A Look Back - and Forward - At Personalization at Scale

I had an email from a former colleague a couple of weeks ago, in which he shared some thoughts on what he and his firm - an e-tailer - are looking for from the use of artificial intelligence.  First and foremost, it is improving conversion outcomes. That is the goal of everything they do, which one might expect to be true across all similar businesses!  Current uses of AI for him are primarily content applications and offer targeting.  But he considers those to only provide the potential for minimal incremental improvements. The holy grail for him is scalable personalization, both for website visitors and through email and other outbound contacts. That was a really timely message, because I had just finished writing up this recent post where I touched on just that topic.

"There is no way my team can scale email or marketing content that is personalized even at a geographic level--we are ... incredibly lean. So, how can AI assist in development of collateral, identification of cross sell opportunities, and new customer acquisition, mostly infused in PPC and organic campaigns."

That took me back. My interest in database marketing and customer data analytics was driven in part by the 1993 book "The One-to-One Future" by Don Peppers and Martha Rogers. In that book, and in a 1999 HBR article titled "Is Your Company Ready for One-to-One Marketing," the authors explored the concept of "mass customization" in some detail.  Mass customization in their view meant individually-tailored goods and services, with the customizations fueled by one-to-one marketing, which meant collecting and leveraging information about individual needs and preferences. Their thesis was that by delivering such customized goods and services, companies would gain increased customer loyalty and long-term value.  Interestingly, they don't really discuss using the information gathered to tailor subsequent marketing contacts. But it is clearly an unaddressed aspect of their "learning relationship" - which makes sense since this was written at a time when the Internet as a commercial platform was still in its nascent stages, and before recent leaps in computing performance and the advent of cloud computing. Around the same time, in the late 90s, I worked for Bob Kestnbaum, who was one of the pioneers of direct marketing, along with other luminaries like Lester Wunderman and Bob Stone (in fact, legend around the office was that Bob Kestnbaum authored the mathematics chapters of Bob Stone's seminal book on direct marketing, "Successful Direct Marketing Methods," although I never verified that...).  Bob often talked about tailoring marketing contacts based on individual information - a frequent metaphor he used involved having a shelf full of various forms of pre-packaged marketing content and an automated system that would select and present the appropriate package to each addressable consumer, utilizing a segment assignment model or a lead scoring algorithm.  Bob thought you would never be able to get to the point where it would be cost-effective to deliver truly one-to-one personalized messaging, but that by aggregating similar "ones" into larger groupings, you could get close enough to drive substantial incremental value. Fast-forward to 2016, and after relentless efforts to individualize online marketing, Marc Pritchard, CMO of Procter & Gamble, announced the company was scaling back individual-specific targeting on Facebook, saying that "We targeted too much, and we went too narrow... now we're looking at: What is the best way to get the most reach but also the right precision?"  Bob Kestnbaum knew of what he spoke...  In a Wall Street Journal article discussing the announcement, Peter Daboll, chief executive of Ace Metrix, was quoted as saying, "If you could run an ad and reach a million people or run a targeted ad to reach 5,000, you have to have pretty impressive returns on that 5,000 to make it worth it." That's really comparing apples and oranges.  More appropriate questions might be: would the return from targeting 1MM people with a single message exceed the return from targeting 20 different groups of 50,000, each with content tailored to that group, given the cost to develop 20 versions of content? Or 200 groups of 5,000? What is the breakeven point at which it is worth it developing multiple content versions? And that leads us back to the potential of artificial intelligence to help in this area. Can AI really deliver on the promise of personalization at scale cost-effectively, as my former colleague desires?   Certainly Generative AI will be able to help with delivering many creative versions with speed and scale at low cost, but that alone won't lead to incremental returns.  It will take additional work in the areas of data tagging, predictive analytics and especially machine learning to determine the right type of messaging to spur changed behaviors, then to assess whether a segment of one or a segment of one million is optimal, and finally to automate the process of delivery. The answer to the earlier question is Yes, AI will eventually be able to deliver personalization at scale in a way that optimizes returns, it will just take time for all the pieces to fall into place in order to get there. The "learning system" envisioned by Peppers & Rodgers, with a bit of a different spin, is coming. And all those visionaries from the 90s and earlier will look especially prescient when it comes to pass.  Now to get to to work!