What Customers Do, Not Just What They Say: Behavioural Signals

Why look at behaviour#

A survey collects an opinion at a given moment, from the customers who agree to respond, with the biases described in our guide to courtesy bias. Behaviour, by contrast, is recorded for every customer, continuously, without anyone having to report it.

Two studies serve as benchmarks. De Haan, Verhoef and Wiesel (2015, International Journal of Research in Marketing) followed 6,649 respondents from 93 companies across 18 sectors in the Netherlands: the correlation between NPS and actual retention was around 0.17, the share of highly satisfied customers (the top two scores) did slightly better, and the effort score had almost no predictive power. Keiningham, Cooil, Andreassen and Aksoy (2007, Journal of Marketing), across 21 Norwegian companies, did not find NPS to be superior in predicting revenue growth.

These results do not disqualify surveys. They indicate that a self-reported metric on its own explains only a small part of what customers actually do. The classic metrics are presented in our guide to NPS, CSAT and CES.

The six families of signals#

Repeat contacts. A customer who comes back about the same issue in the following days has not been helped, whatever score they gave. It is one of the most direct signals of non-resolution.

Drop-offs. An online process interrupted, a call hung up while queuing, a loan application started and then left pending: each drop-off marks a stage at which the customer gave up. Our guide to customer journey mapping shows how to locate them.

Declining usage. Fewer log-ins, fewer transactions, a card no longer used: customers often disengage before they leave.

Departures. Account closure, cancellation, non-renewal. Account closures were in fact among the main grounds for disputes handled by the Moroccan Centre for Banking Mediation in 2024 (15.3% of cases), which shows that leaving can itself become a source of friction. See reducing churn.

External escalations. A case taken to the banking mediator, the insurance mediator or the regulator signals that the internal response did not convince. The volume is small, but each case carries a lot of weight.

Payment and transaction incidents. Failed withdrawals, rejections, double debits: these are technical facts, but they often herald complaints. According to Bank Al-Maghrib's supervision report for 2025, as reported in the press, some of the complaints about means of payment concerned unsuccessful withdrawals at cash machines.

Building a signals dashboard#

Three rules of use. First, express each signal against a base (per 1,000 customers, per 1,000 transactions), otherwise business growth masks everything. Second, read the signals together: a simultaneous rise in recontacts and complaints on the same reason is more conclusive than an isolated movement. Finally, link each signal to a hypothesis about its cause, then test it using the analysis described in our guide to fixing the causes of complaints.

The limits of behavioural signals#

Behaviour does not explain its own motive. A customer may use a product less because they have moved house, close an account following a death in the family, or abandon a journey because they were interrupted. Signals show where to look; verbatims, interviews and reading files tell you why. A customer health score that aggregates several signals should therefore remain an alerting tool, not a verdict.

Governance and data protection#

Combining transaction, contact and survey data creates a new processing of personal data. In Morocco, Law 09-08 requires such data to be collected for specified, explicit and legitimate purposes, and not to be further processed in a manner incompatible with those purposes. In practice:

1. Check that the purpose "improving service and the customer experience" appears in the processing notified to or authorised by the CNDP, and in the information given to customers.
2. Work as far as possible on aggregated or pseudonymised data for management reporting.
3. Restrict access to individual-level data to the teams that need it in order to act.
4. Set retention periods for combined data.
5. Do not base a decision producing legal effects for a customer solely on automated processing intended to define their profile or evaluate certain aspects of their personality: Article 11 of Law 09-08 prohibits it.

The specific banking issues are covered in our guide to data protection in banking and insurance. Where an AI model calculates a churn risk, see also AI governance in banking.

Action plan#

1. Choose no more than five signals to start with, with a written definition for each.
2. Identify the source and the owner of each data item.
3. Check the declared purposes and the information given to customers before any combination.
4. Publish a monthly dashboard that places the signals side by side with reported satisfaction.
5. Open a cause analysis as soon as a signal departs durably from its trend.