“AI/ML expert” here. First of all, I always look askance at claims that revolve around a number of data points (“over 1000 signals”). This is marketing copy that tells you very little about their capabilities. You can throw any old data point into a model; it doesn’t mean it’s contributing useful information. At the same time, there’s no standard definition of a “signal”; you can carve up your dataset however you like to create an arbitrarily large number. Likewise, I don’t see Upstart’s head start as that significant. Larger institutions have massive customer bases to build models on.
I do see value in what Upstart (and similarly, Lemonade) are doing. It is a big cultural and process shift to move away from expert-created heuristics like a FICO score to a trained model. It’s not technically that amazing, so legacy institutions could replicate it. The question is how much inertia they would have to overcome to do so. It would surprise me if they would leave significant gains in efficiency on the table, but some places can be very old-school.
This is where I see a strong potential for Upstart (and not for Lemonade unless they pivot). These incumbents will find it much easier to buy these AI-driven efficiency gains than to build them. If Upstart can provide the engine that helps 50% of the country’s lenders write better loans, they could have a very sticky SaaS offering. As long as Lemonade is trying to actually be the insurance provider, I don’t see whatever marginal improvements their AI can make to their profitability as justifying a P/S multiple so far out of line with the industry.
I do believe that these AI approaches are the future of risk assessment in insurance and lending. But we shouldn’t underestimate the extent to which existing institutions with billions of dollars at stake are already trying this, and we shouldn’t overestimate the improvements they bring. If the story for these newcomers is just that the AI leads to better margins, it’s going to take a long time for that advantage to naturally outcompete the incumbents, and it’s likely to be swamped by other factors. Licensing the capabilities is the way to succeed, IMO.