• The Angle
  • Posts
  • The Eight Paradigms of AI Application Startups

The Eight Paradigms of AI Application Startups

The Angle Issue #228

The eight paradigms of AI application startups
Gil Dibner

Let me start off by saying that we strive to keep an open mind. We continue to actively pursue and invest in companies that do not say the word “AI,” and I can think of several examples of new investments we’ve made this year that really (honestly) do not involve AI. That said, we are not ignoring the tidal wave of innovation that newly powerful AI tooling has unleashed. Our deal flow is awash with AI startups across both infrastructure (such as FalkorDB and Tensorleap) and applications. We’ll talk about infrastructure for AI elsewhere, but today, I wanted to focus on application-level AI companies as we are currently seeing them.

A taxonomy of application AI. One of the unique things about venture capital is the opportunity to see so many talented founders and interesting companies over such a short span of time. Over the past few months, I’ve been struck by how varied the role of AI can be in different companies. I wanted to try to propose a set of paradigms for how AI application companies actually leverage AI. This is, if you will, a sort of preliminary taxonomy of AI application companies.

The eight flavors of AI application startups. As I see it, the AI application companies being started today typically fall into one of eight categories depending on how they leverage AI to create value for customers. These can be grouped into three broad categories. Here goes:

Group 1: AI as a value delivery model. These companies leverage AI directly to deliver novel value to customers.

  • Agentic subsystems. An agentic subsystem uses AI to perform a relatively straightforward (if difficult) service that is consumed as an input to some other service or operation. Often these are consumed behind an API and used by developers to create a bigger offering. These services are “agentic” in that they can sometimes function as autonomous agents (within preset limits) until they achieve their objective. One simplistic example of these would be anomaly detection in the context of industrial quality control. An other would include automated translation of video generation services.

  • Functional copilot. A copilot acts as a companion to a human user to help him or her perform tasks more efficiently. The best example of this is perhaps Github copilot, but we are seeing a very wide range of such systems for various functions (customer support, sales, product management, etc.). Recently, we looked at a company that’s building a copilot for CNC machining.

  • Functional autopilot. An autopilot acts as a replacement for a human user with the intention of completely replacing that function (or, at least, some subset of the people performing it). For example, we’ve seen companies that aim to completely replace the content marketing function for online sellers or various aspects of IT cost management. Theoretically, a company could not have any humans providing the autopilots function.

  • Service/work delivery. In the service (or work) delivery model, the AI enables a software company to completely replace an entire organization in the delivery of a service to a customer. This is basically “service-as-a-software” and was very well described by Sarah Tavel. This approach would include, for example, the many companies offering legal services in the form of robotic lawyers.

Group 2: AI as a penetration vector. We’ve also seen some interesting companies that are using AI as an input more than as an output. They are using the power of AI as a penetration vector into a market. In these paradigms, companies are leveraging AI to enable them to compete in difficult markets and displace incumbents or stubborn internal processes at customers.

  • Data ingest and organization. We’ve seen a handful of interesting companies that are using the power of AI to organize unstructured data as a wedge to penetrate a market. The playbook is typically to incentivize customers to contribute their data, which the startup can then structure and enhance. That data asset can then be leveraged to find ways to upsell accounts and/or lock them into valuable workflows.

  • Data fusion / complex query handling. One of the superpowers of LLMs is their ability to understand data structures and fuse disparate data sources together. We’ve seen a number of startups that are using these capabilities to allow very complex queries that drive unique value for customers. These queries often fuse structured and unstructured datasets and contain complex mechanisms to prevent hallucinations and traceability. On the surface, these applications can look like traditional SaaS but they are applications that would not be available without this sort of advanced AI-powered data infra.

Group 3: AI as software displacement. We’ve recently come across two additional types of AI application companies that constitute a third group: software displacement. These are companies that create value principally by using AI to reduce the need for software development.

  • Custom code displacement. One area where AI excels is in handling novel or complex situations. The many companies that are selling AI-powered robotics are essentially selling this value proposition. The value is not in the robots at all, but in the software which leverages AI to enable it to interact with any range of objects (for example) without any custom coding. The value for this type of company is the reduction of the cost of customization to zero. There are markets (and robotics is certainly one of them) where this can be highly disruptive.

  • Application displacement. The final and most intriguing category of AI application we’ve come across recently is that of application displacement. These companies are trying to completely displace the need for the creation of software itself. They argue - convincingly - that AI is becoming so powerful that it can write software on the fly. Taken one step further - if software can be written automatically - maybe we don’t need to write it at all?

