The Core of AI-First Companies
AI agents and AI workflows may look similar but they couldn't be more different.
In the previous essay, I outlined what an AI-first startup is: a startup that is enabled by and directly benefits from AI progress. AI-first companies are different from businesses that use AI to optimise operations but not in a way that’s core to their product.
In this essay, we’ll explore what products and companies will directly benefit from AI progress, and why AI agents are core to AI-first companies.
The next essay in the series discusses why wisdom, proprietary data and collective intelligence are important for high quality decision-making in AI-first companies.
AI as a core vs AI as an add-on
While every company out there is trying to make use of AI right now, there’s a difference between using AI as a core driver of value creation and using it to optimise some parts of the process.
Consider two video rental businesses facing the rise of the internet: Blockbuster and Netflix. Blockbuster used the technology to optimise their existing processes: creating a website to show store inventory and let customers reserve movies online. They were using the internet as a tool to make their traditional retail model work better.
Netflix, on the other hand, built their entire business model around what the internet made possible through streaming, fundamentally rethinking how people could consume entertainment. They didn't just use the internet to enhance their existing business model — they created an entirely new model that would have been impossible without the technology.
If you never heard of Blockbuster, it’s because Netflix won.
Likewise, there’s a difference between using AI to optimise parts of the existing process without rethinking it (the bare minimum we all should do) and using AI to rethink what’s possible thanks to AI.
To make sure AI is core to value creation, we need to discuss AI agents.
AI agents vs AI workflows
First, what are AI agents? There is no universal definition yet, but it doesn’t mean we should not take AI agents seriously. For the purposes of this essay, here’s how I think about it.
AI agents are systems that are capable of pursuing complex goals using AI to decide what to do.
AI workflows are systems that are using AI to do achieve their goals but without deciding what to do.
For example, an AI agent is Gemini Deep Research: a research agent that can search across hundreds of sources and then compile its findings into a good report. It uses AI not just to write the report, but also to decide which sources to consult, what information to extract, what information to prioritise and what to ignore, and when it has done enough research and needs to stop.
An example of an AI workflow is a feature of any video editor that allows to create a selection of short clips out of one long video. Yes, it will use AI to decide which parts to choose, but there’s a fixed set of steps in the process and little flexibility in changing them.
These two examples might look similar, but they differ in a crucial way. As AI becomes more powerful, AI agents will get correspondingly more powerful, whereas AI workflows are likely to experience more modest gains.
For example, today Deep Research produces good results but at a level of a smart graduate. Nothing to sneer at at all. But once it’s powered by an LLM 10x or 100x smarter (one year, two, three?), it will produce research at the level of a Nobel laureate, making new scientific discoveries. It will consult more sources, discover new insights and connect the dots in ways that it can’t do today.
Simply giving it a more powerful AI model will make the product more powerful.
However, an AI workflow that makes you video clips for social media out of a long-form video will only get marginally better because it’s just following a series of steps to give you a predictable result. It’s like giving AI to your toaster: however smart it gets, it will still be making toasts.
So if you’re building a startup, it matters greatly if at the heart of your product is an AI agent or an AI workflow1. The former are AI-first startups that will get stronger with better AI, and the latter are regular companies that will not benefit from AI progress as much as LLMs get more capable2, doing more complex tasks out of the box.
Three properties of an AI-first system
So, how do you build an AI-first startup?
First, make sure that the value your product creates directly depends on the quality of the decisions that your product is making.
For example, an AI agent that builds software will get better as its underlying AI gets better. This is because the process of building software is a never-ending stream of complex decisions. Which architecture to choose? Which algorithm to implement? Which tech stack to rely on? How to interface with other systems? Ask someone for clarification or make an assumption? How to balance code readability and performance? Software developers (humans or AI) make countless decisions every hour. The smarter the underlying AI, the better the overall result will be.
Second, choose problems where significantly better decision-making translates into significantly better results. For example, drug discovery. There’s no ceiling on how good pharmaceuticals can get. They can always achieve better results, fewer side effects, lower costs, etc. So an AI system that researches new drugs can create value that increases in line with the underlying intelligence.
An opposite example would be an AI system that helps you invent new recipes. Sure, there’s a problem to solve there, but there’s a ceiling on how good your pancakes can be and how much it matters.
Third, enable the system to learn as it works. The underlying AI models will provide raw intelligence and general education — think a Nobel laureate with an IQ of 400 and five PhDs. But your system will need to learn from its own work in its own unique domain, just like any employee learns on the job, regardless of how smart they are.
Conclusion
There’s much more that can be said about AI-first startups, but these core properties of an AI-first system at the heart of an AI-first startup strike me as the most important:
Value delivered as a function of quality of decision-making
High or unbounded potential value to be created
Ability to learn from its own work
This is why AI-first startups will, at their core, have AI agents (not just workflows) that will have complex decision-making abilities with memory they use for learning, working on problems that have significance.
I believe LLMs will get much, much smarter but it won’t mean that we’ll be able to ask ChatGPT to fix climate change or invent a new drug just like that. We’ll still need companies and people coordinating the raw intelligence of AI with available data and the real, messy human world to create things we’ll be impressed by and proud of.
And that’s the task for AI-first builders.
A reminder: This newsletter has two sections:
Unconditionally Human (what you’re reading now)
Startup CEO Succession (the podcast)
You can (un)subscribe to them individually.
A counter-argument here could be that the difference between AI agents and AI workflows is simply the degree of freedom or scope of decision-making: an AI agent has vastly more freedom than an AI workflow. Still, a change in quantity is sometimes a change in quality.
That’s not an excuse not to implement AI workflows wherever possible. Again, to consider the internet/Blockbuster analogy, they should have reinvented their business model like Netflix did, but at the very minimum they should have used the tech to optimise their retail model, as I think they did.