The Marketer’s Guide to Creating CRM Audiences: The Benefits of Plug-and-Play Deep Learning Solutions

Have you ever questioned your team’s CRM audience creation process? Have you ever wondered if your audience creation criteria were specific enough? Or how it could be so time-consuming and tedious to generate your segments? Or perhaps you’ve simply asked yourself: in 2021, is there really no better way to do this?

Or perhaps you currently only send out mass newsletters, but are planning to adopt a more targeted approach soon? This project raises a dozen or so questions, including data processing, audience creation, and internal resources…

Or maybe you’re just curious to learn how AI works in the field of B2C CRM marketing?

Whatever the reasons, I’m here to help. In this illustrated guide, I’ll explain, step by step, the limitations of traditional audience-building methods and show you why deep learning is the only way to effectively build a CRM audience (as well as a few other practical applications in the field of CRM marketing!).

But first, a quick note: this article is a bit long.

A brief overview of the types of CRM communications

To begin with, it is important to establish a few definitions, as companies tend to use different terms for different types of CRM communications.

CRM communications generally fall into one of these categories:

• Automated messages, or triggers: messages that are automatically triggered by a customer’s actions. A typical example: when a customer abandons their shopping cart
• Sequences, or journeys: automated sequences of messages tied to the customer lifecycle. A typical example: a welcome series for new customers
• The campaigns : One-time messages that are not sent automatically, and which are either:
  • sent to your entire customer list (“batch & blast”)
  • sent to a subset of your customer list ("targeted campaigns").

Today, we’re going to take a closer look at these targeted campaigns. They account for about 90% of the CRM messages sent by a brand. They are used to promote new collections or products, sales, specific product categories, and more…

Generally, CRM teams start by sending batch and blast campaigns to their entire customer base. However, they usually quickly realize the limitations of this type of campaign:

• Batch-and-blast campaigns aren’t very relevant to many—if not most—of your customers, since you’re sending the same content to everyone; therefore…
• You can only send a limited number of these campaigns per week if you don’t want to overwhelm your customers with irrelevant messages and give them a negative experience.

So, naturally, you’ll want to move on to targeted campaigns focused on specific topics or offers to share with a small segment of your customer base that is particularly interested in what you have to offer.

To be successful, targeted campaigns must offer highly relevant content

Why do CRM teams run targeted campaigns? There are many reasons:

• Targeted campaigns will, in theory, be more relevant and thus provide a better customer experience with every message…
• … which will have a positive impact on engagement, conversion rates, and revenue…
• … and reduce the churn rate.
• Targeted campaigns also allow teams to promote a wider range of offers and products, thereby providing greater CRM coverage for their entire product or offer catalog, and satisfying a larger number of stakeholders at the same time—think, for example, of your category managers, who want to feature their products in your newsletters, or partner brands that ask you to promote your partnership to all your customers… Targeted campaigns will thus allow you to send a wider variety of messages to a smaller number of customers. And if the overall relevance of the messages you send is higher, you’ll be able to send more messages to each of your customers.

To achieve your goals, your targeted campaigns must meet one key requirement: they must be as relevant as possible.

And their success hinges on the assumption that your CRM team will be able to identify, for each targeted campaign, the customer segment most interested in your campaign.

But audiences built using traditional methods are rarely relevant

So how do you make a campaign effective? For each targeted campaign, a CRM team must identify the customer segment that will be interested in the campaign’s topic(s), engage with it, and ultimately convert.

Does that sound easy? Think again.

As marketers, we’ve all learned to segment our market and categorize our customers into clearly defined groups. It’s become such a habit that we sometimes forget our customers are real people, in all their complexity, with their desires, their quirks, their ideas, their spontaneity—in short, everything that makes them human.

The best among you might not forget… but what options do you have left?

Traditional segmentation tools are based on assumptions that shape the audiences they create.

In practice, here's what it looks like:

Of course, that’s a bit of an exaggeration. I hope this little comic book campaign will still manage to generate sales. But I’m sure you get my point. The traditional approach to campaign targeting, based on a set of assumptions, has many flaws.

The method includes assumptions—twice over!

The first-party data in your CRM database is undoubtedly very rich. You have so much data on your customers, their behavior, and your products.

