19. How do AI startups win?
Starting late 2022, I used to think and wanted so badly to say that AI is a new wave on which many successful startups get built.
My initial position might have been stated as, a redesigned UI purpose-built around AI workflows wins. As long as startups offered these better interfaces for interacting with LLMs, the value delivered to users would be significant enough to compel a switch.
However, as time goes on, I think the answer to whether startups or incumbents win from AI is generally incumbents win, but it depends.In the last few weeks I’ve rotated on a few thoughts that seem to tilt the LLM advantage to incumbents.
With LLMs there’s no schlep or information advantage for startups. Building with the gpt-4 endpoint is so simple that putting it on a roadmap at a company like Adobe doesn’t require leadership to move mountains. Startups don’t win on any kind of sweat or secret dimension.
It’s been said a million times… incumbents have distribution. But stated differently, incumbents are already solving a known problem to a known audience with existing trust in their product. New “AI-first” startups might not have this clarity. All things equal, the incumbent wins.
Incumbents know what works, and have the data to build it. They can design more intelligent and effective AI features, on axes that really matter, into existing workflows.
In other words, application layer AI feels less like a product and more like an enabling feature. There are some use cases where this isn’t the case — for copywriters, Jasper is effectively a huge lever on productivity — but AI fits into app more neatly than AI eats up an app.
Does that mean startups should drop the AI focus? Not necessarily, but they do need to approach it differently.
Less focus on demos and more on demonstrated value. In other words, build products people actually want rather than just feel compelled to check out because of a cool AI feature. I’ve felt this one deeply building a recent project. A demo with lots of oohs that leads to some users and revenue, but usage that wanes as the novelty wears off. As a founder you so deeply want to believe you’re on to something, but being critical here is important.
Work backwards from the problem rather than forwards from the tool. This is obviously a trope, but for good reason. Is a chatbot answering my questions about a webpage is cool? Absolutely. Does it solve a problem for me? Not sure. Would I pay $10 per month for it? Definitely not. It’s early innings of figuring out what shape of problem is suited to AI, let alone use cases and UI designs that can be rethought.
Recognize domain data (and especially system of record data) is an invaluable product ingredient you likely don’t have. Let’s say you’re taking down Zendesk with a new automated chatbot. Great idea, but maybe users care about this chatbot sounding and answering like their team. What stops Zendesk from building the same product but with the last 5 years of conversations to match tone and structure? The question for founders then is, what overlooked or proprietary dataset might you have that unlocks a solution for users?
My directional bet is that the way for startups to win is by redesigning or creating new use cases from the ground-up, and capturing a fundamentally better (or new) process in a data model an incumbent can’t easily adopt. On the Zendesk example, does the data model for a customer service team look different when 80% of your agents are virtual and the majority of tickets are resolved near-immediately?
Even better, startups should jump on workflows that previously haven’t been possible to build, maybe because the latent space was too diffuse or too high-dimensional.
Thoughts on this still developing. Likely more to come soon.
Note: I wouldn’t say these observations apply directly to infrastructure-level startups like Langchain and GPT-index.
Thanks to Peter Zakin for helping think through some of this stuff.