Reaching customers before they leave

Start with a written definition that growth, finance and product agree on before anyone builds a model. Counting closures alone misses most of the loss. Useful definitions describe the drift itself, for example:

No activity

No transaction or login started by the customer for a set window, chosen from how often your customers normally use the product.

Main income moved

Salary or regular income stops arriving, or leaves the account on the day it lands.

Main spending moved

Card and bill payments fall well below the customer's own usual level while the account stays open.

Balance kept near zero

The balance sits just above zero for several cycles, a sign that the customer's money now lives elsewhere.

Choose one primary definition and keep the others as secondary views. A lender may care most about customers who repay early and do not borrow again. Write the definition down with its window and keep it fixed, so one quarter's results compare with the next.

The early signs are in what customers do, and the order of events often says more than the totals. A customer whose salary arrives and is transferred out the same day behaves differently from one who spends it over the month, even when their monthly balances look alike. Monthly summaries keep the counts and lose that sequence. Signs worth tracking include:

1. 1
   Income leaves quickly.Salary or regular income is transferred to another institution on the day it arrives.
2. 2
   Fewer visits.Logins and app sessions become less frequent or shorter.
3. 3
   New links to another institution.A first transfer to an account at another bank or wallet, followed by regular ones.
4. 4
   Spending shifts.Card spend falls in everyday categories such as groceries and fuel, which tend to follow the customer's main account.
5. 5
   Service contacts.Questions about fees, limits or closing balances, and complaints that were left unresolved.

Researchers at Nubank have described transformer models trained on customers' transaction data that improved predictions in large-scale recommendation problems (Braithwaite et al., 2025). The same applies to leaving: keep each customer's history with its timing, because that is where the signal lives.

A warning that arrives after the customer has moved their salary is of little use. Ask the team that will act how many days they need between a flag and a conversation, allowing for approvals and contact rules. That number is the lead time, and it belongs in the brief as a business requirement.

Then judge predictions at that horizon. A model that is very accurate one week ahead may be weaker four weeks ahead, and the four-week view may be the only one the team can use. Report accuracy at the agreed horizon, and record how many days ahead each retained customer was first flagged.

A list of customers likely to leave is different from a list of customers worth contacting. Some customers will stay whatever you do, and some will leave whatever you offer. Others will respond to a call or a better rate, and a few will be prompted to leave by the contact itself. Budget spent on the first two groups is wasted, and contacting the last group does harm.

Uplift modelling estimates the extra effect of an action on a customer's outcome, compared with taking no action. The methods reviewed by Gutierrez and Gérardy (2017) all need data on customers who were contacted and customers who were not. In practice that means a holdout: a random share of flagged customers who receive no action, so the difference between the two groups shows what the programme achieved. Keep the holdout running at a size the business accepts, so every quarter's results can be checked against it.

A ranked list is useful only if someone works through it. Start from capacity: how many customers your teams can contact well in a week, and how many offers the budget covers. Cut the list at that number and rank it by expected retained value, so the effort goes where it is worth most.

Save each week's list as a named group, with the reason each customer is on it. Every team then works from the same names, no customer is contacted twice, and each result can be traced back to the group that was targeted.

Rules that let customers share their data and move money between institutions lower the effort of switching. In the UAE, the Central Bank's Open Finance Regulation requires licensed institutions to give participants in the framework access to customer data and the ability to initiate transactions, in all cases subject to the customer's express consent (CBUAE, June 2024). In December 2024 the Central Bank's board approved Nebras Open Finance, a subsidiary that runs the central infrastructure for open finance (WAM, December 2024). In Brazil, the Central Bank's Open Finance lets each customer decide which data to share, with which institution and for how long (Banco Central do Brasil).

This works in both directions: a competitor can make an offer based on a customer's shared history, and so can you. The time between the first sign of drift and the move itself can shorten, which makes lead time and frequent scoring more important.

1. 1
   The same offer to everyone flagged.It spends money on customers who would have stayed and can push away some who would not.
2. 2
   No holdout.Without a group that received no action, the team cannot separate what it achieved from what would have happened anyway.
3. 3
   Measuring the model instead of the result.A better accuracy score is welcome, but the business case rests on retained value.
4. 4
   Scoring too rarely.Scoring once a month when behaviour changes weekly means many customers are flagged after they have already moved.

Retained value against the holdout

Revenue or balances kept in the contacted group compared with the holdout, over a fixed period after contact.

Cost of offers

Rates, fee waivers, rewards and staff time, set against the retained value.

Response rate by group

Who accepted, stayed or left, by segment and by offer, so the next campaign is better aimed.

Lead time

Days between the first flag and the date the customer would have met your definition of leaving.

Report these together every month, with retained value against the holdout, net of the cost of offers, on the first page and model accuracy in the appendix.

1. 1
   What exactly counts as leaving here, and over what window?
2. 2
   How many days ahead do we need a flag to act on it?
3. 3
   How many customers can we contact well each week?
4. 4
   Which flagged customers are held out, and are they chosen at random?
5. 5
   What did the last campaign retain against the holdout, after the cost of offers?
6. 6
   How often are customers scored, and is that as often as their behaviour changes?

Mantika Tessera learns from your customers' history and shows your teams who to keep and who to grow.