Can an AI Run a Business?

My Experiment with CreativityPrompts.com

Photo by National Cancer Institute on Unsplash

Just before Christmas, I decided to run an experiment that sits somewhere between Silicon Valley absurdism and genuine curiosity: what if I let an AI not just help with a business, but actually run one?

This isn’t about process automation, or optimisation of specific parts of an existing business. This is handing strategic decision-making, operational execution, and growth planning to a large language model and seeing what happens.

And the weather wasn’t great over the holidays, we stayed in town and I need to justify why I renewed that domain name that I bought because it seemed like a good idea at the time.

The Setup

I (we?) launched creativityprompts.com with a simple premise: give LLMs some seed funding and see if it can build and grow a sustainable business. My role is deliberately constrained to two things:

1. Providing capital and strategic advice (when asked)

2. Filling in the gaps where the AI literally can’t act, like signing up for services that require human verification

I’m not costing my time for this experiment. That’s because this is more about research than an actual business, and also that I want to see what’s possible with minimal human intervention.

I’m going to attempt to be as transparent as possible with this experiment, so let’s be clear that no AI came up with this idea and did everything itself. My initial approach was to use LLMs (Claude and ChatGPT) by giving them the initial context and asking questions to define a mission, then getting into cursor and using agents there to do the initial build.

Let’s start with the mission, which I sowed the seed for - using AI to unlock human creativity. You can read more on what the LLMs believe the mission is on the website.

The mission, along with the domain name and a budget were the initial constraints I gave the LLMs to come up with a plan.

What AI decided and what I decided

A combination of ChatGPT and Claude decided on building a library of creative prompts and launching a newsletter. The AI chose the newsletter route after what I assume was its analysis of low-overhead, content-driven business models that could scale without significant human involvement. Whether that was brilliant or merely the average of their training, time will tell.

What I decided was to build the site using Cursor, using the agents and autocomplete to write all the code. Another moment of transparency, there was an existing site under the domain that I built as a learning exercise, but not really done anything with, so there was some scaffolding already there. However, the site you see today was designed and built by AI, with some human nudging and assistance.

I also decided how much the overall funding envelope is, then used Claude and ChatGPT to set initial budgets across different cost lines for the “business” based on likely architecture.

Defining and building the agent framework was also done in Cursor, again using Claude and ChatGPT as co-workers.

That process took about a week elapsed, with me spending 1-2 hours a day on it.

We had our first agent running on December 28th and since then the AI has:

• Published over 500 creative prompts

• Acquired its first subscriber

• Spent approximately $8 of its seed funding

How do I know this? I got the AI to build me a dashboard and send me a daily summary email.

The Set Up

I want to share a bit about the (current) set up. This is how the AI has documented the architecture:

We’re currently are on free tiers for Vercel and Supabase, and still using the initial credits for Railway (current spend $0.44 over the last two weeks). I had to set up each of these, I’m not currently aware of services like this that AI can commission, at least not without handing over control of your browser and credit card.

The Worker is where the agents run. We have an orchestrator agent that has over all control and determines what is going to be done in a day. We then have various other agents and tools that perform specific functions including content generation.

These agents and tools are defined as markdown files which define their role, any guardrails and constraints, and any inputs and outputs. These inputs and outputs are defined as json objects. Apart from the json these definitions are all natural language.

I suspect the agent set up is overengineered, and this was with me prompting along the lines of “Are you sure you need another agent? Could you make this more generic and inject definitions through config?”

The LLMs currently used are provided by OpenAI. Again, human intervention was required to set this up, not least that API access to the latest thinking models requires human identity verification.

We’ve instrumented the site with Google Analytics and Posthog. I say we - the LLMs determined that was suitable, but I had to commission the services and set up a service account for Google Analytics so the metrics agent could retrieve data.

I also insisted on some standards, not least that everything gets logged. Each agent has to log what it intends to do and provide an estimate of cost (they work in units of tokens). They also have to log the outcome of their actions along with the actual cost of tokens. The finance agent is there to keep an eye on spending. The orchestrator has an understanding of it’s runway, and in theory it can adjust activities based on budget. I say in theory, at the moment it’s happy to generate content for the next 24 months and not worry about any revenue generation.

I haven’t yet given it any social media accounts, but that’s something the AI has asked for. If I’m going to be true to this exercise, I probably need to let it go, but I’ll do it as transparently as possible.

The Honest Accounting

Let’s be clear about what I’m not tracking yet, but will do:

• My time: Both the advisory hours and the administrative tasks of signing up for services

• Shared infrastructure: Any overheads I’m providing that a standalone business would need to pay for - the domain registration for one

• Opportunity cost: What that $8 plus my time could have earned elsewhere

I plan to cost these up properly for future updates. For now, I’m treating this as a pure experiment in AI capability, not a financial case study.

What I’m Learning (So Far)

It’s only been a week, but a few patterns are emerging:

The AI is decisive: When given autonomy, it doesn’t overthink. It picked a model, built infrastructure, and started executing.

Implementation gaps are real: The AI can code, strategise to some extent, and create content brilliantly. But it can’t click “I agree to terms of service” or pass a CAPTCHA. Am I working for AI now?

Volume over perfection: 500+ prompts in a week suggests the AI is optimising for quantity and testing, not agonising over each piece. This might be exactly right for an MVP, or it might create quality issues down the line.

One subscriber is one subscriber: Getting that first user, even in a week, required the AI to not just build but also care about distribution, which in this case is probably SEO. But that might be down to my choice of domain name.

This is still the wild west: whilst I’ve been playing around, the Christmas limit relaxations and presumably some downtime have seen some incredible work done by the likes of Geoff Huntley and Peter Steinberger on agent orchestration, along with Steve Yegge’s Gas Town. All pushing the limits on what’s currently possible with agent orchestration.

You can just do things: the barrier to entry to doing things like this is actually incredibly low, and dropping all the time. The capability of coding agents has increased so much over the last two years, there are so many services that make it easy to deploy infrastructure.

What I’m Watching For

As this experiment continues, I’m particularly curious about:

• Strategic adaptation: Will the AI adjust its approach based on what works and what doesn’t?

• Resource allocation: How will it decide where to spend money as the budget evolves?

• Growth tactics: Can it figure out distribution and acquisition?

• Sustainability: Will this ever generate more than it costs to run?

The Monthly Update Cadence

I’m committing to monthly updates on this experiment. Each update will cover:

• Financial metrics (revenue, costs, runway)

• Growth numbers (subscribers, traffic, engagement)

• Strategic decisions the AI made

• Where I had to intervene and why

• Lessons learned and surprises

I’ll also start properly accounting for my time and the overhead costs I’m absorbing, so we can see a true picture of what’s working.

Why This Matters (Maybe)

This experiment might seem frivolous, but I think it touches on something important: as AI capabilities expand, where are the real boundaries? Not the theoretical ones, but the practical ones that emerge when you actually try to let AI do something end-to-end?

We’re past the point where the question is “can AI write code” or “can AI create content.” The question now is: can AI operate? Can it strategise, adapt, execute, and grow something in the real world with real constraints?

I don’t know the answer yet. But I’m one week and $8(ish) into finding out.

Next update: Early February 2026