And yet, when you build audiences using the method described above, a twofold simplification takes place, one that is generally driven by your assumptions:

1. You need to decide which dimensions are good indicators of purchasing behavior. In the example above, Steve decided that gender, recent purchase history, and location are the most important dimensions. But why? How can you be sure that it wouldn’t be better to focus on age, first name, or browsing history?
2. For each of these dimensions, you need to decide which values to select. In the example above: female, recent purchase of a dress, geographic location. Again, why? Perhaps the most interesting feature of these dresses is that they are strapless. Wouldn’t selecting recent buyers of strapless tops be a better choice? Maybe, maybe not. Who knows?

For these two dimensions and these two values, your parameters will be guided by intuition (at best) or by your assumptions (at worst) and will therefore always be influenced by your personal biases.

So, you can probably guess what I'm going to say next…

The performance won't be exceptional

Because this method is only a very rough attempt to " predict " the effectiveness of your campaigns:

• Some people will be included in the target audience even though the campaign has absolutely nothing to do with them.
• Some people will be excluded from the target audience even though the campaign would have been 100% relevant to them.
• This approach is rarely sophisticated enough to account for seasonality or the product lifecycle. People who pre-order a new video game have a very different profile from those who buy the game several months after its release to give as a Christmas gift. How could a traditional audience-building method take this into account?

You're segmenting your customers instead of promoting cross-selling

Unfortunately, using purchase history to predict future purchases will lock your customers into their initial categories, and you’ll miss the opportunity to show them additional offers—even if those offers might be relevant to them.

The reach of your campaigns can quickly become laughable

One of the pitfalls of adding too many criteria to refine your audience is that it can end up being too small. Marketers often realize that their final audience is too narrow. They then have to make compromises—adding new segments, removing certain criteria—only to end up back where they started: with a broad, poorly defined audience.

You don't predict the likelihood of buying NOW

When you create an audience, it’s not enough to predict its likelihood of purchasing a specific product. You actually need to predict its likelihood of purchasing at the time the campaign is sent.

It’s not enough to say, “Steve has bought a dress in the last six months, so Steve will probably buy another dress someday,” to include Steve in your dress campaign. Over a very long period of time, that statement is likely to be true, but it doesn’t really help with the campaign you need to send out tomorrow.

It would be much more effective to be able to say, “Steve is very likely to buy a dress this week, so let’s include him in tomorrow’s campaign.”

I hope we are now on the same page regarding the limitations of traditional audience-building methods.

Fortunately, there is a solution, and that is, of course,artificial intelligence.

More specifically, deep learning.

Deep learning is THE solution for achieving true CRM relevance, as it can accurately predict the likelihood of a purchase right now

When you think about it, it makes sense:

• CRM marketers have a wealth of first-party data, and artificial intelligence excels at making sense of large volumes of data.
• Our campaign targeting approach shares many similarities with a very popular use case for deep learning: image recognition.

You might be wondering what the connection is between building an audience and image recognition. Here it is:

1. In both cases, we’re trying to answer a simple question… Is this picture of a cat? Is Steve going to book a trip to Thailand this week?
2. In both cases, we have rich and complex data to help us answer the question:
   1. Images are made up of hundreds of thousands of pixels, each of which can take on one of 16,777,216 possible color values (in RGB).
   2. CRM data: between customer attributes, product attributes, and transaction history, you easily have over a hundred data points. Now imagine all the possible combinations… add to that browsing history, campaign open rates, and timing… your CRM data has incredible depth!
3. We could consider focusing on a small amount of data to answer our questions…
4. But we will soon realize that even if we spend a lot of time drawing up a list of criteria, the success rate will remain very low.

5. However, in both cases, it is possible to find a large number of examples where the question has already been answered.

1. For image recognition: there are image datasets that contain cats and are labeled as such, or that do not contain cats and are therefore not labeled.
2. When it comes to predicting purchasing behavior in CRM: combining your customer data, transaction history, and product attributes reveals many instances where your question has already been answered! “Did Pierre buy a dress on December 3? ‘No.’ On December 4? ‘No.’ Did Marie buy a dress on December 3? ‘Yes.’ On December 4? ‘No.’ Did Jennifer…?”

We know that deep learning works surprisingly well for image recognition. With a sufficiently large dataset to train the models, the success rate is close to 100%. For certain specific applications (medical imaging, for example), deep learning models are already more effective than humans!

Deep learning applied to image recognition (and in general) works much like the human brain: just as a child learns to recognize a cat, not by memorizing a list of criteria provided by their parents, but because their parents patiently show them pets and tell them, “This is a cat,” “This is a dog.”

Deep learning applied to CRM works in much the same way. The algorithm will base its learning on your entire database—that treasure trove of first-party data we’ve already discussed—without excluding any data points or getting bogged down by arbitrary rules. It will incorporate all edge cases that your intuitive criteria could never have identified and take them into account to adjust its predictions.

