July 2023
From: Brian and Tobias
Subject: Horizontal vs. vertical use cases for AI
Welcome to July’s Infrastructure Newsletter from Brian & Tobias. Over the past several months, we’ve written about various ways AI can change software as we know it. We’re not alone. Listening to VCs on the echo chamber of Twitter, it seems like we’re already living in a world where white collar work has been substituted for AI and autonomy. The investor hype suggests this future is inevitable and imminent, but in reality the fate of GenAI is quite distant and unknown.
When we talk to enterprise leaders, we hear them asking for true needle-moving, value-creating use cases. Initial excitement is waning. The search for value is surging. Use cases like code generation and automatic marketing copywriting were low hanging fruit, obvious applications of the technology, but nothing incremental with the same “wow” factor has emerged on top of the first wave of apps.
Mild disappointment has become a trend. We’ve heard from numerous engineers that they have a hard time getting AutoGPT to work, despite the hype of autonomous agents. Anecdotally, we’ve heard that ChatGPT usage is fading and there is some evidence to back this up. Web traffic was down 10% and unique visitors were down almost 6% from May to June. Some stranger things are happening as well, including a study that GPT-4 is actually getting worse.
We’ve also noticed a shift in the tenor of businesses around the promise of AI – they are still bullish but want to see where new value will come from. We spoke to several C-Suite executives to prepare for this newsletter. One of them advises a large software PE firm and its portfolio. According to him, scaled software companies are pulling back from initial excitement and looking for the real business value in Generative AI applications.
In our opinions, the issue at the crux of this lapse in excitement is enterprise inability to apply Generative AI to company-specific use cases. We’ve gotten good at horizontal use cases, but the real value will come from vertical ones, a much harder nut to crack.
Initial excitement in horizontal:
The quickest opportunities to get addressed by AI have been horizontal - tasks that apply to many fields like summarizing documents and automatically generating images and text. For consumers, seeing horizontal applications has been mind blowing: write a best man speech in the style of Shakespeare, summarize a long document in just a few bullet points… the list goes on. With these horizontal capabilities came a few revolutionary products – Jasper and CoPilot come to mind. They took a general capability and created a fantastic UI for a specific user. To everyone who writes natural language or code (lots of people), this felt like magic.
In fact, the nature of these solutions as horizontal has been part of what has led to the frenzy – they definitionaly apply to everyone, which has enabled this technological moment to feel inclusive and relevant to many. People saw the power of AI and assumed that capabilities would continue rolling out and automation would begin touching more of their daily lives. This hasn’t happened, and the core reason is that we’re in the chasm between the broad and the specific – broad based apps are abundant, while specific ones are only emerging. After all the excitement about the cool stuff that emerged, people are finally looking around and asking, “which of this will I actually use?”
Why vertical is necessary but hard:
The apps to use may not exist yet, and our opinion is the more specialized and verticalized they get, the more helpful they will be. There are early examples of this. Harvey and Casetext focus specifically on the legal field. We think they’re well-positioned because of their focus on a specific domain. Notably, Casetext was recently acquired by Thomson Reuters for $650M.
Additionally, OpenAI is turning its strategic attention towards personalization, with features like custom instructions and fine tuning, which people at OpenAI have said is coming soon. In the words of someone at OpenAI that we recently spoke to, “OpenAI knows it needs to go up the stack in order to win this race, and fine tuning enablement is a big part of that.”
Vertical solutions give companies the ability to build AI products in ways that are uniquely specific to them and their problems. Because they solve specific problems, they are more powerful and value-driving, but also harder to create. They are by definition idiosyncratic company-to-company, problem-to-problem. There are two core reasons why they are so much more difficult to build:
We are faced with an ironic challenge: AI verticalization requires data owned by specific companies, but those companies often don’t have the infrastructure or data security procedures in place to share the data with outsiders who can use it or leverage it internally. They also don’t always know the right way to package the product in a usable, intuitive way. And finally, large enterprises worry about the efficacy of AI products they build – accuracy, efficiency (speed and cost), data infrastructure disruption, and explainability are all big bottlenecks that make companies more resistant to putting something into production. So, while everyone wants AI to help their businesses in theory, the tactics of bringing this into reality are messy.
What will these vertical solutions look like?
What does this end up meaning? We think there will be an evolution of verticalized apps. In the same way that horizontal apps preceded vertical ones because they were easier and less risky to build, so too will there be an onramp for products to get progressively specialized.
The first vertical applications will be efficiency plays – places where document processing is a bottleneck solved today through humans that could and should be solved with AI. This was the fundamental insight from Harvey and Castext. In the words of Casetext CEO Jake Heller, GPT-4 and other similar LLMs enable us to build “graduate level humans that operate at the speed of a machine.” From company to company, where to “employ” this graduate-level human will vary, but there will be lots of opportunities to staff AI products on document-heavy, labor-intensive, and expensive projects. Building the right infrastructure to help companies easily build these sorts of products will be a big value unlock.
These products will be internal, and external-facing products will follow once there is more trust in the effectiveness and explainability of LLMs. It’s okay if Jasper makes a mistake once in a while; it’s a major problem if a financial advisor, doctor, insurance agent, etc. gives a customer the wrong information or advice. As a result, these external-facing products will first require a human-in-the-loop, and over time the need for that assistance will disappear.
In order, we think LLMs will create value in the enterprise only with (a) better infrastructure to build verticalized workflow and document processing tools and (b) explainability and monitoring solutions to make companies feel comfortable with productionalizing external-facing products. From an investment perspective, we’re interested in companies that play in both of these areas.
The bottom line is that we still have a long way to go to make AI truly valuable in enterprise contexts. Horizontal solutions have generated buzz and hype, but vertical solutions will create value. These are harder to make and require more protected data, meaning companies will need to play a big role in building AI products for themselves. Lots of opportunities exist to help companies do this. We are currently incubating two vertical AI businesses in markets that we think are massive and have acute pains where LLMs can help.
We’d like to extend a particular thank you to John Donovan, former CEO of AT&T Communications, for inspiration on the content of the newsletter. John receives these newsletters just like all of you and was tremendously generous sharing thoughts that contributed to this edition.
We will be skipping our August newsletter – we’ll see you back in September and hope many of you will be at our Summit as well. If you have any questions about Summit, please let us know. Have a great rest of the summer!
Until next time,
Brian and Tobias