Principles for Building AI Products

A framework for disruption and defensibility

The iPhone was released 16 years ago, and we finally have a shiny new paradigm to enable the next big wave of startups. AI is the new electricity, and the opportunity to help it impact all aspects of society is there for the taking.

And yet, it’s unclear what to do. You could help “democratize” ChatGPT with a basic wrapper to save people the work of creating the right prompts and copying-and-pasting text between your tool and their web interface. These can be very cool and go viral, but often have questionable cost structure and defensibility.

The other extreme is to go deep and try to build a new foundation model. This is an expensive and deeply technical affair, and there’s probably only a few hundred people in the world who can pull this off. I’d also bet that all foundation models will be eaten by a single, multi-modal, multi-domain super intelligence anyways, so you’d better be #1 if you’re going to try this.

There’s another (big) problem. It’s just as easy for an incumbent to sprinkle in some AI magic as it is for you, and they already have users, a product, and a board that’s yelling at them to use more AI. As Ben Thompson points out, this actually suggests that AI is often a sustaining innovation (in the Clayton Christensen sense) and not actually a disruptive innovation.

if the innovation was sustaining, then incumbent companies became stronger; if it was disruptive then presumably startups captured most of the value.

AI in Microsoft Word? Those customers are paying money to create documents and improve the productivity of their organization, so adding a copilot there sounds like a slam dunk. They’ll even pay us twice as much! Sustaining innovation. To be disruptive, you need to build something that makes no sense for the incumbents to build. Kodak would have happily adopted a new chemical that made their film photos look better, but they were acting rationally by ignoring digital cameras.

So, what’s an AI product builder to do? Here are a few angles to get the creative juices flowing.

Focus on a smaller niche

The TAM of the internet is massive. Don’t go horizontal with your first product - that’s likely to compete directly with incumbents, and you’re less likely to be 10x better for everyone. Making a new tool for design? Rather than a new, AI-enabled Canva competitor (they already have a lot of AI magic anyways), try focusing on being 10x better at email templates for middle aged folks opening their first Shopify store, or fun customized holiday cards for grandkids to send to their grandparents. You can go super specific and still find thousands of customers. Not only can you serve them better by building specifically with their needs in mind, but you’ll have less competition.

Build AI-first, not AI-enabled

Briefly, AI-enabled is more of a copilot, where you have an existing product that works fine without AI, and you add some AI to make suggestions and increase efficiency along the way. It’s like a sailboat with a bonus steam engine. AI-first is a product that’s completely rethought from the ground up with AI, and it would completely fail if the AI stopped working. It’s abstracting away entire categories of work. Read this or this for more background.

Copilots are easy for existing applications to add, because it’s the same core workflow, and it helps existing customers achieve the same goals in a better way (again, a sustaining innovation). Autopilots are hard for incumbents to add, because their implementation would involve throwing away big chunks of the existing product, and could even serve a different customer. Github copilot is a great feature for an IDE, to be sold to software engineers. Github autopilot would replace the IDE and engineer, and might be sold to product managers. It’s much harder to pull off – the AI has to work really well, and building user trust will be a challenge – but if you can deliver, there’s more whitespace and much lower risk that you’ll be steamrolled by the incumbent.

Crossing the capability chasm will require major AI chops to innovate at the ML layer, which takes time, talent and capital. If it’s a high-value, low-latency scenario, consider having humans in the loop to ensure output quality in the short term, at least to validate demand.

Speed as a superpower

Never underestimate how much faster a small, focused team can move compared to a big, established organization. You might be able to copy a subset of a competitor’s app, and build in the new AI features, before they get around to actually adding the AI features themselves. The hardest part will be getting customers, but if you deliver a 10x experience for some important use case, they will come. At least, until this brief window closes.

A flash in the pan can light a fire

It’s cool to go viral, but if it happens and you’re forgotten about a week later, you’re going to wish that you had more to show for it than a massive OpenAI bill. The spike in attention and usage can jumpstart key product loops that lead to more value and defensibility over time. One great example is Instagram - many camera and photo filtering apps existed and died off, but it was Instagram that ended up building an awesome network effect and additional value on top and winning massively (h/t a16z podcast for the example). It really depends on your product, but defensibility could come from building a community, a new dataset, embedding deeply into a workflow with high switching costs, etc.

Build (or wait for) an actual platform

One of the key underlying problems is that ChatGPT isn’t like the App Store. They both provide a set of primitives that make it faster to build - GPS, internet connectivity, and a touchscreen device that’s always with you from the iPhone, and intelligence from ChatGPT. But the App Store also provides distribution, while ChatGPT doesn’t. The iPhone was a groundbreaking new interface, with its own new real-estate, and populating this real-estate with valuable new applications was a huge opportunity to access and provide value for users.

Some big questions to help frame the characteristics of this platform:

• What’s the interface? What new hardware is needed?

• At what level of abstraction is the user engaging with the new platform/apps?

• Will a user interact with one agent that takes care of everything, or multiple specialized agents?

• How does an agent take action and extend its of capabilities?

Here are two scenarios to make it more tangible.

Scenario 1: Many agents on your phone

This is the most likely scenario in the short-term. Soon we’ll have many more AI-first apps than ChatGPT: AI inbox managers, AI trip planners, AI twitter assistants, AI coaches, AI shopping assistants, and so on. Just look at your phone’s home screen and imagine accomplishing the goal that app serves more efficiently using AI. It’s not a bad experience to unlock your iPhone, open the Amazon assistant app, and interact with it to buy a new product. Each app continues to exist in its own silo, and it’s up to you to pick the right one for your use case.

Scenario 2: One agent in your ear

I think it’s inevitable that we’ll move to a Her-style, always-on voice interface, and that it’ll kill the current app paradigm. At one point it was actually possible to say “Hey Alexa, ask Cortana [the Microsoft assistant] to read me my schedule for today”, but it’s slow and puts a huge burden on the user to remember which assistant knows what information. If I’m looking for my latest message from George, I don’t want to separately ask for my most recent message from him in WhatsApp, and then in Signal, and then in iMessage. I just want to ask for my most recent messages from him and send a reply.

It’s cliché, but I agree that we’re headed to a Her-style platform

In this case, I think Her got it right, and a single “super assistant” is best. It should know everything about me and everything about the world, and do whatever I need as efficiently as possible. Under the covers, there will be a whole world of complexity abstracted away. The reasoning, knowledge, and decision making will be part of the platform, but there will be a huge opportunity for adding new capabilities and information sources. Ordering food, booking a flight, making a purchase, or booking a plumber represent messy interfaces with the real world, and providing these capabilities to the assistant will be hugely valuable for the user, and coordinating and interacting with these headless services would represent an entirely new platform. Massive value would flow to the assistant, but the services could also extract value with their own differentiation (e.g. Amazon has Prime and the fastest shipping).

Closing Thoughts

It’s true that it’s never been a more exciting time to build new products, and we’re at the start of a Cambrian explosion of new products and businesses. Thanks to new tools for design, engineering, and marketing, it’s also faster and cheaper to build new products and experiment. So what are you waiting for, go ship something!