Autonomous AI Founder: Can AI Be A Startup Founder
Thoughts on the future of starting a startup and the role of AI
What if the next Steve Jobs or Elon Musk wasn’t human at all, but an AI?
Imagine an AI that is finetuned on all of Elon Musk’s data to become the ultimate AI Super Founder. It does not sleep, it is curious, it researches, talks to customers, identifies problems, drafts plans, submits proposals, builds MVPs, hires and fires people, learns from data, raises funding, etc.
No one really saw it coming: an AI that can create works of art or even more crazy: code! How does this ‘creative’ ability translate to starting and running a startup? What if AI can be a founder? What if AI is better at being a founder than a founder? How does this impact allocation of investments? What does it mean for employment? What tools will an AI need to maximize its founder potential? These are the question I intend to explore in a series of posts.
Earlier this year I read Matt Schlicht’s Beginner’s Guide to Autonomous Agents. He made a list of industries where these Autonomous AI Agents would exist, but what he did not include was “founder”.
Edit: In Jan 2023, Sam Altman said in an interview with Bill Gates: “Someday, maybe there is an AI where you can say: Go start and run this company for me.”
Why should we pay attention? If AI can step into Hollywood's shoes, why not the startup scene? To put things into context. The movie business is huge, with a market size of $90+ billion. Startups are even bigger! Think of the tech giants like Apple, Microsoft, Amazon, Alphabet and Meta. They all started as tiny startups and now they're worth trillions.
I have seen plenty of examples of AI doing practical tasks like creating websites, writing blog posts, reading PDFs, coding APIs, etc. In this post I wanted to explore whether AI can function at a higher dimension: can it start and run a startup? I condensed and simplified the question into 3 buckets:
Can an AI generate startup ideas?
Can an AI execute startup ideas?
Can an AI lead people in a startup?
Can an AI generate startup ideas?
"There is no such thing as a new idea. It is impossible. We simply take a lot of old ideas and put them into a sort of mental kaleidoscope. We give them a turn and they make new and curious combinations." - Mark Twain
Perhaps AI’s most impressive feat is taking this sentiment to the next level. We are calling it “Generative AI.”
As a founder, you may be asking yourself what business should you start. How do you know there is a business opportunity? How will you find your customers? How much will you charge? Where will you start?
“The way to get startup ideas is not to try to think of startup ideas. It's to look for problems, preferably problems you have yourself.” - Paul Graham’s “How to get startup ideas”
So, perhaps a better question is actually: Is an AI able to look for good problems? Naturally, the AI does not have human problems, but it does have AI problems and it could start there to come up with startup ideas. I tested a few prompts on GPT-4 to see if the AI can self-identify problems. See screenshots of the responses below for the following questions:
As an AI, what problems do you have?
As an AI, what technical tools do you need to help humans better use you?
So clearly, an AI does have problems. Given that there aren’t that many other AIs that could be immediate customers just yet, perhaps a better starting point for an AI founder is to look for real-world/human problems. I have identified three potential ways an AI can approach this:
Explore its own model to find problems
Research the open internet to find problems
Talk to potential customers to find problems
We’ll unpack each point in a bit more detail to evaluate how good an AI is at finding problems.
Explore its own model to find problems
I asked GPT-4 to rank which industries it has the most insights into real-world problems:
Given its knowledge is dominant in Tech, Software and E-commerce, I continued asking GPT-4 a series of questions stacked on top of its own answers:
As an output of the series of questions, GPT-4 identified this specific problem in software development: “Ambiguous Requirements: One of the most common challenges is not having clear requirements. This can lead to significant rework later on.”
As a next step, I asked ChatGPT to propose a solution to this particular problem that could become a startup. It came up with an idea called Requirement Clarity:
Clearly, GPT-4 is able to derive human problems from its own model, but there is no way to tell whether it is a real problem without getting real user feedback.
Research the open internet to find problems
Imagine an AI Agent that never sleeps and continuously digests information from online news, research, forums, social media, and blogs. Its sole goal is to identify and unpack problems in real time that could potentially be startup ideas.
