Reducing customer churn means noticing that someone is about to leave while they are still a customer, before the cancellation arrives. Most customer care tools watch the facts: late invoices, open tickets, a drop in product usage. These are real signals, but they arrive after the customer has already decided.
Transactional data is accurate, but it lags behind what the customer is already thinking. A dissatisfied customer stops opening the app weeks before cancelling, and in the meantime writes emails, opens tickets, chats with support. The tone in those texts shifts before the numbers do: that is where a real window to act still exists.
Why churn is always caught too late
Churn is measured almost always after the fact. A customer gets flagged as at risk when usage drops below a threshold, when support tickets pile up, or when the renewal date approaches with no sign of confirmation. These are useful indicators for building a list of accounts to watch, less useful for deciding what to do with each one.
The reason is structural. A usage drop, a run of tickets, a renewal that stalls are the consequences of a decision the customer already made, often weeks earlier. By the time the system raises a flag, the customer care team is stepping into a situation that has already set: dissatisfaction has had time to build, and the room to act has narrowed down to a handful of options, often just a last-minute discount.
There is also a threshold problem. A 20% drop in usage can mean disengagement, or it can mean a quieter seasonal stretch, and the system treats both cases the same way because it reads a number, without catching the intention behind it. The result is a risk list full of false positives and false negatives, one the team then has to sort through by hand before acting on it.
What usage data shows, and what it does not
Usage data tells you what the customer did: how often they opened the app, which features they used, how much time passed since the last login. It does not tell you why, and that gap matters more than it looks.
Two customers with the same 30% drop in usage can be in opposite situations. The first cut their usage because they automated part of the work with the product and now need to log in less often: a satisfied customer getting more value with less effort. The second cut their usage because they started evaluating an alternative and are gradually moving their workflows elsewhere. The behavioral data is identical, the outcome is the opposite: in the first case a sales outreach is unnecessary and risks feeling intrusive, in the second case it already arrives too late.
The same holds for support tickets. A customer opening more of them can be frustrated and close to leaving, or can be a customer learning to use more features and needing more support precisely because they are using the product more. Ticket volume alone does not tell the two cases apart.
None of this means usage data is useless: it remains the most reliable way to know what is happening inside an account. Its limit is that it answers a different question from the one customer care actually needs answered, the intent behind the numbers. That kind of signal comes from how the customer expresses themselves when they write, a layer that click tracking never reaches.
What the customer says before they leave
The signal that anticipates churn sits in the grammar of the message. Psycholinguistics separates content words, meaning nouns, verbs and specific adjectives, from function words: articles, prepositions, pronouns, conjunctions. Content words change from one message to the next depending on the topic. Function words stay stable over time for the same person, because they get used without thinking, and that is exactly why they reveal more about the psychological state of the writer. Psychologist James W. Pennebaker documented this in his book The Secret Life of Pronouns (2011), showing how the frequency of pronouns, articles and prepositions reveals stable personality traits and shifts in emotional state, regardless of what the person is actually talking about.
Portrait applies this principle to customer conversations. It analyzes the frequency of function words in chat, email and tickets and returns a profile of 50 base psychometric traits, valued from -1 to +1, plus 38 aggregated traits, valued from 0 to 1: correlation matrices already built, developed in collaboration with the Massachusetts Institute of Technology (MIT), that combine several base traits into one indicator. Attitude-Churn is one of them: it combines the base traits tied to the tendency to perceive dissatisfaction with people and products with the ones tied to the tendency to abandon tasks that require attention and perseverance.
The practical point is that this kind of signal does not depend on what the conversation is about. A customer might write about an invoice, a technical issue or a feature request: the Attitude-Churn value is calculated on the grammatical structure of the text, regardless of its specific content. That is why the score works on any text channel, with no need to design special questions or a survey for the customer to fill in.
How to use a risk score
A risk score like Attitude-Churn is calculated on the texts a customer already produces during the relationship, with no dedicated channel required. The most common sources are:
- chat and virtual assistant messages
- support and sales emails
- support tickets and their replies
- free-text notes the team adds to the CRM
With enough text, at least 300 characters, the score updates and becomes a traffic light inside the platform: green when the trait configured in the correlation matrix is detected in the analyzed customer, grey when it is not.
A high Attitude-Churn value signals that the account is worth a look before the cancellation lands. The cause has to be found in the context: a customer with a high score and a history of unresolved technical tickets needs a technical fix, while a discount in that situation wastes margin without solving the problem. A customer with a high score but no open issues might simply be going through a quieter stretch, where a generic sales outreach can feel as intrusive as an unwanted discount. A low value, as in the example above, says the opposite: there is no urgency, and a proactive offer at that point wastes margin instead of preventing a departure.
The same retention action, in other words, does not fit every customer with the same score: the Attitude-Churn value tells you when to look at the account, the rest of the context (open tickets, contract value, purchase history) tells you what to do.