Verticalized AI — The Drivers Behind the Next Wave of Vertical SaaS
In our last piece, Forum Ventures Senior Investment Associate, Naomi Goez, and I discussed the frameworks through which the wave of generative AI opportunities may be evaluated, establishing greater conviction in AI powered verticalization as a moat for defensibility, differentiation, and ultimately, scale.
At a high level, there is massive potential for vertical AI to be the driver for the next generation of great enterprise SaaS businesses, but what is driving this opportunity, and how can investors and founders best create assessment frameworks? To better understand what truly drives this advantage, we take a deeper look into three areas where thoughtful strategy can establish superiority over legacy and horizontal solutions: Data, UI, and a deep understanding of nuanced workflows in complex industries.
Data
First generation LLMs are plagued by limitations that impact their ability to meaningfully penetrate enterprise customers. This is because they are prone to hallucinations, are slow and expensive, and lack attribution, alongside other core problems related to third party data privacy. These vulnerabilities uniquely position vertical AI companies to leverage their data strategy to deliver better model efficacy. Horizontal LLMs or even LLMs built by legacy vertical SaaS companies are flush with data, but in many cases this can be more of a hindrance than an advantage: The more noise in a dataset, meaning, a large delta between relevant data and the downstream model’s task, the greater the need for additional data, which comes at a cost of speed and increased compute expense. Following the same logic, the cleaner the dataset, the less data needed, resulting in cheaper LLM training. While cost-effective commercialization of quantum compute power will likely transform the accessibility of AI in the long term, in the near term, companies that collect a hyper niche dataset and are disciplined in selecting the data to train their models on, can create extremely effective, small models with outlier performance when compared to other larger models. Furthermore, those businesses that are able to establish dramatic network efforts to bolster reinforcement learning or build unique data sets faster are positioned to win.
Our portfolio founder, Florian Fischetti, CEO of Murabi, a vertical AI tool for financial services, shared that there are three main ways in which founders can access high quality data: 1. Working with customers to utilize their internal operational data 2. Via APIs to enrich operational data with external data. 3. Offering products that generate additional (net new) data by digitizing a formerly analog process. Once the right schema is there, Abstract co-founders, Patrick Utz and Mohammed Hayat, describe a “three part pyramid” framework: 1. crawlers on the base, 2. processing and sanitizing the raw data, and 3. deriving insights from the now structured data. Getting to “better” data is a result of investing in all three parts of that pipeline, where data engineering comes first and data science second.
User Interface
With the advent of generative AI, the entire way in which humans interact with software has opened up an opportunity for innovative UI design. The ability for humans to converse with a computer via NLP makes it possible for startups to leverage chat, video, and even motion to displace incumbent software still saddled by legacy UI workflows. A great example of this is clear within the insurance sector, where conversation takes precedence over convention. John Cottongim, Co-Founder and CTO of Forum portfolio company, Roots Automation, shares about the plethora of opportunities created by conversational interfaces powered by NLP where users can engage with insurance platforms, documents and data, departing from the confines of traditional UI workflows and allowing for a more intuitive, user-friendly experience: “Imagine — a broker interrogating hundreds of unstandardized applications and submission documents by asking their Digital Coworker to extract key information and present back a summary of relevant data needed to make a decision, or an agent helping a policyholder effortlessly navigate their coverage details through a simple conversation with their Digital Coworker.”
This new UI layer does not come without added challenges however. As with any new transformational technology, there is a layer of trepidation that needs to be overcome before users build trust with AI features. Fischetti further shares that the right UI can be thought of as a measure of de-risking the possibility of a system making errors, predominantly through a co-pilot. This is because one of the key things the UI can do is surfacing actionable intelligence to operators when and where they need it. Echod by a stealth AI cybersecurity startup founder in our network, combining improved search functionality enabled by NLP with personalization will enhance the accuracy of output and make solutions stickier and more intuitive.