These are the eight paradigms of AI application software I think I’ve encountered recently. But I am sure there are more. I would love to know what you think of the categories listed above - and especially if you think there might be some I’m missing. Like everyone else, I’m trying to figure out the new world of software, and would love some help.

In the meantime, if you are building in one of these categories or in one we didn’t list, please do reach out…


Revenue Durability in the LLM World
Everything about LLMs seems to make revenue durability more challenging than ever.

A Digital Fabric for Maritime Trade
Why we invested in Portchain.

Three Keys to the Kingdom
The sometimes-competing and sometimes-aligned goals that early-stage founders must manage.



AI = Apple Intelligence. At WWDC, Apple unveiled its vision for AI - “Apple Intelligence” - as a key component of iOS 18. Siri is getting a refresh, documented by AppleInsider here, as is Apple’s calculator app, all powered by ChatGPT. Ben Thompson from Stratechery has a typically lucid take on the announcement today, highlighting the fact that AI is a complement to Apple’s core business (rather than disruptive, like it is to Google’s core business), which gives Apple wider latitude in how to incorporate AI into its core product offering. This is especially true given that the capabilities of AI are still evolving. As Thompson argues (emphasis mine): “Once again, we see how Apple (along with Google/Android and Microsoft/Windows) is located at the point of maximum leverage in terms of incorporating AI into consumer-facing applications: figuring out what AI applications should be run where and when is going to be a very difficult problem as long as AI performance is not “good enough”, which is likely to be the case for the foreseeable future; that means that the entity that can integrate on-device and cloud processing is going to be the best positioned to provide a platform for future applications, which is to say that the current operating system providers are the best-placed to be the platforms of the future, not just today.”

Don’t sleep on Waymo. Waymo may only be live in a few cities (notably SF), but we should all pay attention to the extremely positive reviews the service is getting. When you witness the future, pay attention. Aditya Agarwal, partner at South Park Commons, shared a great rundown of the experience on X earlier this week.

The future is solar. From this article in CleanTechnica: “A record-setting 11 gigawatts (GW) of new solar module manufacturing capacity came online in the United States during Q1 2024, the largest quarter of solar manufacturing growth in American history.” That’s big. Most of it is utility-scale solar in Florida and Texas. Total U.S. solar capacity is expected to double over the next five years, growing to 438 GW by 2029.

Falcon Heavy splashes down. Fourth time’s the charm. On its fourth test flight, the Falcon Heavy from SpaceX saw a soft splash down, a massive accomplishment and another huge step, further cementing SpaceX’s dominance in launch capabilities.


Weird vibes. Benn Stancil is right, the vibes are weird. Stancil does a great job capturing the strangeness of the current moment in the world of startups. Entrepreneurs are simultaneously raising massive rounds for AI-powered applications and LLM infrastructure, and also wondering if the startup game will even continue past the end of the decade, or even the year (will AGI just come and alter the playing field entirely? What other deus ex machina is hiding around the corner). I don’t agree with it all, but Stancil is putting voice to an unease that many in the industry are feeling right now.

Zero to Hero. Andrej Karpathy has posted another lecture on YouTube, this one is a 4-hour walkthrough of how to recreate a GPT-2 model. This is part of his “Zero to Hero” lecture series, meant to bring anyone up to speed in AI. It’s an impressive feat and an incredible gift to the startup world. Worth watching if you’re also getting up to speed on the space.


The AI Agenda. Jamin Ball from Altimeter does it again, with another super useful post in his Clouded Judgement Substack. He highlights that one big source of friction in procurement cycles right now is an unclear answer to the critical questions: “does this vendor push forward our AI capabilities or not?” If the answer is yes, great! But if the answer is no, not only does the software company risk losing the deal, the buyer/champion risks losing their job. Everyone is on the hook, vendors and champions, to tell their AI story right now. So if you’re a vendor, make sure you can help your champions become AI heroes…and keep their jobs!

Rule of 15. Michael Kim, founder and partner at Cendana Capital, joined Samir Kaji this past week on his podcast Venture Unlocked. The whole interview is worth a listen. Kim is an investor in some of the most well-known, and most performant, emerging managers of the past decade. His main piece of advice to managers - get your ownership early. He shared a simple rule: you need an outcome that’s 15x your fund size to 3x your fund. What does that mean? Keep your ownership high, and your fund size small. More here.


Tensorleap will be hosting a meetup on “Exploring AI Frontiers: Explainability in Multimodality and Unlabeled Data” at MIT Innovation Headquarters in Cambridge. Register here to join.



or to participate.