Take, for instance, the early days of the COVID-19 pandemic. As news outlets began reporting on a mysterious virus emerging in Wuhan, China, and the subsequent challenges in tracking and containing its spread, an AI Agent could have quickly identified several problem areas and business opportunities (likely faster than humans). Recognizing the global need for rapid testing, contact tracing, and remote work solutions, AI could have highlighted opportunities for startups to develop quick diagnostic kits, efficient contact tracing apps, and robust telecommuting software platforms.
In this example, Sully Omar tested AutoGPT to research a fake shoe company:
What is interesting is that is not only doing research but also ‘thinking’ and ‘iterating’ as new information is revealed.
This is where real user feedback becomes important.
Talk to potential customers to find problems
YCombinator famously preaches to founders to talk to their customers. As discussed in this Hackernews article, the primary goal is not to do sales, but to listen and learn about the customer’s needs and continuously iterate on your concept as your understanding of the problem matures. Similarly, The Mom Test by Rob Fitspatrick advises entrepreneurs to ask questions about people's lives and behaviors instead of seeking opinions on their product idea, to get unbiased and actionable feedback.
So, how will an AI founder find and talk to its customers?
Peter Levels created ideasai.com, a platform that generates startup ideas that are upvoted and downvoted by people. This is a simple version of talking to customers as it indicates which project could potentially appeal to users. Imagine an AI spinning up variations of this with industry-specific knowledge.
Using the AI landing page generator called LampBuilder, I manually created a landing page for the Requirements Clarity idea from the previous section. It is a taster of how an AI can spin up landing pages if it has access to the right tools. Here is a link to preview the site.
What is powerful about AI Agents is that numerous conversations can take place concurrently to gather actual user feedback. It is possible to spin up a swarm of agents that talk to customers on social media or live chat on a landing page. Not only can the AI chat, but it can devise a strategy to test various demographics and content while learning in real-time what the user wants.
Can an AI execute startup ideas?
Minimal viable products
Expanding on the previous section of “talking to customers” - one problem is that often you need to put something in the hands of your customer to get valuable feedback. Again referring to Y Combinator, they put a big emphasis on having a technical co-founder. Take for example the story of how Paul Buchheit built Gmail by focusing on feedback from early users. It would have been hard for Paul to guess what people wanted without the actual feedback.
It is exciting to see how far the GPT-4 can code basic apps. In this example Mckay Wrigley created a demo showing how GPT-4 builds a basic note-taking app from a voice command including authentication, backend, Github upload, and deploying to Vercel.
Another example of how fast AI-enabled coding is moving is this demo from Replit’s CEO, Amjad Masad:
Teams are seeing massive jumps in productivity using AI as highlighted in this Tweet by Allie K. Miller:
Taking this a bit further, Itamar Golan’s MetaGPT is a multi-agent framework that specifically focuses on the software development process.
Finding product market fit
"Finding product-market fit is not an algorithm." - Marc Andreessen
Assuming the AI is able to identify and define a real problem, the next step is to find product market fit - the holy grail of startup land! The idea is to start experimenting and test if you offer any real value for a target customer. Secondly, you want to also establish a growth thesis to answer the existential founder question: where can I find more users with the same problem?
Literally, nothing is more relentlessly resourceful than an AI Agent with access to the appropriate tools. If an AI can browse the internet, scrape data, submit forms, send emails, make posts, make calls, create proposals, incorporate, hire and fire, make payments, and sign agreements it is effectively unstoppable as a machine. Being relentlessly resourceful is not just about grinding ahead, it is also about learning quickly, adapting, being creative, and having the ability to identify and act on the right opportunity.
A different way to ask the question is whether an AI can pivot if it identifies a more promising opportunity. I tested GPT-4 on this ability by expanding on the Requirements Clarity idea from the previous section. Again I asked GPT-4 a series of stacked questions:
In the scenario above I used GPT-4 to simulate a persona revealing new information. This alone is a powerful tool, but obvious real user input would be even better. Clearly, the AI can respond to new insights and iterate. Now imagine hooking it up with a live product where it can make metric-driven decisions on the fly.
