AI Horseless Carriages - A multipart series exploring how today's AI apps remain trapped in outdated software paradigms, unnecessarily limiting the true potential of their underlying models, is sponsored by Agent.ai - Discover, connect with and hire AI agents to do useful things.
I recently came across this article from Pete Koomen a General Partner at YC that struck a chord. Here is what he had to say -
When I use AI to build software I feel like I can create almost anything I can imagine very quickly. AI feels like a power tool. It's a lot of fun.
Many AI apps don't feel like that. Their AI features feel tacked-on and useless, even counter-productive. I am beginning to suspect that these apps are the "horseless carriages" of the AI era. They're bad because they mimic old ways of building software that unnecessarily constrain the AI models they're built with.
Before we dive deeper, some background on the term "horseless carriages". People called the first cars "horseless carriages." They weren't seeing cars for what they really were. Instead of recognizing this amazing new invention, they just described it as an old thing (a carriage) minus something (the horse). This limited how people thought about and designed cars at first - they literally built them to look like carriages without horses!
This is not something new - anytime we have a new innovation, the early implementations are "horseless carriages". The "horseless carriage" of early websites was essentially the digital brochure - static pages that simply transferred print media content online without utilizing the unique capabilities of the new medium.
Similarly, when mobile phones came out - Early mobile apps were often just desktop websites squeezed into phone screens or wrapped in app containers without rethinking the experience. Companies created "wrapper apps" that displayed their regular websites in what was essentially a dedicated browser, forcing users to pinch, zoom and endlessly scroll on tiny screens. This approach ignored the fundamental differences between desktop and mobile contexts - ignoring that mobile wasn't just a smaller screen but a completely different way people interact with content. It wasn't until the mobile-first revolution around 2010-2013 that developers began designing specifically for mobile capabilities (touch interfaces, on-the-go usage patterns) rather than trying to force-fit existing web experiences into a smaller form factor.
Paul goes into a deep dive on his struggles with Gmail and it's implementation of Generative AI. You can read all about it here. We can all relate to this. How many of us have used Gmail's Generative AI to write an email and found that the effort to prompt it is more than the effort to write the email.
When Koomen tested the feature by asking it to write an email to his boss, it produced "a perfectly reasonable draft that unfortunately doesn't sound anything like an email that I would actually write."he AI-generated message was overly formal and verbose compared to the brief, casual message he would have written himself.
The Gmail team built a horseless carriage because they set out to add AI to the email client they already had, rather than ask what an email client would look like if it were designed from the ground up with AI.
I think a lot of SaaS companies will go through this exercise and transition and a lot of existing vendors will struggle as they have a lot of "legacy" applications that will over time face competition from AI first native applications that look very very different from our current SaaS applications.
From personal experience, a good example of this is Salesforce Agentforce. When Agentforce first came out, I was extremely excited to try it out. I had been building GTM agents for 6 months and anything I could do with Salesforce would add a lot of value to my clients. So I signed up for their trailhead and spent a few hours completing the trailhead and what I came away from that experience was just how complicated it was. Not because Agent building was complicated, but because Salesforce was complicated. Especially as someone with limited background in Salesforce.
Here are some examples of steps I had to follow that showcase how the complications of Salesforce bleed into anything you want to do with the platform.
This was just the initial instructions and I was already lost.
This got me thinking about other examples of AI horseless carriages in products I use. In my next edition, based on my experience, I will share ideas on how I think some products will evolve over time, as we move from the AI horseless carriages to AI cars.
