A Causal Perspective on Personalized Decision-Making in Business

A large part of how companies use data today is centered around customer experience. Companies want interactions to feel relevant, timely, and seamless. They want customers to see products they are actually interested in, receive messages and offers that are relevant to them, and move through services without unnecessary friction.

To improve customer experiences, companies increasingly rely on personalization. Many customer-facing businesses now use some form of it. Ecommerce platforms rank and recommend products based on browsing behavior and purchase histories. Marketing teams use customer data to decide which users should receive a campaign, which message they should see, and when they should receive it. Subscription businesses use customer data to decide which retention action appears most relevant for each customer, such as sending a reminder, offering support, providing education, making a retention offer, or doing nothing.

All of these are examples of using data for personalized decision-making. The goal is to make interactions more relevant and useful for customers, with the expectation that this will ultimately improve business outcomes we care about, such as conversion, retention, engagement, customer lifetime value, or revenue.

This last part is important. Personalization is often described as showing the right customer the right message, product, offer, or experience at the right time. But underneath that phrase is a decision problem. For each customer, at a given moment, a company has to choose among several possible actions. It could recommend one product rather than another. It could send a message now, send a different message later, make an offer, provide guidance, or do nothing at all.

The real objective is to choose, for each customer, the action that is expected to have the greatest positive impact on the outcome or outcomes we care about. That is what makes personalized decision-making a fundamentally causal problem.

Why predictive modeling is not always enough

Let me step back and explain this better.

In practice, personalization is often approached predictively. By this I mean that companies use historical customer data to predict future behavior and rank customers, products, messages, or offers by expected relevance or response. A recommender system might estimate which products a customer is most likely to click on or purchase. A marketing system might identify customers who have historically been responsive to campaigns, or rank them by their predicted likelihood of opening, clicking, or otherwise engaging with marketing. A retention system might flag customers who look similar to past customers who canceled, or identify customers who look likely to accept a discount.

These predictions can be very useful. They help companies make experiences less generic. They can make search results more relevant, recommendations easier to navigate, and communications more targeted. A customer who has been browsing running shoes probably should not see the same homepage as a customer shopping for office furniture. A new user struggling with onboarding may need a different experience than an experienced user who already understands the product.

But it is important to be clear about what these predictions do and do not tell us.

A model that predicts which product a customer is likely to buy tells us something about the customer’s expected behavior. It does not, by itself, tell us whether recommending that product will increase the probability that the customer buys it.

A model that flags customers who look similar to past customers who canceled tells us who may be at risk. It does not, by itself, tell us whether sending a reminder, offering support, providing education, or making a retention offer would reduce churn for that customer.

A model that identifies customers who are likely to open or click on marketing messages tells us who looks likely to engage with those messages. It does not tell us whether sending the message improves the customer or business outcome we ultimately care about, whether another message would have worked better, or whether contacting the customer was better than not contacting them at all.

This is the bridge that is easy to overlook. We often move from predicting what a customer is likely to do to assuming that acting on that prediction in some way creates the best outcome. But that is a strong assumption, and it is often not justified.

A predictive recommender system may surface the item a customer is most likely to buy or click. However, the item the customer is most likely to buy is not always the item that creates the greatest incremental value from being recommended. Some products do not need to be recommended because the customer would have found them anyway. Other products may have lower baseline purchase probability but higher incremental lift when shown at the right moment.

The same issue appears when companies personalize offers or promotions. A customer may arrive on the website, abandon a cart, or show signs of hesitation. A predictive model may indicate that this customer is likely to redeem a discount or purchase after receiving an offer. However, being likely to use a discount is not the same as needing the discount to make a purchase. Some customers would have purchased anyway, so the discount reduces margin without meaningfully increasing conversion. For others, the discount may not be the right action; a different experience may have created a better outcome.

This distinction matters for business outcomes. If we recommend the product a customer is already most likely to buy, we may use valuable space on something they would have found anyway. If we offer a discount because a customer looks likely to redeem it, we may reduce margin without meaningfully increasing conversion. If we choose a retention action because a customer looks likely to click, accept, or redeem it, we may send messages or offers that create annoyance, cost, or margin loss without actually reducing churn.

A causal perspective on personalization

The type of analysis that directly addresses this is causal analysis. From a causal perspective, personalization is about estimating the effect of different possible experiences. For each customer, there is a set of actions we could take: which product to recommend, which message to send, which offer to make, whether to contact them, when to contact them, or whether to do nothing. The question is which personalized experience changes the business outcome we care about the most compared with the alternatives.

For example, suppose a subscription business wants to use personalized decision-making to improve retention. For a given customer, it could send a product education message, offer support, provide a discount, send a reminder, or send no message at all. From a causal perspective, the question is what would happen under each of these possible actions. Would education increase retention compared with sending no message? Would support create a larger improvement than a discount? Would a discount create enough incremental value to justify the margin loss? Or would any intervention have little effect because the customer would have stayed anyway?

The best action is the one that creates the largest improvement in the outcome we care about compared with the alternatives. For one customer, that may be product education. For another, it may be support. For another, it may be a discount. And for another, the best action may be no intervention at all.

And although I have focused on one example here, the same logic applies to many other personalized decision-making problems. At its core, personalization is often about choosing among possible experiences to improve outcomes for each customer. And whenever we ask which experience we should choose, we are asking a causal question. This is why causal inference matters for personalized decision-making with data in business.