How Artificial Intelligence is Boosting Direct Mail Response Rates

18 June 2020 / By Jeff Tarran
Reading time: 4 minutes

Boost Response Up To +15%

Direct mail was the first mass direct marketing channel. The “traditional media” designation should not lead you to believe that there have not been significant improvements to deliver better leads more efficiently and effectively over the last few years.

Those improvements include practically every aspect of a mailing: digital production that integrates variable data on a per-piece basis; incorporating intent data and triggers to contextualize messages and offers; prospect modeling based on social media activity; CTAs integrating smart speakers… and now artificial intelligence.

Initial tests by our partner modeling teams indicate that response rates can improve +10–15% over other practiced modeling techniques through the implementation of AI when applied to customer modeling for audience selection. This article will provide insights into how we are using artificial intelligence to enhance and improve model building with our clients.

 

An Introduction to Model Building—In Laymen’s Terms

Modeling has always been a mixture of art and science. It starts with a client providing a list of customers. It could be a complete client list or a specific set of clients — longest tenure, most profitable, newest, etc.—depending on the parameters of the project.

Before modeling begins, a random sample of the customer data is pulled out of the client customer file and will serve as a holdout or “validation” file. Model accuracy is validated using “split validation.” Literally, splitting the data into two parts, one to perform the analysis and one to test the analysis against. More on that later.

The art part comes from the modelers understanding of the target and the campaign objectives. They use that information to guide the model to make sure targeting requirements are not overlooked. For example, a home equity product can only be targeted to homeowners, so model builders make sure the model only outputs homeowners.

The science part starts with matching the client’s database to a large, national, compiled, or transactional database. A modeling technique, often neural net, provides insight into the characteristics that make up best prospects and, importantly, how those characteristics relate to each other as predictors of behavior. The “model” is the algorithm that describes those characteristics in numerical terms.

The model is then applied to the entire prospect database, ranking every prospect record from the most likely to the least likely match to the model algorithm ideal.

Before we spend money mailing to the prospect list based on the model, the model builders perform an internal test—split validation—by matching records pulled from the model against the data held out of the modeling process (the validation file pulled from the client database as described above.) The more matches to that validation file, the better the model.

 

The Modeling Process: Traditional vs. AI.

Traditionally, the model builder will manually build a few different models based on tweaks to the upfront targeting instructions. After running the models and determining the match rate with the validation file, they will further analyze the quality of the matches and may elect to revise the model requirements and rerun the process. Model builders will use a variety of approaches, techniques, combinations of techniques and pre-selections on universes in order to find the optimal model.

Artificial intelligence takes that process to light speed. Using AI, thousands (or more) models are run at the same time and the process applies its own minor tweaks to assess far more variations. AI is also used to rapidly assess the performance of models using split validation, so the process incorporates selecting the best model as well.

Human intervention is still valuable. Living, breathing model builders still do sense-checks of the quality of the data upfront, pre-selections that fit the client’s business model and outcome needs, and then an evaluation of the best fit model derived by AI. But AI allows them to analyze and produce models at a much higher velocity with more time spent on optimization and less time spent on traditional busy work.

 

A Win-Win—Faster Modeling Leads to Better Models

Model building the traditional way typically takes 2–3 weeks from when data is received. When AI is deployed it can take less than half the time to do a better job. We are telling our clients a week, but in some cases, we are seeing AI modeling results in as little as 3 days.

Faster model building also means more opportunity to optimize. According to Alexa Sundberg, VP of Data at Gunderson Direct, “If you’ve ever commissioned model building for marketing efforts, you can appreciate how AI model building velocity allows more flexibility to change assumptions and parameters, or develop hybrids of best models. Instead of working toward a hard stop deadline, there is the luxury of working towards the best optimization in a reasonable time period.”

 

What’s Next?

Right now, a limited number of database vendors are able to apply artificial intelligence to their modeling efforts. It takes learning and ramp-up time to implement AI, both to adopt new technology and to train the analysts and model builders. In this case, the benefits are clear so we expect AI modeling will probably be the norm in the not distant future.

Is there AI modeling in your future? Please drop us a line if you’d like to learn more about how we can help your direct mail efforts.

Jeff Tarran
About The Author

Jeff Tarran

As VP of Account Services, Jeff works with our clients to analyze business problems and develop direct marketing strategies that achieve their goals. A 20-year veteran and strategic thought leader in direct marketing, Jeff has headed two independent direct response agencies in the Bay Area after starting his career at Foote Cone and Belding. He earned a Dual BS in management and communications at Syracuse University and his MBA in marketing at Columbia University in New York.

.