Next Best Action Marketing: How to Implement Hyper-personalization with Machine Learning

Reading time: 9 minutes

These days, people don’t like to be overwhelmed by information. They don’t notice or report online ads and posts, unsubscribe from newsletters, reject calls, and mute app notifications. Or simply forget to read bookmarked posts because…there is just too much content on the internet.

As described by marketing consultant Mark Schaefer, content shock  develops when humans can no longer consume an ever-growing amount of content.

And he’s not the only one talking about it. Other specialists are mentioning it too. Suneel Grover from SAS notes that it’s “important to objectively realize that your brand or product isn’t the center of your consumer’s world.” Most marketing-centric content is perceived as an interruption to a customer-oriented experience.

This changes the game for marketers. Businesses now must engage with customers on their terms: through preferred channels and at a preferred time, with offers that speak to their hearts. The shifting attention from a product to a consumer is the central idea of the next best action strategy. 

https://lh4.googleusercontent.com/EI8Ua3bLBlv8wGrc595ixaKZKIMYy_Mnx7ibY8uGgkjXvnpYvB5_5QZC2wEWvU06WNYRbrs0Q5-0F3Z55AHOjNWYjdNG7NQmufBy-WNkY6axwSukaHp9i4uJ1KA6bdhMGN2aC0xu

Next best action, or Consumer-centric approach to marketing. Source: Deloitte

But how do we know which customer to reach out to, and the right moment to do so?

In this article, we’ll discuss what the next best action strategy is and how businesses define the next best action using machine learning-based recommender systems. We also contacted data scientists working at startups, financial services, and EdTech companies to discuss how machine learning can provide the knowhow to make customer interactions lucrative for both parties.

The customer comes first: the next best action paradigm

The path a prospect takes before making a purchase is called a marketing funnel. This journey contains various steps – touchpoints, aka avenues of interaction between a brand and a customer. Interactions may take place in different environments – channels. The funnel for each customer is unique as each customer learns about a company or its services at their own pace and in their own style.

A customer may open the About us section on a website, add items to a cart, read a case study, contact customer support, subscribe to a newsletter, etc. Actions from the company side may include sending an email with hot deals or a coupon, calling a customer, showing a pop-up message on a website page, polling on a social media page. Whether a customer responds to these actions and goes down the funnel or rejects them with irritation depends on how the company ascertains their needs.

The next best action (NBA) marketing strategy aims at finding the optimal action a company must take during a customer interaction that will unobtrusively and smoothly lead a certain sale prospect to purchase. For instance, a user starts with the section showcasing sneakers in a mobile app, then reads reviews, bookmarks a few models, adds two pairs to a cart, and abandons it. The best next action may be to send a notification with a promo code for forgotten items. Another task is to define when to send the notification and how to not make people feel like they’re being chased.

Potential actions for a company may include:

  • offer to register and get a coupon,
  • subscribe to a newsletter,
  • open a live chat on a website page
  • send an SMS,
  • show a personalized banner,
  • call a customer,
  • not bother, or
  • send a customized email.

The next best action is the touchpoint with the highest probability of sale, as DataRobot accurately points out.

https://lh6.googleusercontent.com/QSpDLOBJ1nNFR4Rf3BfvJBnr2b3OOApwa9J4ZqWlD0_d7JZzPAzPYLQFh4WTFxTLMBoDoeyMeu1_Z-GoROxkUlNmiFGvWpiVV-2ncv5aczZzvX_es5efU_u0_sQvx0W9SCW6Lpqq

Predicting the probability of sale across touchpoints. Source: DataRobot

The strategy works well for both inbound (apps, websites, call centers) and outbound (direct mail, email, SMS) communications.

This next best action paradigm encompasses the next best offer whose main goal is to define which new product would delight existing customers and recommend that product. With the next best offer, a company gives a customer a good idea of what else to buy, and most importantly, does it at the right time.

Both products and services can be recommended in terms of the next best offer:

  • an upgraded subscription plan at a special price for a given period,
  • accessories to fit a chosen outfit,
  • another comedy series based on one previously watched, or
  • transfer from an airport once a user booked a flight.

