Debunking the top myths surrounding machine learning models for smart customer engagement
We’ve all heard about machine learning (ML) and its promise to change how brands interact with customers who desire to be engaged in relevant, authentic ways. Defined as a computerized algorithm that learns from past data to predict future outcomes, ML technology is being utilized by retailers, banks, restaurants, and a myriad of other enterprises to optimize the customer experience by enabling personalized engagement at scale. But as ML has advanced, so too have the misconceptions and myths surrounding it.
Myth #1 | Machine learning models can be effective out-of-the-box.
Spurred on by ML technology providers claiming their solution requires no uptime before activation, this misperception plays on an enterprise’s desire for immediate impact and results. While some solution providers claim that their ML model’s exposure to existing data allows it to work immediately for a berth of “similar” enterprise types, this approach fails to take into account how varied purchase patterns are, even in seemingly similar business models.
For instance, take two e-commerce enterprises—one that sells shoes and another that sells tablets. It’s easy to assume that the same predictive behaviors will apply across the board since they are both e-commerce. However, shopping behavior is very different based on the type of product they are seeking and the brand they are interacting with. Shoes may be (or may not be) more of an impulse purchase than a tablet. Tablets may be (or may not be) more of an emotional purchase than a pair of shoes. A customer shopping on a premium brand site may be less price-sensitive than when she shops on a bargain site.
And so, ML models need to be tuned for each business to consider very different predictive behaviors. In order to do so, an ML solution provider must dig into the enterprise’s specific data and customer behavioral patterns to develop a custom model and give the model time to learn and become more accurate. Further still, the solution providerSo should provide visibility into the predictive values considered in order to prove the efficiency and accuracy of the model.
Myth #2: Machine learning can magically solve all of your needs.
Multiple factors contribute to this myth. First, is the tendency of sales representatives to invoke “ML” as the catch-all answer to difficult questions. Second, is the portrayal of AI as human-level general intelligence in futuristic movies and in sensational media.
ML can be extremely powerful for very specific tasks, but, as of 2019, it still requires a great deal of human intervention to be used effectively. For example, ML can accurately predict who is likely to make a purchase on your website. The output of that ML is a propensity score. It takes human intervention to effectively orchestrate the usage of that score to deliver personalized price incentives complete with personalized messaging.
Myth #3 | Machine learning can only map outcomes based on customer-type, not individual customer.
A remnant of traditional marketing practices teaches that the best way to reach target markets is to segment audiences into customer personas (e.g. affluent women in their 30’s with children in the household) and target them as a segment. This practice is still powerful, but it can be dramatically enhanced with ML.
Properly tuned ML models can be used to target activity down to the individual, offering promotions and even individualized price incentives based on the customer’s historical purchase patterns and predictive short-term behaviors. These factors can predict likeliness to bounce, likeliness to purchase within the session, current shopping mode, level of price sensitivity, concerns over shipping costs, confusion about which product to buy, the best product or promotion to offer, and how soon the next repeat visit will occur if no purchase is made. All of these factors can be used to deliver fine-tuned personalized experiences.
Myth #4 | Machine learning is the antithesis of a “humanized” customer experience.
You may subscribe to the belief that the best personalized experiences can only come from fellow humans. I believe this is true today and I believe this will continue to be true for several years to come. But human interaction is limited by manpower, with the delivery of 1:1 experiences contingent on a 1:1 ratio of employee-to-customer.
In past years, this problem has been addressed with segmentation and rules-based marketing algorithms, and as consumers, we have all grown accustomed to “pretty good” personalized experiences that would clearly fall short if compared to direct human interaction.
ML is taking scalable personalization to the next level, and the reason it can work so well is that savvy marketers are using the best of both – human intuition combined with ML driven insights to create scalable interactions that feel (and are) personalized to each customer.
Things to Consider when Evaluating ML Solution Poviders
When seeking out an ML solution provider, it is important to make sure they offer a true ML solution, capable of accurately predicting consumer behaviors to provide lift in sales while improving customer experience. Too often, solution providers will eagerly characterize rules-based algorithms or manual services as ML in a black box. Some solution providers may even try to get away with characterizing vaporware as ML in a black box.
Rule of Thumb: “Black Box” ML solutions are most likely not effective ML. It may be mislabeled or it may be a generic model that was designed for the lowest common denominator. In either case, it probably won’t work for your business as well as you hope. Here are some questions you can ask to give you confidence in their models.
- How effective are your predictions? A strong ML solution provider should be able to demonstrate lift over a control group in an A/B test. This is an important factor, but it should not be the only factor. A “black box” ML solution provider may ask you to trust the model based on past performance with other customers, but that may not apply to you.
- What is the process of planning and building your ML models? The process should include engaging with you to better understand your business so they can custom-tune the model. The process should include iterative evaluation of data inputs to fine-tune the model as well as a process of monitoring effect on model accuracy over the first few weeks of deployment.
- Can you show me predictive values for your models including the number of true positives and false positives? Ask the solution provider to show you a sample accuracy report. Ask them if they will provide on-demand access to this report with your data should you engage with them. Ask them to run a “silent mode” pilot with your data to prove the accuracy of their model in the context of your business before you deploy the models to affect customer experience.
At ZineOne, we custom-build our ML models for each unique enterprise to deliver effective ML excellence to our customers. Discover how our Intelligent Customer Engagement (ICE) hub can transform personalization and engagement in your enterprise.