Part two - read part one here.

We are trying to figure out how to direct artificial intelligence into growing a company, "growing" meaning revenue and with our company's revenue equation being:

Revenue = qualified traffic x conversion rate x average order value x repeat purchase x availability

As we wrote in part one, our core thesis is something like this:

  1. Businesses are process machines.
  2. Growth comes from a small number of high-leverage processes and so achieving growth with AI must involve the AI being aimed at those (and not at generic efficiency).
  3. The best targets are close to money, frequent, variable, high ceiling, and measurable.

This post focuses specifically on growth, which has one important feature - results are lottery-like. What typically happens is that most experimental results produce no or negative effects (a bunch of emails with no responses, for example, or some wasted ad spend). But every now and then there will be a true outlier which can change the business for good.

You can think of this plotted on a graph - lots of results clustered at the left (zero results) and a few dotted towards the right of the line - the true outliers, representing what statisticians might call a "fat tail".

Note: The simple reason for this is that "growth" is actually a bunch of things - consumer psychology, copywriting, novelty, timing, offer, audiences, competitive context, social proof, product. And they are multiplicative, not additive - so making one or two of the "correct" changes can vastly change the outcome.

Given this... It's a fair bet that the smartest way to drive successful growth experiments is just to experiment more, with the right stuff.

What is the right stuff?

When we started on this journey, we'd hoped that AI could help us answer this question. And it did, to a point, but we've since smoothed out some of the ideas and found that the sorts of things we need to experiment with to get growth are likely to fall into one of a couple of buckets.

Big Bets

These are experiments which might be termed non-linear - they don't follow logically from our existing set of business processes. They have extremely high upsides and they are very variable - often they'll be ideas which we haven't considered before and therefore exist outside of our current processes and systems.

Big bets have the ability to change the machine of the business - downstream effects on existing processes can be transformative, or they may require entirely new processes to be built from scratch. They likely change the ceiling of the business.

Growth Assets which Compound

We found AI to be excellent in ideating these. In one sense, they are efficiency plays - consider, for example, the idea of creating a library of every marketing hook (the first 3s of the ad) our ads have ever used, categorizing them, and making them searchable and swappable. That would save us a lot of time!

But more interestingly having a tool like this specifically in our advertising toolkit compounds, because it enables us to save time, make better ads, sell more stuff and feed the results back into the machine. That's a capability compounding over time.

AI has a role to play here because the creation of these growth assets has become - basically - trivial. AI can vibe-code the app I described above, and we can immediately benefit from it. And so building out a list of experimental growth assets is something we're currently doing, with the overall objective of building them all and seeing which has the biggest effect on our existing processes.

Growth Loops

This was where we thought our agents would be able to add the most value - something which we're still working through.

Loops are increasingly used by developers to prompt AI to complete complex, multi-step tasks, by taking advantage of AI's autonomous functions and ability to try different paths to get to a goal. They may fall into two camps:

  1. Execution loops, where the AI just does something regularly so that a human doesn't have to, which contributes to revenue in some way.
  2. True growth loops where outputs feed inputs - this is likely some kind of successive adaptation or "adaptive search" effort, where the AI can be tasked to hit a goal and can automatically adjust parameters in search of that goal.

Loops have an important characteristic that makes them ideal for agents - they are typically boring or high-effort to run, which means humans don't run them as often as they should. Our agents can take this work off our plates, with one important caveat - since we're likely delegating a lot of responsibility, we want to consider the cost of getting things wrong.

What's AI adding for us?

We're testing our agents across all of the growth experiments outlined above - as thinking and research partners for big bets, as developers of compounding assets, and as execution tools for growth loops where we can see significant value being left on the table by our human limitations.

The takeaway here is the same across all of those mechanisms - AI is collapsing the cost of running multiple experiments (including at the same time), to allow us to test more and (hopefully) grow faster. Being able to do this is important, as readers will have noticed in the mention of a "lottery" above - if you have more tickets, you're more likely to win.

We'll report back as we see results!