NGDATA cites data and analytics expert Frank Bria, who highlights that the next best offer adds the most value to the customer – as measured by an increase in their overall customer lifetime value.

So, the next best action is about finding ways of communicating with a particular customer to increase the probability of a sale, while the next best offer focuses solely on personalized product recommendation. And the right channel and the right timing are the requirements for making these decisions. The strategy also allows brands to reduce the chances of mistakenly targeting the same user via several channels.

Types of the next best action strategy systems

You can choose from two approaches for enabling the next best action: rule-based or machine learning-based recommendations.

Rule-based recommendations

The first, conventional way of providing recommendations is to rely on business rules. A rule-based recommender system consists of a knowledge base with rules created considering domain expert knowledge and an inference engine. Rules are written as “if-then” statements: if a customer adds nubuck shoes to a cart, offer shoe dye; if a customer subscribed to a newsletter, send a welcome letter with a 15 percent discount for the first purchase. Once an interference engine detects one of the scenarios/facts described in the knowledge base, it defines whether the situation is associated with a given rule, and the rule is executed.

Such a system won’t respond to unknown scenarios. Users will have to add new rules to optimize a system and keep it up-to-date. The need to manually set and edit rules is the main drawback of these systems. Also, recommendations can be generic since they’re based on assumptions.

Machine learning can help to overcome these limitations.

ML-based recommendations 

Machine learning is about creating algorithms that can find hidden patterns in data and learning from it without explicit programming. The ML-based recommender system will assign probabilities of sale for different touchpoints based on discovered subtle patterns of customer behavior or for specific products.

“Machine learning increases your chance to better discover customer behavior. From the movie The Great Hack on voter profiles to selecting movie recommendations from Netflix and TED, to what coupons a Whole Foods or Walmart should offer, machine learning can uncover hidden insights on your customer profiles,” weighs in David Yakobovitch, principal data scientist at Galvanize and host of HumAIn podcast.

That way, ML allows businesses to provide more granular recommendations based on each user’s preferences and history of interaction with a brand.

Customer behavior is never stable: A person might no longer prefer the products or communication options that they did yesterday. So, the system must adapt to these changes to continue making accurate predictions.

Self-calibrating ML models enable recommender systems to do exactly that. Peter Song, a machine learning engineer at a fintech company and a blogger, emphasizes the flexibility and ease of implementation of ML-based recommendation systems. “The ML model can be changed as you feed fresh data over time. This flexibility will let the model pick up new behaviors and provide suitable recommendations for individuals or teams.” Peter also notes that implementing the ML system can be easier than using traditional code-based logic: “It will be challenging to establish rules and write codes accordingly.”

Types of recommender systems

Another task is to choose the type of analytical system that will define how your company will be acting upon knowledge gained.

A solution may give recommendations for the internal crew or to the user:

Decision-support systems

In this case, employees rely on knowledge from internal solutions and choose the best strategies for communicating with customers.

Chief marketing and analytics officer at Unify Consulting Dave Albano in his tells about a case from his practice. It can serve as an example of how analytical insights can be embedded in employees’ workflow: “A large technology company client that I work with built a planning tool that guided their sellers’ daily activities by presenting a prioritized action list based on AI-generated recommendations. This ‘smart assistant’ prepares sellers for a more targeted engagement based on their customers’ likelihood of purchase. These recommendations save them time and increase customer purchase and renewal activity.

The ML model for the recommender system Peter Song is working on will power the solution that will help marketers select the best-performing affiliate partners: “The Google Analytics data fed into the ML model knows which affiliate partners customers came from and if they led to a conversion. This solution will allow employees to forecast best-performing communication channels or affiliate partners.”

Automated recommendations

That’s the kind of functionality all of us have experienced while shopping or searching for products and services on the Internet. Recommender systems “on the customer side” are commonly used by retailers, fashion brands, online travel agencies, streaming services, media, and others.

For instance, German retailer Conrad Electronic uses SAS customer intelligence solutions with AI and ML capabilities to score online customers based on their current and historical behavior to learn their preferences and show offers they are likely to accept.

