The Churn Signals Already Sitting in Your CRM

A churn post-mortem reads the same way in most businesses. Somebody opens the account record and works backwards, and the history looks like a warning written months in advance and delivered to nobody. Contact tailed off in March. Two tickets sat unresolved through April. The sponsor of the original purchase moved on in May, and their replacement never came to a review. None of it was hidden. It sat in the CRM, in five places, read by different people at different times and joined up by no one.

The gap is not detection technology. It is that nothing was responsible for holding those five facts together and putting them in front of a named person while there was still time to act.

A signal is a condition with evidence, a confidence level and a named owner

The word gets used loosely. A signal a customer success lead can use has four parts.

• A condition, stated plainly enough to argue with. "Contact with this account has been well below its own normal pattern for six weeks, including a scheduled review that did not happen." A number is not a condition.
• The evidence, dated and linked. The review that slipped, the tickets that stayed open, the contact whose job title changed. If a person cannot click through to the record a claim came from, they will not trust it, and they are right not to.
• A confidence level, with the reason for it. Some are close to certain and checkable in seconds, such as the renewal contact having left. Others are inference from thin evidence. Saying which is which is most of the difference between a finding and a guess.
• A named owner. One person who decides what happens next, with a recommended intervention to accept or reject and a stated condition for closing the signal. Risk that belongs to a team belongs to nobody.

The signals that genuinely live in CRM and activity data

Five hold up in practice. Each carries a caveat that settles whether it is worth building at all.

• Falling engagement, measured against the account's own pattern. Meetings, calls and replies compared with normal for that account, not a company-wide threshold. Engagement with your team is not the same as engagement with what you sold them, and the two can move in opposite directions.
• Unresolved service, not ticket volume. Reopened tickets, resolution times drifting from the account's own history, tickets with no owner, escalations that never came back. Volume on its own is closer to a usage measure than a risk one.
• Adoption decline, but only if usage data reaches a system you can query. If usage lives in the product and has never been connected to anything, it is not a signal, it is an assumption. Reaching it is integration work and should be scoped as such before anyone promises detection built on it.
• A changed champion. The person who ran the buying decision leaves, changes role, or stops replying while somebody more junior attends. Often the earliest signal available, and visible only if contact records are maintained, which is the data that decays fastest in most CRMs.
• Silence after a complaint. An escalation followed by nothing is not a resolution. It is the point at which a customer decided the conversation was not worth continuing.

The last is the hardest to build, and the reason is instructive. Systems record events, not the absence of them, so absence has to be computed against an expectation, and the expectation has to be set per account. An account you speak to quarterly by design is not at risk because nobody emailed for three weeks. One global threshold produces false alarms from your best-behaved accounts and misses the one that mattered.

A score out of 100 cannot be acted on

Most tools here answer with a number, sometimes a colour. It looks like progress and it does not survive a renewal conversation.

• Nobody can interrogate it. No one walks into a review with a number and a downward arrow. The first question is why, and the model will not say.
• It compresses different causes into one output. A departed champion and three unresolved tickets need completely different responses. Averaged into one number they become the same instruction, which is no instruction.
• It cannot be corrected. When a score is wrong there is nothing to fix. When a condition is wrong you can see the evidence behind it, correct the source and tighten the rule.

Underneath all three sits a subtler point. A model can be entirely confident about the wrong evidence, so confidence is not the control that matters. Grounding is: can this finding be traced to specific records, and do those records say what it claims.

Suppression and duplicate handling decide whether anyone still reads it in month three

This is missing from almost every demo, and it decides whether the work survives its first quarter. A signal that fires weekly stops being a signal. People rarely switch it off, which would at least be honest. They stop reading it, and the queue looks healthy while nothing in it is being worked.

• One open signal per condition per account. New evidence updates the existing finding rather than creating a second beside it.
• Re-fire on material change in evidence, never on a schedule. A weekly reminder that an account is still quiet is not new information.
• Dismissal is recorded with a reason, and the reason is read. "The sponsor is on parental leave until October" is the input that stops the same false alarm arriving in November.
• A cap per owner per week. If more risk surfaces than the team can work, raise the bar on what qualifies rather than adding a filter in somebody's inbox.
• A signal closes on a recorded intervention, or a recorded decision not to intervene. Not because a number drifted back up.

Where the approval line sits

Detection is a read job. It reads records, assembles findings, explains the evidence and recommends. It does not contact the customer, and that separation is deliberate.

Recommend is where this sits, and for renewal risk it is very often the correct permanent setting rather than a stage to graduate from, because deciding what to do about a struggling account is the part worth a person's time. The value is that a person sees a grounded finding on Tuesday instead of finding out at renewal.

Two boundaries are worth stating. Authority to read the CRM does not carry authority to message a customer: that is a separate grant, per channel, per audience and per message type. And nothing about a detected risk changes who is accountable for the account. These are the retention capabilities designed to sit inside the CRM you already run, set out on our AI for customer success page.

Built into the CRM, the signal, the record and the action are one thing

A separate platform holding its own copy of your customer data can only compute a signal from what crossed the sync boundary. The call note explaining why a customer cancelled their quarterly review does not cross it, and that note was your earliest evidence. So the signal is calculated from the fields that travelled rather than from what you know, and the finding lands away from the record, which means somebody moves it across by hand.

Built into the CRM you already run, the finding sits on the account, the evidence links to the tickets, emails and contact changes it was drawn from, and the follow-up is a task in the queue your team already works. Where growth rather than risk is sitting in those same accounts is a separate question, covered in finding expansion revenue in accounts you already have.

A detected risk is not a saved account

This is where most retention programmes stop measuring. Detection is easy to celebrate because it produces visible output, and on its own it produces no result. What matters is whether an intervention happened, how quickly, and what followed. Track four things, and baseline all four first, because if you cannot state today how late you find out that an account is in trouble, you cannot say afterwards whether it improved.

• Findings accepted or dismissed by a named human, with the dismissal reasons, which are the tuning data.
• Time from detection to first intervention. The single number most worth moving.
• Intervention completion, meaning the action was carried out rather than assigned.
• Renewal movement afterwards, reported as renewal movement.

Be careful with the phrase churn prevented. You cannot observe the account that would have left, so most saves are asserted rather than evidenced. Report the movement you can see and be sparing with attribution. A programme that claims credit it cannot support loses the argument the first time a finance director asks how the number was produced.

The next step

The first question is not which detection tool to buy. It is which risk conditions your own data can support today, which need an integration first, and who owns the decision when a finding appears. Detection of this kind would be one of the two production agents activated in a customer success wave of AI Accelerator, our 12-month department-by-department programme.

To establish the ground before committing to anything, start with a diagnostic. The assessments there are short and fixed in scope, and they say where your data would hold and where it would not, including where the signal you want cannot be built from what you record today. No obligation. For where retention work falls in a wider rollout, see sequencing AI a department at a time.