How VCs do diligence on ML startups
Understand how VCs evaluate early stage ML startups. How to design your business as a founder. How to evaluate your own readiness.
Hey reader, welcome to the š„ free edition š„ of my weekly newsletter. I write about building startups and careers in ML. You can learn more about me here. Feel free to send me your questions and Iām happy to offer my thoughts. Subscribe to this newsletter to receive it in your inbox every week.
Productizing your ML offering requires more time and effort than a simpler SaaS product. Even though ML offerings are delivered as SaaS, the underlying mechanics are different. Martin Casado and Matt Bornstein at A16Z wrote a phenomenal article on it.
As an early stage founder, you'll find yourself in a position where you have to fundraise with no real traction. What points should you focus on? How do VCs evaluate early stage ML startups?
Here are 8 criteria that a good VC will use to evaluate your startup:
1. Are you a seller of picks and shovels? Or are you a gold miner?
In the business of AI, it's better to be a seller of picks and shovels vs being the gold miner. For example, you're better off being the company that labels images vs the company that has to accurately predict a future event.
Some of the most successful AI companies have taken this path:
Scale AI labels your data. Itās up to their customers how they use that labeled data to do any task.
DataRobot provides a way for data scientists to build models. Itās up to their customers figure out what models to build and how to use them.
Datadog monitors your cloud infrastructure. Itās up to their customers to decide what to do information.
Snowflake provides data warehousing. Itās up to their customers to figure out what to do with the warehoused data.
Fivetran enables analysts to build ETL pipelines from different sources to different warehouses. Itās up to their customers to figure out what to do after building these pipelines.
2. How do you price the product
Pricing is an integral part of the product. It shouldn't be an afterthought. It's not something that the sales team would take care of separately.
The exact dollar amount is not relevant in the early stages, but picking the right unit of work is critical.
For example, are you pricing it based on the number of users? amount of data? number of workflows? license? You need to make it very easy for your customers to understand. What is it going to cost them today? What will it cost them if they buy more? Let them consume the product in the way they want.
As an AI product, you're in the business of converting raw data to meaningful information. A good pricing model for AI startups would use the following levers:
Number of ways you're processing that data: This translates to number of data pipelines they're renting from you. It's MRR or ARR.
Speed and volume of data being processed: This translates to how much is flowing through each pipe per unit of time. It changes from month to month, but it will smoothen out over time. Prepaid contracts will help as well.
Snowflake has been very successful in building a consumption-based pricing model. There's an upper limit on the number of users, but there's no upper limit on the amount of data that can be generated.
3. What's the buying process for your product
To make AI work in the wild, it takes effort from both sides. You and the customer will have to work hand-in-hand to make it successful.
If the buyer is expecting an AI black box that will be 100% accurate at all times, then your AI product will never work. The basic premise for the customer needs to move away from this.
A better premise would be "Will this product help me automate stuff that I'm doing manually today?".
Good AI products automate the boring tasks. Savvy buyers understand this.
Even if the buyer is savvy enough, a bad buying process can ruin the whole thing. Can the customer buy the product and start using it without talking to you? Or is this a product that needs a lot of setup work?
Both types of AI products have been successful, but you need to understand how your buyers buys AI products and design your buying process for them. Right from the first touchpoint all the way to signing the deal. Make it easy for them to see the product, test it out, understand how it work in their specific case, and then let them use it.
4. Do you know how to recognize recurring vs non-recurring revenue
AI startups usually end up doing a lot of one-time work in the early days. This services-type work is low margin. That's completely okay. What's not okay is founders trying to pass that off as recurring revenue.
Good VCs know that you're the early days, so there won't be smooth graph of recurring revenue. Not in enterprise SaaS anyway. If you try to dress it up like that, it just shows that you're don't understand how your revenue engine works. That's a tough thing to look past as an investor.
At the end of the day, you're building a business and the VCs are funding a business. You need to show that you understand the basics and that you have a plan to get to recurring revenue. If you 5 good pilots running with large companies, that's a great position to be in.
5. What's your defensibility
Spoiler alert, it's not your algorithm. I see early stage founders obsessing about how their "proprietary IP" is the key differentiator that's going to propel them into the stratosphere. Every algorithm that you'd ever need has already been invented. I'm not saying new algorithms won't be invented, but don't hang your hat on this tiny little leaf. Your proprietary IP won't save you.
Data belongs to the customer, so the size of your database is not defensible either.
It's not about having a few extra features either. If they're good enough, they'll get copied anyway. That cute patent of yours won't save you.
Real defensibility comes from figuring out a way to become essential enough to your customer's core business.
Having said that, here are a couple of things you need to sidestep:
If your customers think you're a magic wand that will act as a competitive differentiator to them, they would want to keep you a secret. This means they won't provide references. They don't want others to know they're using you.
If you're too essential to their existence, then you'll become a risk. They'll want to build this tool themselves.
At Plutoshift, we use many SaaS tools to run our business e.g. conference calls, chat, sending notifications, CRM, sprint planning, support tickets. We could probably build those tools internally, but it's not a good use of our time. They automate away the boring tasks for us.
As Jeff Lawson says in his book Ask Your Developer, build the customer facing stuff yourself and use off-the-shelf tools for everything else. Because you see, nobody cares if you reinvent Zoom or Slack.
6. Are you selling to SMBs or Enterprise
Both avenues have worked for AI products, but it can't be both simultaneously. Not in the early days. You need to pick an avenue.
If your AI product needs a lot of setup work, you can't afford to sell it for $99 per month to SMBs.
If your AI offering is a self-serve product without a sales process, you can't sell it to the Enterprise. Because they need services-type work to solve messy business problems.
If your product is built for SMBs, it would mean you need to have a self-serve product. A tool that can be used by the user without ever talking to you. It's a good distribution strategy that has worked well.
Enterprise would get you a lot more revenue, but the sales cycles are long. Data is really messy in the real world. If your product is built for the Enterprise, make sure you design a good sales process. Because they can pay you real money if you solve a real problem. You'll have to put it more effort to make AI products work in the Enterprise, so you may as well get paid for it.
7. Are you horizontal or vertical
There's no universal AI model that can do everything, so you need to work that into your business. If you build a generic horizontal product, nobody will be particularly passionate about it. If you build a tool for everybody, then it means you're building it for nobody specific.
If you want to do horizontal from day one, then you need to be self-serve. Low priced product focused on large number of small customers. Your product should mean the world to a small group of customers before it means anything to others.
If you find yourself customizing your AI product a lot, it's time to go up-market. Way up-market. You need to get paid for this work. And you need to high ACV to justify your high CAC.
In order to go really big, you eventually have to mean the world to a lot of customers in many different sectors. You'll have to do enterprise. You'll have to build sales teams. You'll have to make customers come to you.
8. What's the biggest company in your sector
A good VC will always try to think about what would happen to you make it big. What does the biggest success in your sector look like? If the biggest company is not big enough, then it won't be exciting.
If you're building a CRM product, the north star is Salesforce. If you're building a data warehouse product, the north star is Snowflake. If you're building a video conferencing product, the north star is Zoom.
If the north star is a big successful company, you'll benefit from the tailwinds. You might eventually become a category creating company that's one-of-a-kind. But in the early days, you should know what you want to be compared with. If you don't pick one, something will be picked for you by the market.
Hello, thank you so much. This was very insightful. I do have a question though. Can you please explain why it is better to sell picks and shovels instead of mine gold?