If you assume that the AI will have access to “office hours” with investors, advisors, management, and team members then all decisions will get valuable insights from actual humans in the process. As seen in this Tweet by Paul Graham, it may just well be the serendipitous conversations that unlock new insights or strategies:
It is also possible to consult various AI on decisions. Jeremy Nguyen made an interesting post comparing Bing, Bard & Claude next to each other for various tasks.
Founder-led sales
Beyond building an MVP or figuring out product-market fit, often founders need to shift gears doing sales. There is no better way to know if you are creating something people want than by charging for it. In Peter Levels’s bootstrap slides, he suggests, “to put BUY buttons on everything all over your site and app and see what people want to buy or not. Then remove what doesn’t sell.”
In the early days, founders often sell an incomplete solution, especially in B2B sales. A big part of the early sale is the relationship with the buyer and their perception of the future value. For consumer use cases a founder might run into a cold start problem. Andrew Chen talks about how to scale and start network effects. He refers to how Tinder had everything built, but could not figure out critical mass until they hosted an event on campus where it was required to install Tinder to join the event. Through this, they realized that by engaging 500 of the most influential individuals at USC to use the app, they could establish a robust, self-perpetuating network.
Ignoring the relationship aspect for a moment, there is a couple of things the AI can do better than founders. It is highly data-driven, analytical, precise, and iterative with a lot more parameters. What does this mean for sales? Imagine a potential buyer coming to the Requirements Clarity website (the example startup from earlier). While there is no founder available for a one-on-one call, the AI could leverage a live chatbot to learn more about the customers' needs, simulate potential results against those needs and keep iterating until expectations are aligned as best as possible. Take for example how Twig is doing this on the customer support front:
Customers’ needs often adapt to the ability of the startup or founder to deliver on certain expectations. The nature of the relationship between the AI founder and the customer would change if the abilities and limitations of the seller also changed. Theoretically, it would not be too much to ask the AI to help the buyer prepare a presentation with actual data to present to the decision-makers.
Can an AI lead people in a startup?
Will humans follow an AI?
Expanding on the human touch points of doing founder-led sales by looking inwards to the leadership team of a startup.
"When people are financially invested, they want a return. When people are emotionally invested, they want to contribute." - Simon Sinek’s Start With Why
But, that is not always the case. People follow leaders for all kinds of reasons. Sometimes it is just pure confidence in their leader. Koos Bekker is perhaps not widely known in Silicon Valley, but for a long time, he was the CEO of Naspers (a major stakeholder in Tencent). At Naspers Koos Bekker gained so much respect that employees would simply do what is asked, because “Koos said so” [“Koos sê so” in Afrikaans]
In my opinion, emotional investment does not have to come from a charismatic messenger but can come from clear expectation management, transparency, and trust in decision-makers. I can imagine a scenario where the employee can reverse interview the AI to really get behind how it is thinking about a problem and its guiding principles. If we assume that the AI founder keeps the context of previous decisions and makes decisions from a defined set of principles, it will start to think in a more predictable way that could sit well with employees considering joining the team. Now, I would imagine if I joined the team led by this AI founder I would want to stick to this purpose and it should understand the consequences of fundamentals.
Communication and growth
I think leadership expectations in the light of AI-managed teams will change dramatically. If AI starts to take on leadership roles successfully, human leaders won’t be able to keep up with AI leaders on the small things that can make a huge difference for employees. Imagine being able to get real-time guidance on how to do your work, feedback on where to improve, and automatically booked training sessions to upskill as needed.
"Clients do not come first. Employees come first. If you take care of your employees, they will take care of the clients." - Richard Branson
If you assume that AI can manage employee expectations better and groom employees with special care, it is easy to imagine that you would unlock the full potential of your workforce on a level that was not previously possible.
