cft

The Only Three Scalable AI Startup Archetypes

Traditional SaaS Business Models Don’t Work for most AI Startups


user

Praful Krishna

3 years ago | 5 min read

AI has been top of mind of the startup world for at least five years. Hundreds of billions have been invested, thousands of entrepreneurs and intrapreneurs — some of the best minds of our generation — have tried their hands with varying degrees of success, and none of the pure AI companies has made it to IPO.

The Problem

We all have learnt, some the hard way, that artificial intelligence cannot be packaged into a product. The factors that scale a typical venture backed company, say in SaaS, don’t apply to AI.

  1. First, to be useful an AI must train on data specific to the customer and to the problem it is trying to solve. Despite the various claims of so-called “pre-trained” models, their accuracy is nowhere near reliable e.g. GPT3. In general, AI can’t be copy-pasted easily.
  2. Those that have attempted an industry-specific approach hoping to train on customers’ data have met with an impenetrable wall of data security requirements. For example, most enterprise customers want vendors to install within their firewalls, sometimes on premise.
  3. SMB and consumers are open to sharing their data in return for some functionality, which is the fundamental promise with SaaS. However, to succeed, an AI company will have to sign up many many users before their algo becomes effective. There is no incentive for the early user to volunteer their data.
  4. Then there is the problem of deterministic behavior. Say you sign up with an AI that tells you ‘two plus two is five.’ It’s alright. You design your processes around the result, like subtracting one. Over time, the AI gets new data and one fine day changes to ‘four’. You must redesign your processes now. And then again if for some reason it goes to ‘three’. (See the link below for more on this.)
  5. Next, is the ‘Buzzword Compliance’ problem. I remember advising an entrepreneur once that his solution does not need AI at all. He candidly objected, ‘Then how will I raise money?’ I tried to explain that he is better off positioning himself as a SaaS company if he wants to scale, but that didn’t go too far. These days everything is called AI, and that just muddles the field.

The list continues, but let me pause here.

Flight to Data

Many of you would remember the iconic image in 2017 that depicted data as the new oil. They hit the nail on its head. Today, because of the problems I list above, technology hardly matters in the business of AI.

In the enterprise world, companies are primarily building their own solutions and keeping data close to their chests. In the consumer world, behemoths like Google and Facebook are using their data with AI to build new services for consumers. Partly for SMBs, too, but largely the SMB world is woefully underserved by AI.

In all this, startups have had little role to play. Even the talent is going towards these large, data-rich companies. The State of AI 2020 report gives interesting statistics on how professors are leaving academia to work for giants like Google, Amazon and Microsoft. There is a direct correlation between such exodus and drop in AI startups from the campus.

There is no shame in starting a solutions business. Needless to say, an IBM or a Microsoft has huge advantages in terms of their customer relationships, existing infrastructure and depth of bench.

A small solutions business faces many other inherent challenges as well. Still, there is some value to differentiated IP with startups. VCs are also showing some signs of being open to funding solutions companies, however they are still looking for infinitely scalable product ideas.

If you stick to products, there are only three AI startup archetypes that scale like a SaaS company.

The Ball Bearing (Fungible AI Problems)

Almost every machine — the physical kind — uses some ball bearings. They come in various sizes and specifications, but the basic idea is to connect two parts that rotate with respect to each other.

The engineer designing such a machine usually does not worry about ball bearing at all. They just assume that the right type is available when they need it.

Image by author
Image by author

The Ball Bearing Archetype of AI startups is to solve a small problem: 1. that everyone has, and 2. whose solution can be trained using publicly available data. Examples — processing receivables, analyzing contract, marketing analytics, customer review processing, etc.

The essential characteristics of how a customer will use this, and hence the AI behind it, does not change from customer to customer.

You can start by proving your model with data that does not belong to any customer. At a certain level of accuracy, you can attract more customers to share their data with you. Even if they don’t the model works just fine. Lookup the ‘Virtuous Cycle’ in this article to understand this better.

The Handbook (AI Enabled Data Products)

While ball bearings, nuts and bolts are standardized parts for most machines, gears are a completely different story. An engineer would likely design their own gears. They would do many calculations and optimizations to do so.

They would also rely on other engineers’ knowledge typically published in papers, manuals or handbooks.

The Handbook Archetype of AI startups is to use AI to create data products. Your aim is to help the engineer figure out their gears quickly by analyzing all available data about gears and packaging them in a useful format.

For example, the weather service, healthcare insights, predictive maintenance, etc. You control the AI, you get the input data, the customers only pay you for the output.

The Wrench (AI Tooling)

To continue the engineering metaphor, no matter what our engineer is doing, they need a wrench to do it. They don’t think twice about the wrench, yet pay a lot for it and other tools like it.

The Wrench Archetype of AI startups is to create tools so that others can build AI solutions. Think of the various ML Ops tools, work benches, feature libraries and SDKs that help in creation, production and operation of AI/ ML.

You will still ride the AI wave — I spoke about how everyone is trying to create their own AI solutions earlier. You still have to understand the concepts deeply. But you won’t be selling AI. Almost all of these tools are barely intelligent, which means they don’t need customization and can scale well.

If your idea does not fit these three archetypes, it doesn’t mean it is not valuable. I hope you use this article only for a perspective. In the end, you should follow your passion and you will find a way to make it work.

Upvote


user
Created by

Praful Krishna


people
Post

Upvote

Downvote

Comment

Bookmark

Share


Related Articles