AI is giving marketers new ways to investigate customer behavior, uncover relationships hidden in their data, and challenge existing assumptions. Turning those discoveries into better marketing decisions still requires human judgment, behavioral understanding, and testing.

Years ago, when I was teaching personalization, data mining, and customer behavior, one of the examples I used was the work of U.S. retailer Target. It’s still one of the most fascinating demonstrations of what businesses can discover when they look beyond individual transactions to examine patterns of behavior.

Target wanted to identify women likely to be pregnant, ideally before other retailers did, because pregnancy presents an opportunity to establish new purchasing habits. Its statistician, Andrew Pole, analyzed customer transactions and identified about 25 products that, when considered together, could help predict whether a customer was pregnant and estimate her expected delivery date.

The work became widely known through Charles Duhigg’s 2012 article in The New York Times Magazine, “How Companies Learn Your Secrets.” What made it remarkable wasn’t simply the prediction. It was the analytical sophistication required to uncover relationships that wouldn’t have been obvious from looking at individual purchases.

At the time, this kind of work required specialist expertise, extensive datasets, statistical modeling, and considerable technical resources.

Today, some of that analytical capability is accessible to marketers who’ve never written a line of code.

We’ve been looking at the wrong AI opportunity

Most of the discussion about generative AI in marketing still revolves around content production: writing emails, generating subject lines, creating images, building campaign variations, and speeding up workflows.

There’s nothing wrong with these applications. But they’re largely about executing the marketing we’ve already decided to do.

In my recent MarTech article, “AI made marketing faster than it made marketing better,” I argued that producing marketing more efficiently doesn’t necessarily make it more effective. Faster execution has limited value if the underlying decisions haven’t improved.

What if AI’s greatest contribution to marketing is helping us understand more instead of just producing more? 

Businesses have accumulated years of customer data, including transaction histories, browsing behavior, purchase frequency, product preferences, discount usage, average order values, and engagement signals.

Yet much of our analysis still focuses on reporting what’s already happened. 

• Which campaigns performed best? 
• Which segments generated the most revenue? 
• How many customers converted? 
• What’s our average order value?

These are useful questions, but they’re largely descriptive.

They don’t necessarily tell us why customers behave differently, what relationships might exist between behaviors, or which assumptions we should challenge.

That’s where AI-assisted analysis gets interesting. Instead of looking back at what happened, we could look forward to what all of our data means.

From reporting results to discovering relationships

Imagine an ecommerce marketer looking at customers who bought once but haven’t returned.

The conventional approach is to segment them by recency, product category, or purchase value, then create a second-purchase program.

Perfectly reasonable. But what if the most useful behavioral differences aren’t immediately obvious? Consider three hypotheses that conventional segmentation might overlook:

• Customers who purchase across two categories are more likely to return than those who buy several products from one category. 
• Discount-led first purchases are associated with different repeat-purchase patterns depending on the product category. 
• The time between a customer’s initial browsing session and first purchase is associated with their subsequent value.

With the right data and analytical tools, AI can help marketers explore datasets, identify associations, compare customer groups, write analytical code, and develop hypotheses for further investigation.

Instead of asking only, “How many customers made a second purchase?” we can ask, “What’s different about those who returned compared with those who didn’t?”

Then we can take it even further: “What other differences should we investigate?”

That’s one of the most valuable questions we can ask. Marketers, like everyone else, are influenced by their existing assumptions. We analyze the things we already consider important, using our most familiar segments and metrics.

AI can broaden that exploration, identifying possible relationships we might not think to investigate ourselves. It doesn’t mean the relationships are meaningful. But it gives us additional places to look.

The technical barrier is falling, but the intellectual one still stands 

We need to be careful about these claims. Generative AI hasn’t turned every marketer into a data scientist. Using ChatGPT or Claude to analyze a dataset isn’t the same as developing and validating a sophisticated predictive model.

AI can misinterpret fields, generate incorrect calculations, overlook confounding variables, and confidently explain relationships that don’t actually exist. Data quality, statistical methods, privacy, security, and validation remain essential.

For complex predictive modeling or causal analysis, specialist expertise is still invaluable.

But marketers can now participate much more directly in analytical work that previously depended heavily on technical resources.

They can explore questions, request comparisons, interrogate findings, build visualizations, and collaborate more effectively with analysts without having to write every query or model themselves.

This doesn’t remove data scientists from the equation. Instead, it makes analytical exploration more accessible and potentially more collaborative.

As that barrier falls, something else becomes more important: the ability to distinguish an interesting pattern from a useful insight.

Because finding a relationship in the data doesn’t explain why that relationship exists.

A pattern isn’t an explanation

A useful illustration comes from an independent analysis of the University of California, Irvine’s Online Retail dataset, which contains more than 540,000 transaction records from a UK-based retailer between December 2010 and December 2011.

The published analysis found that repeat customers represented 65.6% of customers but generated 92.9% of revenue.

At first glance, that’s a compelling argument for investing in customer retention. But what does it actually tell us?

Repeat customers make more purchases than one-time customers. We would expect them to contribute more revenue. The finding describes how revenue is distributed across customer groups, but it doesn’t establish why customers return or whether a particular retention strategy is responsible.