Empathy, integrity, and data-driven decisions
While some founders are known for cutthroat decision-making there is a general belief that machines would be worse - ultimately machines don’t feel.
Empathy, integrity, and emotional intelligence form a foundational component of human interaction, leadership, and relationships. This is perhaps the area where AI is scoring the lowest in the light of being a founder.
What I find interesting is how OpenAI has been able to “intervene” with GPT responses to such an extent that it is able to respond “appropriately as possible”, sometimes to extreme annoyance. I can see how AI models could be fine-tuned to oversee more sensitive conversations or decisions. For example, when firing someone, making public statements, etc. That being said, I think for more sensitive scenarios having a human-in-the-loop would be the most likely safety net.
Touching on my previous point, my gut feeling is AI can bypass emotional intelligence as a prerequisite for being an effective leader. Instead, our trust will be gained through its consistency, data-backed decisions, the opportunity for direct discussions, diverse input from advisors, etc. Similar to the argument that self-driving cars could eventually drive better than humans and humans rendered illegal drivers, I get a shiver thinking that eventually AI will be the only legal founder or CEO.
Quick thoughts on Autonomous AI Agents
It is still early days for Autonomous AI Agents.
"They break all the time, they get stuck, they can’t reach their objective, their abilities are too limited, or their completed result is not good enough.” - Matt Schlict’s Autonomous Agents Suck But Here's Why I'm More Bullish Than Ever:
Matt goes on to explain why he remains super bullish on autonomous agents and how people get stuck on two types of tasks: 1) those only they can do and 2) those they can outsource. While Matt looks at his personal to-do list as a low-hanging fruit for Autonomous AI Agents, one could look a bit further at the to-do list of a founder.
While things are buggy, it is moving fast! In this post, Yohei Nakajima reveals the latest version of BabyAGI, a fun chat interface to interact with the agent.
As a quick reference, Adam Silverman created a list of the top AI Agents. Alex Reibman is a great follow on Twitter if you want to stay up to date with the latest AI hackathon projects.
Quick thought on Y Combinator and AI Founders
YC has invested in over 4,000 founders generating over $600B in combined valuation. It can be argued that YC’s biggest ceiling is access to founders that meet their investment criteria. What if YC tried its hand at mining all its founder data and started backing its own AI founder(s) instead?
So is it time for Autonomous AI Founders?
Looking at it in isolation, never did I imagine that so many aspects of a founder’s role could be outsourced to a machine. It is still very early, but skill-by-skill I can see how almost everything can be done better, faster, and cheaper by an AI. Whether it is problem discovery, talking to customers, making decisions, inspiring audacious visions, or building and leading a team, I can see how AI can do it well or even do it better than most founders over time.
The obvious strengths of an AI are processing vast data, doing things in parallel, perpetual energy, multi-disciplinary continuous learning, real-time tracking and measuring, and predictive scenario mapping. Its weaknesses are emotional intelligence, leading hands-on, lack of human-like intuition/creativity, and building real relationships. What I concluded from this article is the possibility of mediating these shortfalls with special measures by leading with transparency and clear expectation management, having access to key advisors, having a human-in-the-loop for some things, and relying on team culture for relationship building. Just like founders aren’t able to do it on their own, an AI founder will need a team and resources too.
Obviously, we are still a few steps away from fully Autonomous AI Founders. Autonomous AI Agents keep running into useless corners and tooling is still in its early days. There is also the uncertainty of ethical considerations, biased models, and limited access to domain expertise and real-world data.
That being said, I can see how in the immediate short term a more constrained and human-managed AI Agent could get quite far as a founder or co-founder, but might not be able to win. Version 1 is perhaps more of a founder coach or assistant. As we have seen with the example idea, Requirements Clarity, it was easy to spin up a landing page using tools like LampBuilder. Stitching together more skills and tools for AI to leverage and automate the role of the founder is just a matter of time.
In future posts, I intend to experiment with using AI to help with startup idea validation stages as a starting point.