Banking success story

Why don't customers use the credit card they just received?

For a Latin American bank, we built a behavioral model that predicts which customers will start using their card and which beliefs explain it, so the bank knew where to start.

2.6xhigher probability of using the card when customers know what day and how much they'd pay (63% versus 24%)
Client
Latin American bank
Method
Interviews, a survey of 3,150 customers, and usage data
Service
Behavioral Research
Sector
Banking

For confidentiality, the figures and some project details have been changed.

The challenge

Cards delivered that never got used

The bank delivered thousands of credit cards a month, but only 47% of customers made a first purchase within 60 days of receiving the card. Every unused card was a wasted acquisition cost and, often, a customer who never tried again.

The team had plenty of possible explanations (the credit limit, the merchants, the card design, the perks) and no way to know which one really mattered. It needed to understand what predicts whether a customer starts using the card and how much each factor matters, so it could decide where to begin.

What we did

Listen, measure, predict, and segment

We organized the work around the COM-B model, which explains a behavior through three conditions: the person is able to do it (capability), their environment allows it (opportunity), and they want to do it (motivation).

1

We listened

We ran 18 in-depth interviews (with customers who had started using their card and customers who hadn't) and two workshops with the card team, one to generate hypotheses and one to prioritize. That gave us 24 statements about what helps or holds back a first purchase.

2

We measured

We surveyed 3,150 customers in their first days with the card and asked them to rate their agreement with each statement from 1 to 5. With their consent, 60 days later we linked their answers to the bank's records. That way we measured actual purchases instead of what customers remembered.

3

We predicted

With a logistic regression we estimated which beliefs predict a first purchase (p < 0.01), holding the others constant, and we tested the model on customers it had never seen. The result is a list of beliefs ranked by their effect on the probability.

4

We segmented

We grouped customers by their answers and found four profiles with very different usage rates. Three questions in the welcome flow are enough to tell which profile each new customer belongs to.

Why it's different

Two steps beyond market research

Market research describes what your customers think and compares groups. Here we also model which beliefs anticipate a purchase and how much each one matters, and we use that to rank the actions.

The interviews supply the hypotheses, like "if I use it, I'll lose control of my spending," and the survey measures which ones are widespread and which ones actually predict card usage.

What a typical study does
1
Describe
What customers think and do.
2
Compare
How groups differ, with significance testing.
What this diagnostic adds
3
Predict
Which beliefs anticipate a purchase and how much they shift the probability.
4
Prioritize
Which action to take first, by impact and effort.

What this diagnostic answers, and where to see it in this case

The finding

What decides whether the card gets used

How much each lever weighs in the model
Barrier43%Motivationfear of debt and fees
Enabler28%Capabilityunderstanding how payment works
Enabler17%Opportunitytools and places to use it
Enabler12%Motivationperceived benefits

Weight of each lever based on the sum of its statements' effects in the model. Figures changed for confidentiality.

The model holds up outside the sample. On customers it had never seen, in 81% of the pairs made up of one customer who bought and one who didn't, it gave the higher probability to the one who bought (an AUC of 0.81).

How much each belief changes the odds of using the card
The odds are the probability of buying divided by the probability of not buying. Each bar shows how much they rise or fall for each additional point of agreement with the statement, on the 1-to-5 scale. For a customer at a 42% probability, 52% higher odds means moving to 52%.
BarrierEnablerDoesn't predict
In gray, two beliefs that came up in the interviews but add no predictive power (n.s., not significant). Percentage changes in odds aren't symmetric, and a 38% drop (odds divided by 1.6) is a slightly larger effect than the biggest increase. Figures changed for confidentiality.

The strongest barrier is fear of debt

For each additional point of agreement with "a credit card is debt I don't need," the odds of using it drop 38%. With "if I use it, I'll lose control of my spending," they drop 34%.

The strongest enabler is understanding how payment works

Knowing when and how much you'd pay raises the odds by 52%, and knowing that paying the full balance on time means no interest raises them by 46%. Both can be taught from the welcome message on.

The limit, the channel, and the design matter little

Once the model knows a customer's beliefs, the credit limit, getting the card without asking for it, and liking its design add almost nothing. That's why a card redesign didn't make the priority list.

The insight

Fear of debt and not knowing how payment works are separate barriers. Some customers understand the card and still fear debt; others don't fear it but don't know when or how much they'd pay. Each barrier needs its own action, a welcome message that teaches how to pay and a spending limit that gives a sense of control.