Conversion score, calculated for every website visitor in real time, defines their probability to purchase. Customer transaction history and current browsing behavior are taken into account.  Based on this score and other factors, a relevant product recommendation appears on the website, convincing users to complete the transaction.

“If the customer gets distracted and abandons their cart, SAS Marketing Automation then automatically initiates a trigger campaign based on their profile, generating a push notification on their phone and an email with a coupon later that day,” explains the vendor.

Only identified customers can see personalized banners, and they make only two or three percent of all visitors. That’s why the provider applies neural networks that analyze real-time data (i.e., clickstream behavior, length of stay on particular pages) that now can identify 60 percent of the retailer’s visitors.

So, how can businesses prepare for implementing data-driven next best action/offer strategy?

Data for analysis

Since the next best action is about having mutually beneficial consumer-brand relationships, companies must harness internal data about every customer that can be:

Demographic data like age, gender, education, income, or employment.

Online behavior data: clickstream, time spent on given pages, past purchases, service history, payment methods, preferred channels, and many more.

Metadata: devices, browsers, resolutions that were used. For instance, you may find that mobile visits outweigh desktop ones in terms of small purchases, while more expensive products are usually bought via desktops. This can alter recommendations based on metadata.

Social-economic, market, and sales data. This data gives a broader context into factors that may influence what products people buy, in what quantities, and why they buy or don’t buy them. Historical sales data for a specific period of time may help reveal sales trends. Businesses may use information about competitors’ offers, upcoming events, or weather conditions to create deals for products or services that customers are likely to enjoy.

CEO of Chain of Demand AJ Mak notes that their forecasting solution focuses on a multitude of external and internal data points to provide item recommendations: “This can range depending on the business’ needs, but in the past, we have pulled from historical sales data, consumer price index, weather data, market and competitor data, product attributes, styles, and color.” The company provides predictive analytics solutions for retailers and brands.

Traffic source data showing how people landed on a website can come in handy as well. Peter Song is designing a recommendation system to predict which channels work better in terms of a loan application. “For this, I plan to build a machine learning model using the traffic source data affiliated vendors, ads, or organic Google traffic. We collected Google Analytics data for the past two years since that was when we started seeing meaningful traffic amount. That is also when our website was stabilized,” says the specialist.

Finally, once you are sure that you have enough relevant data to rely on and apply data science to, you must understand an implementation strategy for your initiative and justify ROI. Explaining how exactly data science projects unfold is beyond the scope of this article. You may read about it in our ML-focused whitepaper. But we’ll outline the key directions you may consider.

Implementation approaches to the next best action strategy

MLaaS. You can use machine learning as a service (MLaaS) cloud-based platforms that automate data preparation, model training, model evaluation, with further prediction. MLaaS solutions are designed for data scientists of different skill levels and developers, allowing them to solve complex and typical problems. Prediction results can be delivered via REST APIs. Still, this option will work for those who have software engineers and data scientists onboard.

Using off-the-shelf solutions. Another way is to try automated ML platforms. DataRobot, for instance, provides a platform that allows business users without ML expertise (software engineers or data analysts) to easily build ML models. AutoML platforms usually allow for completing relatively simple and custom tasks.

Developing a custom system. This approach would be reasonable if out-of-the-box solutions lack the necessary features. You’ll need to collect relevant data, dedicate some time for data quality and dataset preparation, build a predictive ML model, and either integrate it into an existing solution or into the one developed specifically for a given purpose. As a result, you might get a tool designed for your data and workload. If you want a custom solution but don’t have skilled employees to do that, consider hiring a data science team.

Conclusion

Accurate, personalized recommendations can significantly drive sales while helping companies build their customer fanbase. It’s worth mentioning Netflix to show what these systems can do. Its recommendation system influences choice for about 80 percent of hours streamed.

But having a recommender system isn’t enough. In the first place, adjust the workflow of your customer-facing teams and shift to a customer-centric mentality in the organization.

Add a comment

Comments

avatar