The deep learning algorithm will be able to detect highly nuanced data—such as product lifecycles, seasonality, people’s “styles,” and their “unpredictability” —even though none of these factors are explicitly recorded in your database.

Deep learning can accurately predict that Marie will buy a vintage floral dress, even if she’s never bought a dress from you before—or any“vintage”or“floral”items, for that matter.

Dear readers, please note: not all artificial intelligence systems are created equal

Of course, everything I’ve just mentioned applies only to a well-designed deep learning algorithm. Simply feeding CRM data tables into an image recognition algorithm won’t be enough to generate predictions about purchasing intent.

Similarly, we all know that artificial intelligence and deep learning are buzzwords. So here are a few things to keep in mind when considering how to incorporate artificial intelligence or deep learning into your CRM strategy (!):

• Artificial intelligence is a broad term that can describe any type of automated decision-making process, such as rule-based decision trees—much like the ones I mentioned earlier, which we’ve seen have so many shortcomings! It’s interesting to note that rule-based artificial intelligence (often called “expert systems”) was the dominant form of AI in the 1980s… a period that eventually led to an “AI winter” (a dark age for AI, when enthusiasm for AI waned and investment nearly dried up).
• Machine learning is a much more advanced field of AI. It was research in machine learning that revitalized the field of AI after the second “AI winter” of the late 2000s, which I mentioned just above. Gone are the days of hard-coded, rule-based decision-making. In short, machine learning develops its own predictive rules based on patterns it observes in the data it analyzes. As I mentioned earlier, this approach is essential when it comes to processing CRM data and predicting a customer’s likelihood to buy! However, machine learning still has a major drawback: it requires human intervention to ensure that the data is properly structured and categorized… And of course, it’s difficult to ensure that in a CRM context.
• Deep learning is an even more powerful technique derived from machine learning. At its core, deep learning mimics the way our brains work. Deep learning’s “artificial neural networks” can learn from any type of data, with minimal human intervention, to achieve exceptional results. These networks are designed to make sense of data sets on their own and learn independently based on that!

Bingo! It seems that deep learning is the most suitable form of artificial intelligence for predicting purchase intent.

There are still a few key questions regarding deep learning:

• What datasets are used?
• How do algorithms handle incorrect or incomplete data?
• In a dataset, which data points do algorithms learn from? Do data points from five years ago carry the same weight as the most recent ones? How does this work in practice?
• What goal do algorithms tend to optimize for?
• Is the technology resource-efficient? Will it allow you to assess your customer base in a day or in a minute?

As is often the case, the true value of an algorithm lies in its application and execution. That is why manydeep learningmodels are available for free or sold commercially. While these serve as excellent foundations for your algorithms, they won’t get you very far in your day-to-day work.

That’s also why, in general, you should be wary of AI, machine learning, or deep learning applications that weren’t developed for a specific purpose, but rather for a wide variety of use cases. Einstein, Watson—do those names ring a bell?

Not all AI or deep learning models are created equal.

So how can you achieve truly exceptional results?

For deep learning to be successfully applied to building CRM audiences, it must follow certain basic principles:

1. Accuracy: Results should be as precise as possible
2. Practicality: The setup should be as simple as possible

Without these two qualities, an app would never be used.

Here are some key points on the effective use of deep learning in CRM:

• The app must solve real-world problems.
• The application must be tailored to the work methods of a CRM team.
• The application should be easy for a CRM marketer to use, without requiring any data science skills.

All of these points seem pretty obvious, which is exactly why these applications are called “solutions” and not “problems”! But you might be surprised by how many “solutions” don’t meet all three of these criteria.

***

That being said, let's get down to business.

The algorithm must be able to use all available first-party data, regardless of its state at a given point in time, even if it is unordered and decentralized.

• Practicality: There’s no such thing as perfect data. Most companies have already started looking for a CDP, but the process often takes time… Do they really have time to wait? And what might they discover once their CDP is finally ready?

That their data is still not“clean”: consistent, organized, and usable.

No, there is no such thing as“clean”data.

We have to live with it and still deliver the results that are expected of us. The good news is that this is entirely achievable if the model was developed with this constraint in mind.

• Effectiveness: Data matters. Data is a signal. Even messy data. Let deep learning figure things out on its own and determine what’s important and what’s just noise. That’s its job, after all.

The algorithm should not require any rules or criteria to function.

