Horizontal vs. vertical AI

What kind of AI product should you build?

When you build software, you can build for a specific vertical or target a use case across verticals. I’ve been thinking about which of these is a better strategy for AI products. It’s a widely debated topic and something I’ve been thinking a lot about. Today’s essay is my working hypothesis of how I’m thinking about it. It’s a working hypothesis because AI is moving fast, and we don’t know what we don’t know.

But hey, I’ve always believed having an opinion is better than no opinion at all. Let’s dive in.

Horizontal vs. vertical

A vertical software product caters to the requirements of an industry. For example:

  • Toast: offers point of sale software for the restaurant industry

  • Glofox: helps fitness clubs and centres manage their customers

On the other hand, a horizontal software product offers targets a use case across many industries. These products might include “modules” or “templates” for industry verticals, but they will not offer the level of specificity that industry-specific software does. For example:

  • Salesforce: builds CRM for a variety of industries.

  • Workday: offers general purpose human resource management software.

  • Slack: allows people to message each other at work and is agnostic to the type of industry the business operates in.

One caveat that is important to point out: irrespective of whether you are building a vertical SaaS product or a horizontal products, it is critical to target a narrow segment in the early days of a startup. To that end, it’s better to think of vertical vs. horizontal as a long-term strategy: once you’ve achieved product market fit, are you going to market and sell to a single vertical?

The dynamics of vertical SaaS

The dynamics of a vertical SaaS product is down to 3 things:

Your customer acquisition costs will be lower. If you are targeting a specific vertical, you can be ultra focussed in terms of ad spend, outreach and events. This should lead to more efficient marketing spend, and lower customer acquisition costs.

A straightforward roadmap. Let’s say you are building for an industry like e-commerce. Your roadmap is a lot more straightforward. For example, a Shopify integration is an obvious build if you’re focussing on e-commerce.

The downside of vertical SaaS is that your market size is limited. There may be room only for a limited number of players. You need to ensure that your market is big enough and that you can execute well. If you’re building the CRM tool for property managers, your product must be better than every general purpose CRM tool in the market.

Vertical or horizontal AI

If you’re building in the application layer for AI, I believe you should operate on the assumption that a vertical SaaS product is the way to go.

I define the “application layer” as software that enables end users to take actions with little to no technical set up time. Typically, your users are non-technical users within another business. It does not include foundational models like OpenAI or Anthropic.

My reasoning stems from the few things in AI that I know are certainties:

  1. The supply of software will increase as it becomes easier to build and ship products.

  2. It’s easy to go from 0 → 75% accuracy with an AI product, it’s really hard to go from 75% → 99% accuracy.

  3. Distribution always wins in the long-term.

When I think about (1) and (2), it’s easier to deliver a 10x improvement over the status quo today by focussing on a vertical. If you’re building a product to write better marketing copy, you are going to have go through a tonne of trial and error to deliver true value. This experimentation involves things like the right prompt, enforcing a human in the loop workflow or using contextual data to improve output. The value from this experimentation accrues to a specific use case and vertical.

This means that you will build far more momentum and traction by focussing on a vertical. Once you spend 18 - 24 months building a specific vertical (you should do this irrespective of the vertical vs. horizontal debate), you need to think about growing revenue. There are 3 sources of revenue you can pull from:

  • Increase revenue from your existing user base

  • Acquire new customers from the vertical you’ve focussed on

  • Expand to other verticals

I believe the first two are going to give you a much better return on investment. Acquiring customers from a new vertical means figuring out marketing and product. Besides, it’s almost certainly the case that someone else has built a solution for the vertical you plan to enter. How will you compete? Building the tech might be trivial, iterating through different versions of AI and winning distribution is not.

Horizontal AI products

I do believe there will be lots of very successful horizontal AI products. However, these look more like “middleware” than the application layer. These products will have the following characteristics:

They have a high degree of customisation. They are developer friendly and allow teams to build products with ease for their specific use case. It’s like the ERP system of the future by leveraging the power of AI. For example, Fixie.AI allows enterprise customers to spin up their own AI agents. (I interviewed Matt, their CEO, here).

In some cases, they might solve an unmet need. For example, when you build in AI you need to store data in a different structure. Vector databases like Pinecone are solving this unmet need. It makes sense to be horizontal because the secret sauce for these products (retrieving the most relevant entries with low latency) exists across the board.

These “middleware” products enable other AI-first experiences. For example, I use Pinecone for my product which is an “application layer” product. It’s an important distinction because the marketing and go-to-market strategy looks very different for these horizontal, middleware products. The application layer feels closer to “traditional” SaaS. These horizontal, middleware products look a lot more like cloud services did in the last decade.

To close

For application focussed AI products, I believe the working hypothesis should be to focus on a specific vertical. This actually gives you more optionality than you imagine. In the early days of a product, you should focus on a very tight customer segment anyway. I’m proposing that you build within a vertical for as long as you can, and hopefully, forever. You can always choose to move to a horizontal product at a later stage, but be very mindful when you do.

On the other hand, if you’re building a product that caters to the needs of other folks leveraging AI, you are better off building horizontally for a wide variety of customer segments.