Hot Take on NFX's "5 Levels of AI Spectrum"

January 2024

I’m an avid reader of the NFX newsletter – and highly recommend it to founders! I couldn’t agree more with the "AI Litmus test" and having start-ups focusing on finding the “white hot center” of customers to improve their product-market fit.

Yet I disagree with 2 of the “5 Levels of the AI Spectrum", even if I go with the general audience’s take that AI = ChatGPT. Here’s why. 

As someone who helps funds do AI due diligence, has developed, implemented, invented algorithms, and collaborated with 150+ companies of all shapes and sizes, I see AI as a means (e.g., algorithm derived from data) to an end (e.g., a product that brings value to its customers that, ideally, they’re willing to pay for). Can I propose the following five levels instead? Levels that I’d argue can only be ascended sequentially by succeeding in the previous rung – without skipping steps (!).

Level 0: Data Curious

Most traditional businesses are here. Nothing is data-driven, but they’d like to explore a potential use case to see if/what is possible with the currently available data.

Level 1: Data-Enhanced with KPIs/metrics

Most modern companies are here, collecting data but need help with availability/quality/access; those who succeed here can get insights from product and operational metrics in real-time.

The goal here is to capture the time-to-value along with key customer types and their flows to establish a baseline for a company’s product-market fit, as the customer is paying for the value the product brings them, which may or may not be “AI”.

Level 2: AI-Enhanced to Drive Efficiency

Level 2 here is similar to Level 1 of the NFX article, with an added layer of an algorithm that identifies promising, repeatable steps to reduce the time to value for internal/external customers. Most companies can find a lot of low-hanging fruit here, even with if/then statements, which can serve as a baseline algorithm against comparing performance. Decisions are made with the help of a manual algorithm or via Excel/SQL (or similar) formulas. 

The goal here is to reduce friction and the time-to-value for customers to begin moving the needle on improving a company’s product-market fit.

One example is a heuristic look-a-like model to identify the top 5 promising customers to upsell to that week, to help the Sales team prioritize whom to upsell to.

Level 3: AI-Enabled

Level 3 here is similar to Level 2 in the NFX article, with an added layer of making real-time predictions based on real-time customer engagement with the product. The algorithm is a part of the product.

One example is matchmaking in a marketplace when a new customer joins.

Please note: I see most companies try to skip the previous levels, start on this rung, and waste valuable time and resources developing algorithms for integration into a product they have little/no idea how their customers use and how it will bring the customers more value.   

Level 4: AI is the Product

Level 4 here is similar to Level 3 in the NFX article, packaged up as a standalone product/feature to sell to customers that helps customers make decisions in real-time. 

One example is flagging low-quality parts during manufacturing or early diagnosis of a patient’s condition.

Please note: AI will still be the product even if AI models get retrained automatically, though I would argue that if guard rails and systematic validation are missing, it will be the case of "garbage in, garbage out".

Conclusion

While I agree that there will be use cases for AI we haven’t seen yet, the focus should not be solely on generative AI – and we don’t need another level on the AI spectrum.

I imagine a lot of future development in the computer vision/image/video space – and any future development would still fall into the "AI is the Product" category – or not pass the "AI Litmus test". :)

What’s your take on the article?


References: