Respon
Operations

Churn risk scores hide the evidence: why we replaced ours with dated facts

A 72 out of 100 doesn't tell you what to do. Eight churn signals you can read in support conversations, dated and checkable, plus the routine to act on them.

Jérôme Gambiez · Founder, ResponZ
· 5 min read

A risk score out of 100 asks to be believed, and nobody knows what to do with a 72. A list of dated signals (what the customer wrote, what dropped, who stopped replying) can be checked, corrected and acted on. That is why we removed the churn risk score from ResponZ in version 1.6.0 and replaced it with facts you can verify one by one.

Why doesn't an aggregate score trigger any action?

A score blends signals of different natures into one number, with weights the team can't see. At the moment of deciding, three questions stay unanswered.

  • Why this account? A 72 can come from a customer who wrote that they are canceling, or from an account that is merely quiet. They call for different messages.
  • Is it true? If the score is wrong, you can neither challenge nor correct it. People stop trusting it and it sits unused.
  • What do you say to the customer? Nobody opens a call with "your score is 72". You open with a fact: "you wrote to us twelve days ago that you were going to stop".

We checked our own metrics against the ratings customers left. Those that didn't hold up were removed. The risk score was one of them. It was replaced by the list of what was actually observed on each account.

Which churn signals can you read in support conversations?

Eight signals come out of support, each with a date and, when possible, a figure. You can also collect them by hand, with no tool, from your conversation history.

  • A departure phrase written by the customer. "We are going to cancel", "we're moving to a competitor". Keep the exact wording and the date: it is the strongest signal and the only one that comes from the customer.
  • A plan change. A downgrade, with the plan before and after.
  • A drop in seats or MRR. From how much to how much, over how many days.
  • Inactivity. The number of days since the last login.
  • Reopened conversations. How many were reopened out of the total: a problem that keeps coming back, see the reopen rate.
  • Silence after a bad experience. The customer had a conversation that went badly, then nothing for so many days.
  • Emails that are no longer opened. Across how many sends.
  • A main contact who stopped replying. The person who carried most of the exchanges has been silent for so many days.

None of these proves by itself that a customer will leave. But each one says what to look at and who to talk to, which an aggregate number doesn't.

How do you separate refund requests from intent to leave?

"I want a refund" and "I'm canceling" are not the same sentence. The first is a request about an invoice or a one-off failure, often settled with a single reply. The second announces that the customer is leaving. Mixing them inflates your departure rate and buries the real cases in noise.

In ResponZ, a departure phrase is flagged as soon as the conversation is imported, on every plan and without using a credit, with the customer's exact wording visible on hover. Refund requests have their own badge and no longer count toward the departure rate. If you do this by hand, keep two separate columns in your spreadsheet and never add them together.

demo data
A conversation on the demo account with its departure intent badge.
On the demo account: the intent-to-leave badge and its reason, next to the conversation's opening tone.
A fact is only useful if it is dated"The customer talked about leaving" doesn't say whether it was yesterday or six months ago. Always record the date, the wording and the source. A departure phrase from two weeks ago with no follow-up doesn't call for the same action as one from yesterday.

How do you date, quantify and correct each signal?

Each line of the list follows the same pattern: what happened, when, by how much. For example (fictional, for the format):

  • Wrote "we are going to stop", 12 days ago.
  • Seats: 12 → 8 over 30 days.
  • Main contact went silent 21 days ago.

What sets this apart from a score is correction. An automated reading is sometimes wrong: a customer who writes "I'll cancel" in a conditional sentence, an opening tone misclassified. In ResponZ, a pencil next to a conversation's opening tone, outcome or departure intent lets you correct the reading in two clicks. The original verdict stays displayed, struck through, next to your correction. Even if you work without a tool, keep that rule: the error should stay visible, not be erased.

What routine should you set up: who looks, how often, for which action?

A list of facts only helps if someone reads it on a fixed rhythm. A simple routine is enough to start.

  • Who: the Customer Ops lead or CSM who owns the accounts, with the founder or Head of Product on the accounts most exposed in MRR.
  • When: once a week, always the same day, for example Monday.
  • What, signal by signal:
  • Departure phrase: a human writes to the customer within 48 hours, using their own words. If it is a refund, handle the billing request, not a retention plan.
  • Drop in seats or MRR, downgrade: ask what changed in their usage before offering anything.
  • Silent main contact: look for a second contact in the account and alert the CSM.
  • Silence after a bad experience: a personal message, not a campaign. See segmentation on recent intent to keep these accounts out of marketing sequences.
  • Reopens or unopened emails: raise the recurring topic with product, as in the voice-of-customer loop.

One action per signal, one owner per account, and a review date. That is everything the score didn't give you. To compare with another kind of scoring, see lead scoring from conversations.

Frequently asked questions

Is a churn score always useless? No. A score can help sort a large base. But it must be possible to break it down into facts. If it can't be, it can be neither verified nor corrected.

Do you need a tool to track these signals? Not to start. A spreadsheet with one row per account, one column per signal and a date is enough. A tool saves time once conversation volume is too high for manual reading.

What exactly does ResponZ do? It lists, per account, the signals observed in your conversations and account data, each dated and quantified, and flags departure phrases at import. It does not predict a departure date. See the Retention and churn module.

Your next bad Fin answer is happening right now.

Free · No credit card · Read-only