• Do repeat customers also spend more per transaction? 
• Are certain first-purchase characteristics associated with a greater likelihood of returning? 
• Which of those characteristics represent behaviors that marketing could realistically influence?

We also need to consider how the analysis was conducted, which customers were included, and whether the findings reflect differences in purchasing opportunity, customer characteristics, or other factors.

The figures might be entirely accurate, while the conclusions we draw from them are misleading. That’s the danger of treating an interesting relationship as an explanation. 

We could look at those numbers and immediately decide to invest more heavily in retention campaigns. But without understanding what drives repeat purchasing, we’re simply assuming we know what will generate additional revenue.

AI can help identify associations, suggest possible explanations, and explore competing hypotheses. What it cannot do, simply by discovering a correlation, is establish that changing one behavior will produce a particular commercial result.

Finding the pattern is the beginning of the investigation. Not the conclusion. That’s where behavioral science proves valuable.

Behavioral science helps us ask better questions

Customer behavior isn’t a collection of transactions and interactions. It’s influenced by habits, motivations, perceived risk, cognitive effort, expectations, and the circumstances in which decisions are made.

Consider customers who repeatedly browse products but rarely purchase. An analysis might reveal patterns relating to product category, price, device, or acquisition source. Those findings can help us understand where the behavior is occurring and which customer groups are involved.

But they don’t necessarily explain what’s driving it.

• Are customers struggling to make a decision? 
• Are they comparing alternatives? 
• Do they perceive purchasing as risky? 
• Are they researching products before buying elsewhere? 
• Or are they simply not ready to purchase?

Each explanation suggests a different marketing response.

A customer struggling with uncertainty might benefit from clearer product information, reviews, or reassurance. Someone comparing alternatives might need help understanding the differences between products. Someone who isn’t ready to purchase might benefit more from useful content than another promotional reminder.

Maybe none of those interventions could influence a customer who has already decided to buy elsewhere. Behavioral science doesn’t supply a convenient explanation for every correlation. Instead, it helps us develop plausible hypotheses about customers’ decisions and the factors that influence them.

The combination of AI-assisted pattern discovery, behavioral understanding, and disciplined experimentation is powerful:

• AI helps us identify relationships worth investigating. 
• Behavioral science helps us consider possible explanations. 
• Testing allows us to examine whether changing something actually influences the behavior we’re trying to improve.

Without those additional steps, we’re simply generating more sophisticated assumptions.

Better analysis should change what we test

The distinction matters. Marketers could use AI-generated findings the same way they’ve historically used dashboards: as justification for decisions they’ve already made.

An interesting correlation becomes a new segment. A predicted behavior becomes an automated journey. A relationship between discount usage and repeat purchases becomes a revised promotional strategy. But the more useful application should be to improve the quality of the hypotheses we’re testing.

Let’s return to our repeat-purchase example. Rather than assuming that the high revenue contribution from repeat customers means we should simply send more retention emails, we could investigate which experiences and behaviors appear to distinguish customers who return from those who don’t.

Suppose further analysis suggests that first-time customers who discover products across multiple categories are more likely to make a second purchase.

We could immediately introduce more cross-category recommendations. Or we could question what’s actually happening.

• Are these customers more engaged with the brand before their first purchase? 
• Do they have broader needs? 
• Does exploring multiple categories increase their confidence in the retailer, or is it simply something already-engaged customers are more likely to do?

That distinction matters because it changes what we should test.

If we believe that broader product discovery could encourage customers to return, we might test whether helping single-category customers explore relevant complementary products increases second-purchase rates.

We’d measure the incremental impact on purchasing behavior, rather than simply whether customers clicked on those recommendations. That’s a much stronger use of the insight.

The objective isn’t to turn every discovered pattern into another campaign or automation.

It’s to use the analysis to identify which interventions are worth testing and what evidence would demonstrate that they work.

Better questions could give us a competitive advantage

I find something ironic about the current fascination with AI. We’re spending so much time exploring its ability to generate answers when we should be asking better questions.

• Instead of asking AI to write another re-engagement campaign, we could ask what distinguishes customers who lapse from those who remain active.
• Instead of generating more discount-led subject lines, we could investigate how different customer groups respond to incentives over time.
• Instead of creating another set of lifecycle journeys based on predefined segments, we could examine whether those segments reflect meaningful differences in customer behavior.
• Instead of accepting the explanations our existing reports appear to offer, we could explore alternative interpretations.

None of these removes the need for robust analysis, sound statistical methods, or specialist expertise where appropriate. Nor does it mean we should act on every relationship AI identifies.

But it does expand the opportunities for marketers to investigate customer behavior and develop better-informed hypotheses.

Target’s pregnancy-prediction model demonstrated what could be achieved when sophisticated analysis uncovered relationships hidden within ordinary purchasing behavior.

As analytical capabilities become more accessible, simply finding a correlation may become less of a competitive advantage.

• Can marketers distinguish meaningful signals from statistical noise? 
• Can they interpret customer behavior rather than just describe it? 
• Can they challenge their own assumptions, develop hypotheses, and test whether proposed interventions change outcomes?
• Finally, can they recognize when an apparently interesting finding doesn’t justify any action?

AI can help us discover patterns we might never have thought to look for. But the value isn’t in finding them but in understanding what they mean, deciding whether they matter, and knowing what to do next.