The 7 Types of Generative AI Products

In this exercise, I hope to outline thew new game of AI, how to play it, and ideally - how to win.

There are only 7 types of generative AI products.

Hundreds of new AI products launch each day, in every vertical imaginable. The hype around how generative AI will transform work and life has never been higher.

And yet there are only 7 types of generative AI products.

Breaking down the startup landscape into these seven categories helps us make sense of each new startup launch. It gives us a framework with which to analyze products, the value they create, and the strategies that they should take.

By attributing a product to one of these categories, we can immediately intuit its value proposition, product challenges, business hurdles, and differentiation methods. Because within a single category, they’re all the same.

So in this exercise I outline the 7 types of AI companies and in the process, hope to outline this new game of AI, how to play it, and ideally - how to win.

---

While hundreds of new AI startups pop up every day, they all fall within 7 distinct categories.

They are Chat, Copilots, Semantic Search, Agents, Tooling, Foundation Models, and Infrastructure. The table below outlines the 7 categories and their important properties.

Three Common Problems

The competition amongst generative AI products is insane. Products that don’t protect themselves against competition will likely spin in circles, earning customers and then quickly losing them. Any product category that seems profitable is quickly inundated with countless competitors, so how does one create something of durable value?

These are some of the the major problems AI products face when it comes to winning in the market.

The Commoditization Problem

Commoditization occurs when a market is perfectly competitive and there is little differentiation between products. This drives profit margins to zero.

Most LLMs follow a similar, if not exactly the same, API interface. The model itself is a black box that takes in a prompt and outputs language. Because the interface is so simple, it’s easy to swap models. Or otherwise put, there is very little switching costs once a model is integrated into a product.

This means foundation models have to continually compete with each other to be the best. As soon as one claims to be better than the other, customers can easily swap their usage resulting in heavy churn for incumbents.

The only opaqueness for customers in this market results from two properties

1. Prompts often differ in their effectiveness between models, so switching models often result in lost time testing new prompt variations.

2. Measuring language output quality is still subjective. This means changing a model won’t be perfectly backwards compatible to existing users.

As more businesses experiment with models, we can expect there to be more information on how to optimize prompts for different models. There are already tools that claim to help with this, and I expect them to get better over time.

As for the second point, there are many models that have chosen to focus on a specific task and become the best model for that job. However, even when focusing on a niche, the market dynamics remain the same. Open source models have reduced the barrier to entry when it comes to training domain specific models, so any profitable domain specific task will attract lots of competition. Because of this competition, many companies building domain specific models may be well served to verticalize by building applications on their model and selling direct to end users.

The best form of differentiation in the market for models will end up being private data. OpenAI and Google have already begun licensing private content. This gives owners of good private data sets immense bargaining power over model makers. Many have even decided to build great AI-complimentary businesses by letting models and applications consume their data via API.

The Enterprise Security Problem

For companies without external private data sets, using data generated from customers to create better and better models seems like an attractive business idea.

This would give a big advantage to early entrants in the market who can leverage existing usage to improve their products, creating a network effect style moat over new players in the field.

One problem with this however is that enterprises aren’t allowing AI companies to user their data for training base models. In fact, reports of models “leaking” private information has created apprehension for large enterprises to external AI solutions entirely.

Because of this, enterprise-facing AI companies will have to rely on traditional SaaS lock-in to create defensibility around their business. This requires at least one of two things:

1. The AI solution must require an upfront investment of time and money to integrate the solution (the must be inherent in the type of product, i.e. competitors would require the same integration cost)

2. The AI solution must generate business critical data for the customer that can’t be easily exported or accessed elsewhere

Unfortunately, many AI solutions fail to meet these criteria. One of AI’s strengths is analyzing data. For example, many search and chat apps allow customers to supply their own data and use the AI to interact with it. Since the customer owns the main data source, the AI product itself is more expendable. As another example, copilots often embed themselves as “assistant-like” extensions to existing software. This makes them plug and play, since the user doesn’t have to really learn how to use the copilot. It just works ambiently to help edit or create content.

This alone doesn’t create much lock in. Chat and search apps will have to invest in building entire product experiences around the interaction with data that unlocks other value. Building a deeper product experience will require customers to do more training to get up to speed. This adds a cost to transitioning between products.

Copilots similarly should aim to own the entire product experience rather than simply integrating into existing software. The tradeoff with this strategy is that it requires customers to replace their existing software, which is no easy task. Products that integrate into existing software will gain more distribution in the short term, but over time they won’t be able to build as sticky of an audience.

Cursor is a great example of a company that has followed this playbook. Rather than integrating into VSCode like GitHub Copilot, they’ve forked VSCode and built their own flavor of the IDE which is “AI-native” and installed separately.

The Consumer Retention Problem

For those selling to consumers, retention is a hard problem. Consumers are always fickle, and AI seems to bring the worst out of this problem.

It’s a common trope that making AI demos is easy but building AI products is hard. Many new launches grab the attention of lots of consumers, but fail to convert them into recurring users.

Some of the problems that they face are:

1. Products are too open ended and place a high burden on the user to figure out how to extract value (chat, search)

2. AI is prone to mistakes and interpretation issues (agents)

3. AI output quality is subjective and so hard to optimize across a large user base (copilots, search)

AI software is the first mainstream software to be non-deterministic. Given the state of other software, consumers often expect consistent output and fast performance.

If they can’t gain confidence that a tool they’re using will give them a good outcome 95% of the time, they’ll likely churn.

However, if they do work - consumer software benefits highly from word of mouth and organic product led growth. This means that products that can mitigate the above problems (e.g. Perplexity) have a good chance at hitting exponential growth. For this reason, consumer AI companies are in a race to build the most reliable AI solution at scale.

Closing Thoughts

There have been thousands if not hundreds of thousands of AI startups launched in the last year. However, they all fall under 7 real buckets and are organized in the following stack:

Despite the sheer volume, the problems that they face are all fairly similar. The ones that win will either secure private data sources, build high switching costs, or build enough reliability to win over consumers.

---

If you like this contact, feel free to follow me on LinkedIn or Twitter.