Happy Sunday and welcome to InsideAI! I'm Rob May, CTO and Cofounder at Dianthus, and Venture Partner at PJC. I mentioned last week that I was stepping down from the General Partner role at PJC to start a new company. Today I want to walk you through that thought process because Dianthus doesn't look like most other startups. It's an AI-First company, and everything about it was designed with AI in mind. Given that I've spent the last 6 years writing about AI for this newsletter, and made 100+ early stage investments in AI companies, I've seen a lot of what works and what doesn't. Some of the lessons we are learning about AI businesses show they are very different than SaaS companies in some key ways.
In today's newsletter, I'll walk you through my thought process on Dianthus, why we made certain choices we made, and why I think you will see more companies designed this way in the future. So I hope you enjoy this reflection.
What does Dianthus do?
Dianthus builds AI software for ecommerce, but rather than sell that software to ecommerce companies, we buy ecommerce companies with debt and plug them into our tools so that they can grow faster and increase their margins. This allows us to build models for things that might not make sense to sell to others, and for more targeted specific use cases, but still apply AI/ML at scale. We have an AI-first culture, so we constantly think about how to turn every task in the company into a machine learning model. This machine-assisted leverage allows us to scale marketing faster, and operate the business with more automation and thus more economic efficiency.
What kind of ecommerce companies?
We are starting with D2C companies in the Shopify and Woocommerce ecosystems for a few reasons. But longer term we will expand to be omni-platform and omni-channel. Running this type of business can be operationally complex, but, this is a great area to apply AI/ML to handle some of that complexity. Over time we can also expand into retail, as there are many AI/ML opportunities there as well.
Many of the players in this space today - Thrasio, Perch, etc, have a similar model but are Amazon FBA focused. The benefits of this are that integrating these businesses are much much easier than integrating a bunch of D2C Shopify stores. But there are some problems. Amazon takes a huge cut and thus margins are better in the Shopify ecosystem for some types of products. Using Amazon as your main channel fuels growth but, Amazon has the main customer relationship and some data and information they don't share. Owning Shopify stores gives us first party data which is better for ML models, and more control. That first party data and ability to experiment faster and more directly, leads to quicker learning and more powerful ML. Plus there is a burgeoning anti-Amazon movement among both customers and third-party sellers who feel Amazon is abusing it's power. We think that is the movement to ride long term, and we think it has opened up a space to build a new kind of ecommerce company - a company eventually as strong as Amazon in AI but disaggregated and distrubted across many sites with a common AI/ML infrastructure layer, instead of a single front-end. These businesses are slightly more difficult to identify and acquire, but that's where AI can help.
Why use this AI/ML software to acquire companies instead of sell the software?
This is the biggest thing some of our early investors didn't understand, but the properites of AI sometimes shift how we should think about business opportunities. Buliding SaaS software was about performing a task where the primary value of the software was to enable a human to manage a workflow with the tool. While there are some ML opportunities like that, most of them are different. AI/ML tools are usually making a prediction, or automating some task. There is less human intervention. When there is human intervention, it's in the form of annotating data, validating the model, or correcting the output - things you don't do in SaaS apps today.
In certain cases, like maybe predicting a security intrusion or financial fraud, it might be best to build a ML app that you sell to multiple companies because that cross company data will make it better for every company. In other cases though, the company's own data may be the most important data set, so you capture less value when you sell to them because you are just enabling something with their data. After spending several months digging into AI/ML tools for ecommerce, I concluded that most of them will fall into this latter category. Selling tools to small to mid-sized ecomm brands will be hard, in part because they won't use them correctly either. The cultural/training/workflow issues around this could be problematic. Plus many ecomm brands don't want their data used to help other ecommerce brands, but if Dianthus owns, say 100 brands, we can learn from all that data.
In addition, by building some of our own tools in key areas, we can turn economies of scope (doing different tasks with different data sets) into economies of scale (learning across many tasks and areas and customers). And more importantly, we can build an AI first workforce that uses our tools in a way that pushes our AI models forward.
And finally, for many SaaS applications, you have to water down the UX/UI to make it palatable to the most types of people and workflows and all the different use cases your customers could have. By building tools just to use ourselves, we can save a lot of time and move faster, integrate the tools we build with each other easier, and have more control over how people use them.
Another way to think about it is - hedge funds don't build software that they sell to day traders - they use it themselves. If you have a significant edge, better to use it for yourself. It makes sense to sell software to others when your edge from using it is minimal, and there broad value to many market participants if you sell it to them. If you can build just for your data and your use cases, it makes sense to keep it internal, and for AI, many people misuse or underutilize the software anyway, which is why many AI software firms grow so slow.
Where does Dianthus apply AI?
The short answer is - everywhere. I'll actually go through some of that here and share much of what we are working on, but I do have to keep the most advanced stuff secret for now.