• Effectiveness: These rules would run counter to the very principle behind using a deep learning application. It would amount to reintroducing bias into the predictions. Remember the famous “AI winter” I mentioned earlier…
• Practicality: Developing rules takes time!

The algorithm should not start from scratch.

It’s pretty obvious, both in terms of effectiveness and practicality. Marketers want to see results from day one—that’s non-negotiable. Deep learning models must therefore be able to learn from historical data and have high predictive power right from the start. (To learn more, read the article by one of our data scientists—it’s a bit more technical than what you’re reading now, so you’ve been warned!).

Seasonality, trends, and changes in consumer behavior must be reflected in the algorithm’s predictions.

It’s mainly a matter of effectiveness: if these trends and behaviors aren’t reflected in the predictions, the results will be mediocre at best. Shopping habits are very different during the holiday season and in early spring, for example!

To better account for seasonality, the model must be calibrated to give special weight to the most recent purchases of the product for which it is predicting purchase likelihood!

The algorithm should be optimized based on your preferred success metric

Revenue, in most cases, rather than open rates, click-through rates, etc.

The algorithm must be able to make accurate predictions even for customers for whom there is no recent data.

In any given week, perhaps 2% of your customers will visit your website or make a purchase, generating new data points.

And over the course of 3 or 6 months, maybe you’ll reach 20 or 30% of them?

But what about everyone else—the vast majority of your customers—for whom you don’t have recent data? You’d still like to be able to send them relevant communications. They may even be inactive, making their reactivation all the more necessary. This is the shortcoming of many personalization or product recommendation solutions: they work well for customers whose intent data is very recent, but are completely ineffective for everyone else.

And building on all of the above, the algorithm must be capable of making multi-category predictions and going beyond a simple extrapolation of purchase history.

After all, very few of us buy only dresses, only rugs, or only flights to Miami. If Jean has already bought socks from you, are you going to send him only sock-related promotions? If Marie has stayed at your hotel in Bordeaux in the past, is she doomed to receive your “Weekend in Bordeaux” offers forever? What if Marie doesn’t like visiting the same place twice?

***

As we’ve seen, developing a deep learning model designed to predict purchase likelihood based on a CRM dataset raises many questions. In fact, there are so many that it’s never worth it for a company to build its own model in-house: the level of expertise and development time required would likely deter most companies…

However, for a team of data science experts whose sole focus for years has been to create and refine such a model, it is possible! And this research has many practical applications in the field of CRM marketing.

If deep learning can predict purchasing intent, what are its real-world applications?

The practical applications of being able to accurately predict each customer in your database’s likelihood of purchasing a product or offer from your catalog are endless. Here are a few:

Audience building

The first practical application is, of course, building an audience (as you might have guessed, that’s the topic of this article).

Let’s say your deep learning algorithm can rank all your customers from most likely to least likely to purchase a given product. In that case, creating the audience could simply come down to:

1. Select the product(s) or offers featured in your CRM campaign
2. Select the desired number of recipients OR follow the algorithm’s recommendation (for example, send to everyone with an above-average likelihood of purchase).

Managing Marketing Fatigue

CRM teams send out numerous campaigns every week, sometimes several a day. However, they may have a policy stipulating that no customer should receive more than three messages per week, or more than one message per day.

But what happens if a customer is highly likely to purchase every single product featured in the campaigns scheduled for the same day?

Information regarding purchase propensity could be used by the algorithm to assign this customer to the campaign where their purchase propensity will be highest, thereby ensuring maximum relevance of your messages while minimizing customer fatigue.

A multichannel CRM

If you can calculate a customer’s propensity to purchase a product, there’s no reason why you can’t calculate a customer’s propensity to purchase a product through a specific channel.

This adds a whole new dimension to everything we've discussed so far!

Identify untapped sources of purchase intent

There may be a small group of your customers who are particularly fond of a certain product right now. They won’t show up in your key metrics because the product isn’t yet among the “best sellers ”… but there is strong demand concentrated among this small group of people. This is the ideal audience for a targeted CRM campaign!

This information could, for example, be used to create a heat map of your product catalog, allowing you to easily visualize all these opportunities.

CRM Schedule Planning

Should we highlight two products in the same campaign or in two separate campaigns?

The answer depends on whether people interested in one product will also be interested in the other.

Guess how to get that information…

Would you like to learn more about this topic?

If you've made it this far in the article, thank you! I hope you've learned a lot from it.

If you’d like to learn more, feel free to check out all our articles.