How we modeled it

We flagged each customer by whether they made a first purchase within 60 days of receiving the card, using the bank's records. With a logistic regression we estimated how strongly each belief is associated with that outcome, holding the others constant and controlling for age, income, credit limit, and whether the customer had another card. We required p < 0.01. We trained the model on 70% of customers and tested it on the remaining 30%. In that group, 81% of the pairs made up of one customer who bought and one who didn't were ranked correctly (an AUC of 0.81).

Simulator

How much the probability of using the card changes

With the model's coefficients you can estimate how the probability of a first purchase would change if a belief changed, with everything else held constant. One more point of agreement on all three statements takes a customer from 42% to 69%. Move the controls to see it.

Neither agree nor disagree (3)
Neither agree nor disagree (3)
Neither agree nor disagree (3)
Estimated probability of a first purchase within 60 days
42%
The line marks the starting point (42%).
Same as the starting point (42%).

The simulator uses the model's coefficients, which measure associations. It starts from a customer who answers 3 on all three statements; since the average customer agrees a bit more than that, the bank's actual rate is higher (47%). A message rarely shifts a belief by a full point, which is why the experiment was designed to detect 5 percentage points. Figures changed for confidentiality.

Want a simulator like this built on your customers' data?

Let's talk
Who

Four profiles, four different conversations

We grouped customers by their answers and four profiles emerged. The first-purchase rate ranges from 22% to 81% depending on the profile, and each one has a different barrier.

Cautious
31% of customers
22%
used the card within 60 days

They fear debt and see the card as a risk. They need to feel in control.

Novices
27% of customers
38%
used the card within 60 days

They fear debt far less than the Cautious (37% versus 68%), but don't know when or how much they'd pay. They need clear rules.

Pragmatists
24% of customers
64%
used the card within 60 days

They understand how the card works and use it when they see a concrete benefit.

Natural users
18% of customers
81%
used the card within 60 days

They already know credit cards and how to use them. Removing obstacles is enough.

Where each profile's gap lies
Novices have the cheapest gap to close. They fear debt far less than the Cautious, and what they lack is knowing when and how much they'd pay, something the welcome message can teach.
CautiousNovicesPragmatistsNatural users
Percentage who agree (4 or 5 on the scale) with each statement, by profile. Figures changed for confidentiality.
Priorities

What to do first, by impact and effort

We plotted each action by the effect the model attributes to it and the effort and cost it would take the bank, as estimated with the card team. The top left shows what's worth doing first.

  1. 1Payment calendar in the app (if you buy today, when and how much you'd pay). High impact, low effort.
  2. 2Welcome message with the pay-in-full rule. High impact, low effort.
  3. 3Spending limit the customer sets in the app. High impact, medium effort.
  4. 4First-purchase reward. Medium impact, high cost.
  5. 5Redesigned activation flow in the app. Medium impact, high effort.
  6. 6More partner merchants. Low impact, high effort.
Impact estimated with the model; effort and cost estimated in the prioritization workshop with the card team. Figures and details changed for confidentiality.
Validation

The experiment that tests it before scaling

The model shows associations, and the experiment measures the real effect. We designed a ready-to-run test with three randomly assigned groups of new customers for the two cheapest actions.

Group A

Current welcome message

The message the bank already sent. It's the point of comparison.

Group B

A welcome that teaches how to pay

Shows what day and how much customers would pay if they buy today, and that paying the full balance on time means no interest.

Group C

The same, plus a spending limit

Adds a limit customers set themselves in the app, designed for those who fear losing control.

Metric
60 daysFirst purchase after receiving the card
Sample
1,900 customersper group, 5,700 in total
Minimum detectable effect
47% to 52%a 5-percentage-point increase
Rigor
80%power, with a 0.025 significance level per comparison

Each group is compared with the current welcome message. Comparing B with C would take a larger sample.

For decision-⁠makers

What the bank now has to decide

A survey alone shows what customers think. Add a predictive model and the bank also knows in what order to act and where the budget should go.

Instead of 24 unranked hypotheses, the team has the beliefs that matter most, how much each one can move the probability, the profiles worth reaching first, and an experiment ready to measure the result before scaling.

Want to know what predicts your customers' behavior?

Let's design the diagnostic that shows you where to act first.