Wisdom and data in AI-first startups
On top of raw intelligence of AI models, AI-first startups need diversity of (AI) thought, proprietary data and access to human wisdom.
In Part 1, I argued that AI-first startups will be as different from existing tech companies (think Airbnb) as they are different from their predecessors (think Hilton) because they will be both enabled by AI and directly benefiting from it.
In Part 2, I argued that the core of AI-first companies that allows them to directly benefit from the improvements in underlying large language models (LLMs) is that they rely directly on quality decision-making to create value. For example, building software is a prime use case for AI (tools like Cursor and Replit) because the process of building software involves making a large number of good decisions.
In this essay, I will look more closely at the decision-making process in AI-first companies. I will argue that there are specific actions entrepreneurs can take to make sure their companies make better decisions.
Good decision-making
There are four major parts that underpin quality decision-making.
First, intelligence. A person or a system who is smart, knowledgeable, educated and can think clearly, will make better decisions. This is why top firms fight for graduates from best universities who did well there: academic success is taken as a proxy for raw intelligence that they can apply for other tasks. Likewise, this is why LLM progress is measures on various intelligence benchmarks, and why it’s so exciting when newer models show exceptional results.
Second, diversity of thought. A group of smart people working well together will outperform one bright individual through collective intelligence. It’s an emergent property, meaning that no single individual possesses it, but it shows up when a group works together. By supporting and competing with each other, as well as offering different perspectives, a group can do better than any one individual. This is true for people and I believe it’ll hold for AI, too1.
Third, data. To make good decisions, we need to have a lot of reliable data. For AI systems, much of it comes from general training, but the rest must come from internal company sources. Data will remain one of the competitive advantages for AI-first companies if it helps with decision-making.
Fourth, wisdom2. In particular, the bit that helps us to understand what matters from what doesn’t and see the world from multiple different perspectives. While AI can be incredibly smart, I’m yet to see any evidence that it can be wise, and yet we can’t make good decisions in complex situations without wisdom. This also includes an ethical and moral foundation that can’t be fully expressed in rational rules.3
So, if we have intelligent people/AI agents working well together as a collective, leveraging high-quality data and having access to real-world, hard-earned wisdom, they’ll make better decisions than any smart AI out of the box.
Implications for AI-first startups
This means a few things for AI-first startups.
First, as I argued in Part 1, they will naturally and effortlessly benefit from more powerful intelligence available to them. This will set them apart from non-AI-first companies but it will not give them any particular advantage over AI-first competitors.
Second, AI decision-making systems will benefit from having multiple agents coordinating and competing with each other instead of having one AI agent trying to do the job. This means AI-first companies will need to design their systems and processes accordingly4.
Third, AI-first companies need to use data as their competitive advantage. If intelligence is cheap and fast, data isn’t so easily obtained. For example, Boardy.ai, an automated service that makes intros5, has impressive tech but also it collects very specific data: what every person is currently looking for, in fine detail. So competing with Boardy would require not just rebuilding the tech but also getting the critical mass of users willing to share their data.
Fourth, AI-first companies need to have humans in the loop to allow AI to leverage human wisdom. When it comes to questions of what’s the right and wrong thing to do, we need lived experience and wisdom, not just raw intelligence, especially if it’s trained on data containing the entire history of human biases.
In conclusion
So, here’s how I think about starting an AI-first company. It must be creating value through better quality decision-making, directly benefiting from better AI models available to it. Instead of using a single AI model or a single AI agent, it needs to create a diverse team of agents, supporting, challenging, competing with and learning from each other. To create a competitive advantage, it must secure access to data that will help it make better decisions and learn faster. Finally, it must involve wise humans to bring lived experience into the process.
Of course, this is just one of many ways to think about AI-first companies. I’d love to hear your take. Leave a comment or hit “Reply” and let me know what you think.
I’m currently speaking to a number of startups looking to incorporate AI into their businesses, and to a few founders looking to start AI-first companies.
Check out the presentation I built to outline what I offer and what I’m looking for.
If you would like to discuss bringing AI into your business, drop me a line.
This is why, I believe, the team at
built an AI agent that refines ideas by giving them to different AI agents to critique from different perspectives and then synthesising the results instead of asking one AI model to do all the work.Check out my essays on Being Human in the Age of AI and Cultivating Wisdom in the Age of AI.
Several of Dostoevsky’s novels elaborate on what happens when people try to rationalise why they shouldn’t be bound by societal or moral norms. It inevitably ends in tragedy for everyone, just ask Raskolnikov. There’s no reason whatsoever to think any AI system has any notion of ethics or morals at this point, making humans in the loop a critical part of the process of good decision-making.
There’s a loose parallel here with microservices, a pattern in software development that allows to build complex systems as a large number of nimble and relatively independent microservices working together instead of a large and unwieldy monolithic system.
You should try it, it’s very impressive. Pretty much indistinguishable from a conversation with a real human.