There are three key parts to this business. First we have to identify acquisition opportunities, select the appropriate ones, negotiate a deal, and get it closed. Second, we have to integrate the business into everything else we have. Third, we have to operate and grow the business. We have architected the org structure to map to this workflow, with a VP of Acquisitions and a VP of Onboarding and Integrations, and both of those teams are separate from the core operational team.
If you take the first section - acquisitions, this area is ripe for machine learning, but it's an area that wouldn't make sense to build a standalone tool to sell. If I built a machine learning model to predict ecommerce success based on various market and product factors using data about trends, word vector analysis of how companies position against each other, search and ad data, it would be a tough product to sell. But if our business is identifying acquisitions, it's a very useful tool and provides a strong advantage. And it gets better as we make more acquisitions and can ground the model in real sales data we have about trends and customers. This is where some defensibility comes from - our model of what to acquire gets better over time as we acquire more, and it's really difficult for someone to get access to the same data to catch up. Plus, as we build specific capabilities, we can hone the model on certain industries, platforms, or customer types where we feel we have an advantage.
When we acquire companies, we can also look at the marginal value of additional data. How do the acquired data sets augment, complement, or interact with what we already have? Does the target company have data they don't realize is valuable or is underutilized? After looking at hundreds of companies so far, it seems most companies are underutilizing what they have.
On the integration side, one of the challenges in buying so many small brands is figuring out how to put all the systems together, or when to keep them separate. Much of this can be done with RPA, but my first thought when I saw OpenAI Codex was - the path for this kind of tech is to use it for these kinds of integration tasks, that are straightforward, but boring and slow and expensive to code. I think overtime Codex technology can be used to ease the burden on systems integration, but that will be a while. This isn't an area we are initially focused on for a lot of machine learning but, over the long term the opportunities will emerge to apply it.
This is where there are so many opportunities to apply AI/ML. Some are obvious like cross channel ad optimization, predicting whether a discount will drive more overall revenue, ML versions of existing tech like retargeting. If you have data (or own multiple brands) then cross product customer analysis and promotion can be helpful. On-page conversion optimization via ML is a big area.
Recommendation systems are a well established area where it's probably better to buy than build. Similar things exist around checkout cart abandonmen, customer insights, and promotions of all types. There's a whole bunch of stuff like this we can use and the determination of build vs buy comes down to the use and value of data and how well that matches to the pieces we consider our special sauce.
What is less obvious is some of the automation options. We built a tool that helps brand managers automate ads with GPT3. The tools suggests advertisements, then can place them, monitor them, and allocate spend all with machine learning. Yes we could buy something like that, as there are some cool companies doing it, but I think we benefit more and write better ads when we can feed it more internal data, and I think we can push that technology forward faster than an independent company can.
And there also things I haven't seen anyone try - like market positioning using word vectors. Can you figure out the gap in the market using NLP and suggest value propositions and wording that will cater to a segment of the market not being serviced very well? Can you monitor shifting market attitudes and stay ahead of them?
There are plenty of automated things around customer support, inventory, and supply chain. These are areas we won't build our own stuff initially but, as time progresses, it might make more sense to at scale, and as our ML capabilities improve.
The one area we will most stay away from that has high AI/ML potential is shipping/logisitics. The physical nature of this makes it so hard to do well until you have decent scale that we just aren't going to address it for quite some time.
The financial piece of this is probably where AI will impact it the least. Raising a little equity and using mostly debt for the acquisitions is well understood, and AI isn't going to do much there. Probably the best financial application of AI will be helping us figure out the growth potential and thus the price we can pay. But this is still an exciting thing to pay attention to because I think you will see many more acquisition-focused AI startups. Buying into something gives you data and workflows to start applying ML to, instead of having to build up something from scratch. What we are doing in ecommerce with Dianthus, that makes a lot of VCs scratch their heads and say it doesn't look like VC, will be a best practice for certain types of AI ideas in a few years.
The benefit though is that we can still be very innovative, but scale faster by acquiring EBITDA, and staying cash flow positive the whole time.
What about the non-AI pieces?
For all I've said about AI, my favorite thing about this model is, it can all be successful without AI. Just buying and growing businesses by being smart, even if you apply zero AI, can be a good model. If we have AI that works well and juices all those returns, that's what can turn a good business model into a great one.
Despite being an AI-first company, we've actually started with a team of people who have deep ecommerce expertise. My role as CTO is understanding what they do, and AI-fiying it. The industry expertise is just as important as the AI expertise in a company like this.
And of course there are obviously a whole bunch of areas to avoid, like products with unusual supply chains or other issues that require ecommerce expertise that is unique but can't benefit much from AI.
So in summary, I think machine learning models can have the most impact in ecommerce when the full data stack is owned by a single company and the tool is built for that data stack. And then acquiring companies which fit that data stack will allow us to capture the maximum amount of value. If you want to be part of something like this - we are hiring! And if you run a Shopify store and would like to consider Dianthus as a good home, we would love to chat with you.
Thanks for reading.