My Algorithm is Better Than Yours

My algorithm is better than yours. My algorithm performs better on the precision/recall tradeoffs. It surfaces fewer false positives. It converges to an answer faster. Perhaps it requires a bit less data. Those statements might all be true. But none of these advantages confer a competitive sales advantage in the market. They aren't technology innovations leading to a go-to-market advantage.

I first observed the use of large scale machine learning at Google. In the early and mid-aughts, the advertising ecosystem bloomed. Hundreds of ad networks vied for publisher ad impressions. Each one promised a better targeting system, new algorithms, unique data, better performance, more revenue. 

How did website publishers respond this rush of competition and overwhelming, undifferentiated marketing? They compared yield. At the beginning, a publisher might benchmark five or six different ad networks. Today, software exists to do this at scale. Supply side platforms select the optimal ad network for this ad unit on this page at precisely this time of day for this user. Rather than trusting an ad networks claims, a publisher verified them. 

In the advertising world, publishers bear almost no switching costs. If a new targeting mechanism arises, just click a button or change a line of Javascript. 

But in the SaaS world, switching costs are real. Customers must wait for contracts to expire. Or pay two competing vendors simultaneously. Cost prohibits A/B testing. 

So does training time. Unlike the advertising world where copious training data can be obtained inexpensively, SaaS products often require integration into customer systems. Customers must exercise patience. And the SaaS startup suffers slower sales cycles. 

In addition, internal dynamics create further friction to testing. Will the VP of Sales instruct her account executives to use two different CRM tools in parallel to see which is better? Hardly. 

The advantage of a better algorithm doesn't appear in the sales process in SaaS for the reasons above. But better ML is still important. An algorithmic advantage accrues to the customer success teams, and reveals itself in the retention metrics.

Once a company has committed to an ML-powered SaaS product, trained the system, tuned it to their liking, switching costs increase. The very same reasons above that slow sales cycles lengthen customer lifetimes. That's the competitive advantage that many SaaS companies will benefit from. 

And if that's the case, fast, expensive, and aggressive customer acquisition will become the dominant strategy in the early days of ML-disruption. The business that can buy the greatest number of annuities the fastest will win. Particularly, if those startups retain more of those recurring revenue dollars with superior math.