Historically, startups had an innovation advantage because incumbents were slow to adopt new technologies.
The battle between every startup and incumbent comes down to whether the startup gets distribution before the incumbent gets innovation.
On the surface, AI seems different from past secular trends like mobile because incumbents are readily adopting it.
But with AI, mere adoption is not enough. It’s not sufficient to slap the AI label on something or add an AI chatbot to a product. Even ‘integrating’ AI with existing products is often subpar.
It’s been ~2 years since the ‘ChatGPT moment’ when AI entered the zeitgeist. How many incumbents actually have a good implementation of AI? ๐ค
This raises two questions:
- How do we define a ‘good’ AI implementation?
- Why do so few incumbents do it well?
This is an exploration of those questions.
What’s a ‘good’ implementation of AI (or any new technology)?
With new technologies, the first phase is always doing the old thing in the new way: “Let’s do everything we did before, but with AI!”
The most common example is layering AI into existing experiences. This can achieve incremental improvements, and seems a necessary step in exploring what’s possible (particularly for incumbents locked into the way things used to be).
Incumbents benefit from awareness and distribution advantages, but have large existing businesses with customers to keep happy and complex dependencies that prevent them from innovating.
On the other hand, thousands of unburdend startups perform a distributed search experimenting at the edges of what’s possible to find what works. Most don’t work out, but some do. Most importantly, they all push us at least a little bit further in some new way.
While incumbents are largely constrained to doing the old with in the new way, real innovation doesn’t happen until the new way (AI) is used to do things that weren’t possible before.
So, how would we define a good implementation of AI? It’s true innovation. The definition of innovation according to Wikipedia:
Innovation is the practical implementation of ideas that result in the introduction of new goods or services or improvement in offering goods or services.
~ Wikipedia definition of Innovation
I love this definition, and I’d add two things for AI innovation:
- Some significance threshold for improvements. For AI, 100X on some dimension seems like a good target (100X fatser, 100X cheaper, etc.).
- The product must improve as AI improves.
Examples in the wild
Despite the chat interface, ChatGPT and Stable Diffusion were early examples of this. But what examples are there from incumbents that are widely available to consumers?
Even examples from leading companies like Microsoft seem mostly to be incremental improvements to their Office products.
Since Microsoft acquired GitHub, Copilot is one of the best early examples of AI implementation from an incumbent, though it’s a new product rather than existing.
Another potentially innovative implementation by an incumbent is Apple Intelligence (not yet released, so TBD).
The secret sauce for successful incumbents
Two things GitHub Copilot and Apple Intelligence have in common also gets to why it’s so hard for incumbents to innovate: They’re leveraging data/capabilities that they uniquely have access to, in order to do things that weren’t possible before.
Not all incumbents are in that position, which might determine the ability to maintain dominance going forward with a built-in moat.
Why do incumbents struggle innovating with AI?
For over a year now the market expects every company to have an AI strategy. For companies without visionary leaders (most companies that aren’t founder led), this means execs and managers are asking their teams to brainstorm ways they could integrate AI into their existing products, with many on the second or third wave by now.
But that’s the problem. It’s really hard to retrofit existing products with AI, especially without a broader company vision for the future. I learned this first-hand building one of the first AI products at my last company in early 2023.
How AI is different
AI is predicated on different assumptions, and many outdated assumptions are necessarily built into pre-AI productsโฆa million invisible decisions that didn’t feel like decisions at the time.
But AI bends the curve of what’s possible, so starting from scratch helps unburden you (which is hard enough) to discover where true innovation is possible. Then you can figure out where to layer in the existing advantages (if you have them).
When you do it right, an AI-native product should feel simple in comparison, yet have much of the same capabilities along with new ones. At first it might be worse in some ways, but the overall product will impove as AI does.
Build…wait…copy?
When considering new features, ‘build, buy, partner’ were the main options in the past. But for AI startups, it’s more like ‘build, wait, copy’ where you can:
- Build it now
- Wait for AI to sufficiently improve
- Copy an innovative approach from another startup
Caveat: With AI, ‘wait’ is a good tell that something isn’t worth building if there’s a good chance AI will be able to do it soon enough. You want to focus on the hard/complex things a general model won’t handle any time soon.
There’s so much innovation happening now that there’s a ton of inspiration from startups across industries solving problems in new ways, which is why it’s so valuable to learn from other startups at the edge where you can.
