The AI Revolution is Still Early
ChatGPT is well-known, but this chart shows that enterprise AI adoption is still modest. One way to interpret this is that the AI revolution is overhyped. Another way to interpret this is that the AI revolution is still early and sticky enterprise use cases are concentrated among early adopters.
Software developers are the poster child of early adopters. They quickly embraced coding assistants and helped tools like GitHub Copilot reach $100 million in annual recurring revenue (ARR) shortly after launch. Now, GitHub Copilot is set to hit $2 billion ARR by year-end. If you're a developer, the AI revolution has already happened for you and you're already seeing a significant productivity boost from code-gen tools. You scoff at other developers who refuse to use code-gen tools. You shake your head disapprovingly at them. It's already obvious to you that they'll be left behind.
Other jobs will follow this pattern. AI tools will succeed by fitting into work routines, not as standalone chatbots. They'll be like coding assistants, built into existing workflows.
Build workflows, not chatbots
There are at least two ways to productize language models:
- General-purpose chatbots like ChatGPT and Claude that have unconstrained inputs and outputs
- Embedded workflow tools like Github Copilot that have a custom UI/UX for inputs and outputs (optionally embedded into some existing product surface area like an IDE)
These approaches are not mutually exclusive but solve different problems.
Chatbots work well for broad research and casual use. However, they are not well suited for predefined workflows. This partly explains why many people don't use ChatGPT regularly.
Daily predefined workflows are the exact kind of thing that enterprise use cases are built on. For these we'd have to unbundle general purpose chatbots into specialized tools.
I particularly like this chart from swyx showing various examples of this unbundling.
This is a natural evolution of the AI landscape. Embedded workflows reduce cognitive load as users don't have to figure out how to craft complex open-ended prompts for their use case. Good old-fashioned buttons and specialized UIs encapsulate context and make LLMs easier to use.
New technology cycles take time
I like to remind myself — and anyone who claims the AI revolution is not real — that new technology cycles take time.
Tim Berners Lee developed the World Wide Web protocol in 1990. The first web browser, Mosaic, was introduced in 1993. However, it took almost a decade after the development of the world wide web for Amazon to go beyond selling books and for e-commerce to become a mainstream category.
The AI revolution hasn't happened yet. It's still so early.