The Business of AI Agents
where incumbents will lose
We’re in the home stretch of what is undoubtedly the year of AI. ChatGPT crossed 100m+ users early in the year, people are talking to chatbots for an hour a day, and generative AI has made its way into seemingly every conversation in the startup world.
There have been two important storylines this year.
First, there has proven to be a real demand for AI products, both from consumers and businesses. It’s evidenced in the ChatGPT and Character AI metrics, Github Copilot’s rapid ascent to $100m+ ARR, and the number of startups that have seen huge growth in users and revenue this year.
The second, and arguably more interesting storyline is that incumbents have been quick to embrace AI. Just look at how public companies have all jumped on the AI train during earnings calls.
This is unlike the transition from on-prem to cloud, where existing software players were more skeptical and dismissive of cloud as a distribution and business model. This matters because it points to the fact that LLMs are not inherently disruptive to existing businesses.
If incumbents continue to embrace AI, then are there opportunities for new company creation? I’d argue the answer is yes, but less from chat-based AI copilots and more from AI agents.
AI Copilots – Strategy of the Incumbents
I’ve been thinking about this tweet from a few months ago more and more recently.
“a copilot is the strategy of incumbents”
In my mind, the two things that incumbents cannot easily change are (1) their GTM motion and (2) their business model. New startups are able to displace existing vendors typically when they innovate on one of the two. An example of (1) is Snyk taking a developer-first GTM motion to take on application security incumbents who sold top-down. An example of (2) is Salesforce’s cloud-first approach against Siebel in the CRM market.
The challenge for new AI startups is that AI copilots do not necessarily require new GTM motions or business models. It’s why incumbents have been so quick to embrace AI. For example, Github’s core product is priced per user per month, and unsurprisingly, Github Copilot has the same pricing model. It’s effectively a new product offering that fits seamlessly into the existing business model.
Where AI Copilots Shine
Without a doubt, AI copilots have a significant opportunity to bolster existing software offerings. I’m bullish on features like Adobe Firefly, which augment existing workflows with AI.
But for startups, I think the opportunities will be more narrowly confined to areas with limited software incumbents:
Vertical software – One way to avoid software incumbents is to pick a small enough niche. Today, there are a handful of vertical SaaS giants, but far fewer than there are distinct professions. A company doing this well is EvenUp which focuses specifically on injury lawyers, a customer base that is not inundated with software tools. Similarly, there are other such verticals that AI can significantly expand the market for.
New markets – Alternatively, I think AI copilots have to go after entirely new markets, which is what companies like Character.ai, Jasper, Synthesia are doing. They are taking some of the unique advancements in AI today and attempting to create a new category where companies don’t already have a significant footprint. However, a key challenge here is building something defensible, when it’s likely that a handful of players will be significantly well-funded.
The Bigger Opportunity – AI Agents
If AI copilots are software that helps humans do work, then AI agents are essentially software that does the work entirely. Agents are designed to operate independently until some desired objective is achieved. In this way, AI agents more closely resemble how humans do work, whereas AI copilots focus on automating tasks that fit into the human workflow.
I won’t go into the specifics of Agent architectures because that warrants a post of its own, but recommend this post as a starting point. The short version is agents take in some high-level objective and inputs, then break it down into simpler tasks, and iterate until they feel they have adequately reached a solution.
To do this well, Agents need to plan, prioritize, and self-critique along the way. Additionally, they need to access different sources of data or search the web for answers. Foundation models are the engine that powers many of these decisions and reasoning activities, but a lot more needs to happen for Agents to be successful.
Agents are a new type of product that has the potential to disrupt existing software business models because the value they provide is not tied to # of human users.
Consider Github Copilot, priced on per user per month basis. This makes sense since Copilot is helping each individual user do their job more effectively. However, an AI agent for coding could be priced on actual output. An example would be Sweep which prices based on # of closed tickets. This distinction may ultimately pushes these companies in two very different directions.
We’re still at the very early days of AI agent value creation. The bar is high for agents given that they can’t just be “helpful”, but rather they have to reliably solve user requests which can be highly variable.
Despite what I think will be a slow development and adoption curve, agents could be the new shape of startups and the next wave of disruptive companies.
If you’re building Agent-based applications or supporting infrastructure, I would love to chat. Feel free to reach out at nanilal@canaan.com or on twitter.