AI in Direct Mail Marketing
"By the 1950s, we had a generation of scientists, mathematicians, and philosophers with the concept of artificial intelligence (or AI) culturally assimilated in their minds," Rockwell Anyoha wrote in "The History of Artificial Intelligence," in Harvard University blog in 2017.
The concept of AI goes back more than 70 years. It is considered a general term for complex computing. In 2017, Gartner Group predicted that 87 percent of AI projects would never generate any ROI. Their estimates have continued to suggest 50 percent failure in the report, Gartner Predicts Half of Finance AI Projects Will Be Delayed or Cancelled By 2024. (Gartner Predicts Half of Finance AI Projects Will Be Delayed or Cancelled By 2024)
In the early '90's, we landed Moore Business Products as our first database client. The project was to take customer history data and make it more accessible to the marketing department. They were able to see customer counts for people buying business forms and out-sourced products like furniture and other office supplies.
"What good is this if we cannot pull names for contact?" Though we'd more than satisfied the system specification, we also realized that this was an important question that needed attention. Soon, our query database also included a list selection system which would grow in capability with each new client.
In 1995, we were asked to develop a customer ranking/selection model for The Hudson's Bay National Credit Office. There were huge challenges in trying to load and utilize 250,000,000 transactions running only on a Pentium processor. The largest hard drives were just one gig. At the time, best practice in customer selection was to build one grading system to differentiate best, average, and worst customers. I pointed out to John Travis, marketing director, that HBC credit card could be used to buy general merchandise but also gas and car insurance, travel packages and other affiliate products. "Wouldn't you want different rankings depending on what you were offering?" I suggested.
Our customer query database was crudely connected to a modeling system. HBC IT accused us of violating the specifications by including a modeling system (rather than just building a simple customer score).
Next, we attracted the attention of Cabela's. Our first real catalog company interested in more advanced customer selection. We had already connected SPSS CHAID with query software to produce modeling extracts which could use past mailing list history and subsequent purchases. We were able to build a fishing model a few weeks before it was due, but the results caused a huge crisis.
In a normal customer ranking, every model would suggest that those who had purchased most recently were the best bet to select again. In our first attempt, recency was nowhere to be seen. Turned out, due to the highly seasonal products, hunters were the most recent, but most were not fishermen. Cabela's told us that our model produced $2.3 million more profit than they expected. A later, more carefully crafted test saw us beat their segmentation by 73 percent profit/piece.
Almost 30 years ago, we found a way to build machine learning (a more specific form of AI) and achieve ROI of millions of dollars per use.
Five years ago, USPS paid me almost $20,000 to inspire their sales force with a speech touting direct mail. I received a standing ovation. "Last night I told my wife I was going to take retirement, but after your speech, I know I have something to say to the millennial ad managers I face every day." (This was said immediately after my speech.)
I know direct mail works because I've participated in dozens of hold-out tests comparing customer spending between those who were mailed and customers who were not. About this time, the Gartner AI failure study was published. What puzzled me was how we could achieve massive success with machine learning while others saw only failure?
I spent almost three years pondering this question. I spent considerable time learning about how AI/machine learning was being implemented, what it was being applied to and why it was not working. My research suggested that the failure rate was probably much higher than the 87 percent when applied to marketing. I discovered that both mass and digital marketing did not really know who their ads were presented to. They also did not know who might have seen them but ignored them. (Admit it, we all ignore them as much as possible.) They also rarely could connect even the interested people who perhaps took the trouble to visit their website with the actual purchase they made. Attribution, even for people who actually buy something, all the way from advertising to purchase is almost impossible.
Andrew Willshire wrote in his article, "Attribution Is Broken, Here's How to Fix It," that "most brands should forget about individual-level attribution and focus on the aggregate response. Chasing individuals around the internet produces more data than can be managed, yet not enough to solve the problem."
The failure rate was beginning to come into focus...
The answer shocked me...
The reason we could do machine learning was firmly connected to the medium of direct mail marketing! Why?
In direct mail, we know who we are going to expose our advertising to, we select their individual address based on association, demographics, and purchase history. We start with a mailing is delivered list which has specific characteristics and count.
We also know that our mailing is delivered nearly 100 percent of the time. USPS now actually offers individual household delivery data back to mailers (called Informed Visibility). We also know that it goes to a decision-maker within a household. Even with four kids, we did not let our children bring home the mail and we certainly did not let them decide what would be kept or thrown away. Finally, we know that direct mail forces a decision to be made. Direct mail does not throw itself away, the decision-maker is required to choose between trash, purchase or set aside for later.
You might think this is hardly serious engagement, but I suggest this is the highest level by non-buyers by any other media. The TV is on, but you go get a beer, you drive by the billboard looking at your speedometer, the ad appears on your monitor but you have learned to ignore sidebars, or the ad appears in your social media feed and you simply thoughtlessly swipe it away. All other media advertising literally throws itself away when you ignore it.
What does all this have to do with AI? When you build a model, you need to train the system. The computer compares data and finds important differences. You've probably helped build the "where is the stoplight" data. In the picture, a human will see stoplights in the upper left, they do not have lights (so a machine likely would not recognize them), but a human knows better from experience. If your self-driven car were pulling up closely behind the truck, the stoplights ahead would be blocked. The upper right, backwards lights could still alert your car that a controlled intersection was ahead.
When IBM wanted to teach Watson to play Jeopardy, they had a labeled dataset of 127,000 previous answers with which to evaluate their algorithms. Watson's performance was compared against the right but more importantly the wrong answers.
IBM had worked for years trying to figure out Dragon Naturally Speaking speech recognition software. Remember reading paragraphs into your computer microphone? Google got it working much better in just a couple years. Rather than trying to figure out a person's individual speech patterns, Google collected the corrections. When a person stopped to fix their text message by hand, this gave Google insight into what needed to be fixed in their speech-to-text system.
And now we have the answer to why we could make millions of dollars for clients while universities and computer giants could not. The knowledge of which households have seen our message plus the engagement level of direct mail has a built-in labeled data set. Not only are the buyers known but also those households who saw, engaged, and decided not to buy. No other media can claim to know the engaged non-buyers. No other media can generate this labeled data set and I firmly believe that no other media can achieve this level of AI.
Direct may be slow, expensive, difficult but it is the most cutting-edge medium for discovering who your real customers are and what makes them buy. Before you spend millions or tens of millions to explore AI, understand that without direct mail, it is nearly impossible to achieve ROI.
The views and opinions expressed are solely those of the contributor and do not necessarily reflect the official position of the ANA or imply endorsement from the ANA.
John Miglautsch is president of Wisconsin Direct Marketing Association.