologies.

The speaker mentions a company called MCH, which is developing an AI agent for quality assurance (QA) testing. Unlike previous companies like Rainforest QA, which struggled to balance efficiency with job security for QA teams, MCH's AI can fully replace QA personnel. This allows them to market directly to engineering teams without needing to appease QA staff, thus simplifying the sales process.

The speaker reflects on their experience with a recruitment software company, Triple Vet, which faced similar challenges in gaining acceptance from both engineering and recruiting teams. The software aimed to streamline the hiring process for software engineers but encountered resistance from recruiters fearing job loss. However, advancements in AI now allow for the development of comprehensive recruitment solutions that can operate without traditional recruiting teams.

The discussion continues with the mention of a company called Prior, which is successfully implementing AI to handle the entire recruitment process, including technical screenings. The speaker notes that as AI technology evolves, companies may no longer need to build traditional recruiting teams, thus eliminating the friction previously experienced in the hiring process.

The speaker concludes by referencing a company called Cap.AI, which has developed a highly effective chat tool aimed at automating developer support. This reflects a broader trend where AI is increasingly capable of handling tasks traditionally performed by human teams, further illustrating the shift in how businesses operate.

The discussion highlights the increasing use of AI bots in customer support, which effectively handle technical queries by ingesting developer documentation, YouTube videos, and chat histories. This advancement has led to smaller DevRel teams, as these AI systems provide high-quality answers. The speaker reflects on their experience with Power Help, an AI customer support agent company, noting that while the market is crowded with over 100 competitors, most utilize simple zero-shot LLM prompting that fails to replace complex human workflows. Only a few companies are developing sophisticated software capable of managing intricate customer support tasks, collectively holding less than 1% market penetration, indicating a significant opportunity for growth in this sector.

The conversation shifts to the concept of hyper specialization in customer support software, suggesting that while a general-purpose AI customer support agent may emerge in the future, the current landscape is characterized by highly tailored solutions for specific industries. For instance, Gig ML is mentioned as a company handling 30,000 tickets daily for Zepto, effectively replacing a thousand-person team. This specificity is crucial, as different verticals, such as delivery marketplaces, require distinct support systems. The speaker emphasizes that the demand for customized solutions has led to the emergence of numerous SaaS companies, collectively valued at $300 billion, rather than a single trillion-dollar entity.

The discussion touches on the limitations of firm growth, referencing Coase's theory of the firm, which posits that companies will only expand to a point where further growth becomes inefficient. This principle leads firms to specialize in their strengths, creating networks and ecosystems within the economy. The speaker reflects on insights from Parker Conrad at Rippling, who emphasizes the importance of understanding the capabilities of employees and the potential of AI in HR software. Conrad's perspective highlights that while AI can perform various tasks, the true value lies in its ability to process and understand information, which is essential for managing a large workforce effectively.

The discussion highlights the increasing power of tools for managers and CEOs, particularly through platforms like Rippling, which aims to create a comprehensive suite of HR tools. This could potentially allow managers to scale their operations significantly, possibly leading to the consolidation of multiple billion-dollar SaaS companies into one. The speaker emphasizes the transformative potential of these tools in extending managerial capabilities.

The conversation shifts to the impact of AI on the ability of leaders to manage relationships within organizations. The speaker references Dunbar's number, which suggests a limit of around 150 meaningful relationships for individuals. However, with AI tools, this limit may be extended, allowing leaders to engage with a larger number of employees effectively. An example is given of a CEO who created a voice chat tool that could connect with all 1,500 employees, illustrating how technology can facilitate broader communication.

The discussion introduces a specific AI tool that can call employees and extract meaningful insights from their conversations, providing CEOs with concise summaries of important information. This tool is noted to be significantly more effective than previous methods of gathering employee feedback. The speaker raises questions about the future capabilities of large language models (LLMs) and whether they might evolve to perform actual thinking, thereby altering the dynamics of organizational leadership.

The conversation reveals that Rippling employs over a hundred founders who manage various SaaS verticals within the company. This unique structure is designed to foster innovation and specialization, with a focus on recruiting talented individuals who can build on Rippling's platform. The speaker compares Rippling's approach to that of Amazon, emphasizing the importance of shared infrastructure and the potential for significant revenue generation from new product launches, such as time tracking tools.

The discussion highlights the importance of having a strong market presence, as exemplified by leading software companies like Oracle, Microsoft, and Salesforce. The speaker emphasizes the need to balance customer acquisition cost (CAC) with lifetime value (LTV) to ensure sustainable growth, particularly for startups aiming to establish themselves in competitive markets.

The speaker introduces Salient, a company utilizing AI for voice calling to automate debt collection in the auto lending sector. This innovation addresses the high turnover and low job satisfaction in traditional call centers, where workers often handle tedious tasks. Salient's AI technology has shown promising results, achieving high accuracy and gaining traction with major banks, marking a significant advancement in the automation of customer interactions.

The conversation shifts to the rapid growth of voice infrastructure companies, such as Vapy, which enable businesses to quickly implement voice solutions. The speaker notes the excitement surrounding these technologies, which can be operational within hours. However, challenges remain regarding customer retention, especially with the emergence of new voice APIs from OpenAI, raising questions about the long-term viability of these platforms.

Reflecting on the evolution of AI voice applications, the speaker observes that just six months prior, the technology was not sufficiently advanced, with issues like high latency and unrealistic voice quality. The discussion recalls the first Y Combinator batch in winter 2023, where initial LLM-powered applications primarily focused on generating text, lacking the sophistication seen in today's AI voice solutions.

The speaker reminisces about early LLM applications, such as Speedy Brand, which simplified content generation for small businesses. These applications, while innovative at the time, were basic and faced challenges from more advanced models like ChatGPT. The speaker notes that many of these early apps struggled to compete with the rapid advancements in AI technology, leading to a 'boiling frog' effect where the gradual improvements in AI capabilities went largely unnoticed until they became significantly superior.

The discussion highlights a remarkable progression in AI technology over the past two years, with advancements leading to the development of full-on vertical AI agents capable of replacing entire teams and functions within enterprises. This rapid evolution is unprecedented, and the speaker expresses amazement at the pace of progress, noting that the landscape is shifting from a single dominant player, OpenAI, to a more competitive environment with contenders like Claude emerging.

For aspiring entrepreneurs, the speaker emphasizes the importance of identifying 'boring, repetitive admin work' as a potential vertical for AI startups. This common thread suggests that there is significant opportunity in automating mundane tasks, which could lead to billion-dollar AI agent startups. The speaker advises that founders should pursue areas where they have personal experience or connections, as this can enhance the likelihood of success.

The speaker shares specific examples of successful AI applications that emerged from identifying repetitive tasks. One example is a startup developing an AI agent to bid on government contracts, inspired by a friend's tedious job of refreshing a government website for new proposals. Another example involves a company creating an AI agent for processing medical billing in dental clinics, which originated from a founder's experience working with his dentist mother, who highlighted the tedious nature of claim processing. These instances illustrate how recognizing mundane tasks can lead to innovative AI solutions.

The speaker draws a parallel between robotics and vertical SaaS, suggesting that profitable AI solutions will likely focus on 'dirty and dangerous jobs' or, in the case of vertical SaaS, on 'boring butter-passing jobs.' This insight reinforces the idea that the most viable opportunities for AI applications lie in automating tasks that are often overlooked due to their mundane nature.