As Abstract’s Utz and Hayat highlighted in our conversation, many of these vertical SaaS workflows are replacing the reliance on 5+ archaic websites, a hodgepodge of sticky notes, word docs, and spreadsheets, and thus, it is crucial to think through how to utilize cutting edge UI advancements to deliver an innovative project management workflow.
Deep Understanding of Workflows
As discussed at length by many other early stage investors, finding true founder-market and founder-product fit lays the essential groundwork for success. However, the journey of building a differentiated and scalable vertical AI product demands a more granular exploration — one that delves into the intricacies of industry-specific workflows. This goes beyond aligning a solution with existing market needs and is concerned with recognizing the latent gaps that an industry-tailored solution should bridge. We tend to bucket those gaps into two areas: first, bridging external and internal gaps, which can only be done by studying the various parties within a value chain and providing them with purpose-built tools for seamless interactions and transactions, and in itself can create a flywheel of adoption across said value chain. Second, bridging internal workflows and dismantling silos that impede operational harmony.
The former was discussed in a recent conversation with Dan Moshkovich, VP of Marketing at Chargeflow, which raised $14M across two Seed rounds: adoption of vertical software tailored to one transacting party will propel the other transacting party to head in the same direction, each seeking uniquely designed solutions to address their respective pain points. An interesting way to think about this would be observing a value chain holistically and identifying the stakeholders whose needs are already met, to reveal the parties who may be compelled to use new tools in order to keep an even playing field. In the case of Chargeflow, an automated chargeback management solution for ecommerce merchants, there is a clear moat established by the fact that banks, the “other transacting party”, have their own toolkit to address their side of the same problem, and the deeper the market penetration is on their end, the greater the catalyst is for merchants to seek better solutions.
The latter was dissected in our conversation with Roots Automation’s Cottongim, where the underlying premise is recognizing that different departments within the same organization often have disparate data structures, applications and processes, which make it difficult for a universal AI system to integrate all aspects without custom engineering. This can only be done via sector expertise, which then creates two types of opportunities for vertical AI solutions: 1. Becoming a complete system (i.e., Roots Automation’s Digital Coworker), and 2. Building a purpose-built modular component (i.e., their InsurGPT component) that plugs directly into key business processes and user workflows for faster and lower-cost deployment.
A final consideration here touches on the competitive landscape, particularly as it pertains to understanding not only who is building what, but also why they are positioned to bridge operational gaps. What we are seeing now is uneven concentration within specific sectors and job functions, that could get in the way of even the most intuitive UI or great dataset. Notably, we are already detecting friction across healthcare focused notetakers, legal co-pilots, and real estate sales assistants.
On the healthcare front, a current example would be NextGen Healthcare, Amazon, Microsoft, and Oracle, all launching AI based clinical documentation tools. While conversations around the technology’s accuracy, accountability, and equity have sparked in the face of such prominent players dominating the field, this landscape leaves very little space, if any, for a young startup to throw their hat in the ring. On the legal assistants and real estate sales front we are seeing a slightly different scenario at play: the utilization of two of the earliest LLM use cases — customer service and data querying — catalyzed the creation of dozens of new startups all solving the same problems. This was initially a bigger issue across horizontal application layer solutions, where generalist marketing and business intelligence tools became abundant and therefore, vulnerable, and we are now seeing a similar pattern unfold in certain verticals. We reached saturation so quickly that Instagram users have already been getting flooded with “ChatGPT for Law” ads. What can be learned from this almost immediate market saturation is that evaluating barriers to entry and the potential advantages of incumbents in specific sectors, is just as important as building a superior product, particularly in such a dynamic climate.
We remain excited and hopeful about the intersection of AI and verticalization, and will continue to deepen our understanding of how an investor can bring value beyond capital to founders in the space.
If you are building a vertical AI solution — pitch us here.
To be included in our market map, please fill out this form.
A special thank you to our contributors: Florian Fischetti, Patrick Utz, Mohammed Hayat, John Cottongim, Dan Moshkovich.