An example from POPSMASH
The hard reset is also an opportunity for new insights. For example, with POPSMASH we opted to NOT build most design functionality since we assumed AI would be do it soon automatically (or make it easier). These were ‘critical’ features that every incumbent and competitor had.
But we didn’t build them. How many sales would we lose?
Interestingly, only a couple customers even asked for the features, and those who did needed something very specific–just a sliver of the design features competitors had.
This turned into an advantage, since now we’re using AI to intelligently handle the minimal design functionality customers. There’s also a huge benefit to customers: there’s a TON of complexity our customers never have to endure with us that’s tablestakes at competitors.
It turns out when you innovate dramatically (~100X improvement) on some dimension, customers care less about the old ‘essential’ stuff.
This insight and others wouldn’t be impossible building AI into an existing pre-AI product. As a result, we’ve grown 81% in the last month and 35% in the last week (and also why it’s taken me so long to finally finish writing this ๐ ).
The edge of possibility & the AI constant
AI is moving so fast that you have to build at the edge of what’s possible. It’s like falling forward, since you know the the inevitable progress of AI (like gravity) will get you the rest of the way (and if it doesn’t, maybe you’ll at least bounce off the ground ๐ ).
Discovery vs. invention
The other side of that challenge is building something that automatically gets better as AI does. One thing that’s fundamentally different about building AI-native products is that it often feels more like discovery than invention.
That’s when we get down to the real issue. ‘Incumbents’ are companies, which are groups of people. And all of these people have motivations, managers, goals, and deadlines to meet.
It is very hard to have the space to experiment and explore unless you’re a massively profitable company with a healthy R&D budget. But even in those cases, those explorations often are not tied to the product teams delivering customer value.
Who are incumbents (AKA targets)?
Incumbents typically means big companies leading a space, but with AI ‘incumbents’ are any company (including startups) who:
- Have customer base sufficiently large a startup might target it.
- Are NOT AI-native: Have an existing business prior to early ~2023.
- Have a busines predicated on the way things used to be (customer expectations and employees tho make it hard to change)
- Are not founder-led, thus have a harder time adapting product vision
- Note: Some non-founder leaders can do this, but it’s much harder without founder credibility, thus relatively rare.
This is who the new AI-native startups are competing with, and often with much fewer people thanks to the efficiency of AI.
Because if you’re in this boat, you often don’t have the flexibility, vision or runway to sufficiently adapt. So you’ll fool yourself into believing that slapping AI onto your existing product (even thoughtfully) is sufficient, but it’s not.
So, what’s an incumbent to do?
How do you fix it if you’re in this situation? The best bet is likely having a new team (with fresh perspectives) build a new, separate product from scratch that is AI-native.
Then test a small amount of traffic until success is within striking distance of the existing product, knowing that as AI improves so will the product. One nifty opportunity is if the existing product is paid, make the new product freemium since it’s likely a ‘light’ version of the existing product in the beginning. At a minimum, your teams will learn a lot about building with AI.
If POC proves itself, you can look at what integration is possible, a migration path for customers, etc. But that’s typically also a good opportunity to simplify your business and get rid of tech debt since we’re only at the beginning of the AI revolution and you’ll need much less complexity to move quickly and adapt to a volatile future.
A fish climbing a tree
But…often that’s not even enough because 0-1 people are often different than 1-N people. Incumbents traditionally need people who are best at iterating on existing products, so those are the kinds of people they have.
But it takes a unique kind of vision and creativity to start from scratch and build something better. I expect we’ll see many more incumbents acqui-hiring tiny (<5) innovative AI startup teams teams to jumpstart new products, then leverage their existing distribution advantage to win.
Closing thoughts
Whether incumbents can get innovation before startups get distribution is truer than ever. But I don’t believe most incumbents have sufficiently adopted this new secular trend.
Incumbents may adopt AI superficially enough to check the box, but in time the minimal impact to their bottom line by losing customers to AI-native startups will show the truth.
This is especially true because AI startups often innovate to the degree that they demolish barriers to entry incumbents never realized existed, thus vastly growing markets (along with creating new ones) in unexpected ways.
The interesting thing is that this isn’t an AI research problem anymore: We don’t need significantly better AI, AGI or anything like that. Even today’s models are capable enough to make improvements for the next 5-10 years.
The problem is on the applied side: We need visionary, creative people to leverage AI to improve everything around us. And that’s the rare thing in this new age of abundance. But maybe it’s you, so start building ๐
