The essentials on weak signals in customer conversations
- A weak signal is a rare, recurring and growing fact. Rare: it concerns a small share of conversations. Recurring: it comes back in the same form from one call to the next. Growing: its frequency increases from one period to the next. One dissatisfied customer is not a weak signal, a hundred customers asking the same unusual question are.
- It is found in what is said or done, never in an intention attributed to the customer. Four forms can be detected reliably: a phrase that keeps coming back, the same reason for a repeat call, a request reworded several times, a break in an expected sequence.
- A sample cannot see them. A signal present in 0.5% of calls appears twice when 400 calls are listened to. Across all the calls of a centre receiving 20,000 a month, it represents around a hundred.
- The method comes down to four steps: cover every call, name the signal in observable terms, track its frequency over time, confirm by listening before drawing conclusions.
- An emotion score is not a weak signal. It varies with the nature of the calls more than with the customers, cannot be verified, and becomes hard to defend as soon as it concerns an employee.
What a weak signal is, and what it is not
The expression is often used to mean "something interesting we had not seen". That is too vague to be useful. In a contact centre, a weak signal is defined by three properties, and all three are required.
It is rare. It concerns a small fraction of conversations, often less than 2%. That is what makes it invisible in the usual indicators: it moves neither average handling time, nor first contact resolution, nor the quality score.
It is recurring. It comes back in a recognisable form from one conversation to another: the same question, the same word, the same sequence. An isolated fact, however striking, remains an anecdote.
It is growing. Its frequency increases from one period to the next. A rare and stable phenomenon is background noise, worth knowing about but announcing nothing. It is the slope that makes the signal.
What it is not matters just as much:
- A spike is not a weak signal. An outage that triples calls in a single day is a strong signal: everyone sees it, and the problem is handling it, not detecting it.
- An isolated keyword is not a weak signal. "Cancel" spoken in a call says nothing without context: is the customer cancelling, or asking how to cancel an option?
- An intention attributed to the customer is not a signal. "This customer is about to leave" is an interpretation. "This customer mentioned a competing offer and asked for the end date of their contract" is a fact. Only the second can be counted, verified and tracked.
The four observable forms of a weak signal
Across thousands of conversations, the weak signals that prove useful almost always take one of these four forms. They all have one thing in common: they concern what was said or done during the exchange, and each can be found again by listening back to the passage in question.
1. A phrase that keeps coming back
Customers use the same unusual expression without having compared notes. At a health insurer, several members talk about the "new reimbursement form" when no form has changed: it is the member portal screen that was modified, and it is hard to read. At an energy supplier, the word "adjustment bill" appears in calls from customers on monthly direct debit who have never received one.
What makes the signal is the expression itself, noted exactly as the customer said it.
2. The same reason for a repeat call
A customer calls back for the same reason as the previous time. Taken one by one, these repeat calls look like bad luck. Grouped by reason, they point to a process that does not keep its promise: an announced letter that never arrives, a reconnection that takes longer than what was said, a cancellation that is not recorded.
This is one of the most profitable forms to track, because an avoidable repeat call has a direct and measurable cost.
3. A request reworded several times
Within the same conversation, the customer asks their question again in a different form, two or three times. The advisor has answered, but not the question that was asked. When this behaviour concentrates on the same topic, for example the terms of a new plan at a telecom operator, it signals an offer that is poorly understood or a sales pitch that does not answer what customers are asking.
The observed fact is the rewording, not some assumed irritation on the customer's part.
4. A break in an expected sequence
Some conversations follow a predictable flow: identification, statement of the request, solution, confirmation. The signal appears when an expected step is missing or an unexpected step is inserted. In banking, customers asking at the end of the call "and that will definitely go through today?" after a transaction that is supposed to be immediate. At an insurer, an advisor announcing a call back within 48 hours in a journey where no such call back is planned.
This is the subtlest form, and the most valuable: it often reveals a workaround of the procedure before it becomes established practice.
Why a sample never sees them
Traditional quality control listens to between 1 and 3% of calls. To measure script compliance, that is debatable but workable. To detect a weak signal, it is structurally impossible.
Take a centre that receives 20,000 calls a month and listens to 2% of them, i.e. 400. A signal present in 0.5% of calls appears there twice on average, scattered across several evaluators who each listen to their own batch. Nobody can connect two occurrences they did not hear themselves. Across all the calls of the month, the same signal represents around a hundred conversations: enough to be seen, counted and tracked.
| 2% sample | All calls | |
|---|---|---|
| Calls analysed per month | 400 | 20,000 |
| Occurrences of a signal at 0.5% | 2 on average | About 100 |
| Visible to a single person? | No, scattered across evaluators | Yes, grouped and counted |
| Growth measurable from one month to the next? | No, the difference is lost in the noise | Yes, 100 then 150 can be read |
This is why weak signal detection is one of the first concrete benefits of analysing every conversation, well before automating the score.
The four-step method
1. Cover every conversation
This is the prerequisite, not an optimisation step. Without it, the next three produce nothing but noise. It also requires reliable transcription of the terms specific to your business: a badly transcribed offer name makes the signal related to it disappear. Hence the value of choosing your transcription model and declaring your business vocabulary.
2. Name the signal in observable terms
A signal can be tracked if it is phrased as a fact that can be observed in the conversation. The difference often comes down to a single word:
| Phrasing to avoid | Observable phrasing |
|---|---|
| The customer is frustrated by the delay | The customer says they have already called about the same request |
| The customer is hesitant to sign up | The customer asks for time to think after the price is announced |
| The customer does not trust the advisor | The customer asks for written confirmation of what has just been said |
| The customer is thinking of leaving | The customer mentions a competing offer or the end of their contract |
The right-hand column has three advantages: it can be verified by listening, two people will read it the same way, and nobody can accuse you of attributing a thought to a customer.
3. Track frequency over time
A signal is read as a rate, not a count: the share of conversations in which it appears, period after period. Two precautions avoid most false alerts:
- Keep a stable denominator. If the volume of a call type doubles after a campaign, the number of signals will double too without anything having changed. Relate the signal to conversations of the same type, the same team or the same contact reason.
- Break it down before drawing conclusions. A signal rising overall may come from a single team, a single channel or a single product. That is often where the explanation lies.
4. Confirm by listening before drawing conclusions
Before acting, read or listen to a dozen occurrences. It takes half an hour, and it is what separates a signal from an artefact: a transcription that confuses two words, an analysis question that is too broad, a single very talkative customer. A tool that does not let you go back to the exact passage of each occurrence does not allow this step, and its signals remain hypotheses.
The most common false signals
- A change in call mix. A new queue, an outbound campaign or a billing period changes the composition of calls, and every rate moves with it.
- Transcription errors. A badly transcribed offer or competitor name creates a signal that does not exist, or erases one that does.
- Concentration on a few customers. Three customers who each call ten times are not thirty signals.
- Correlation mistaken for cause. Two signals rising together may have a common cause outside the conversation. An association shows where to look, not what to conclude.
Why an emotion score is not a weak signal
Many tools offer to spot "frustrated" or "angry" customers from a sentiment or emotion score. The idea is appealing. In practice, it is poorly suited to detecting weak signals, for four reasons.
It varies with the nature of the calls more than with the customers. A complaint starts from an unfavourable situation by definition. An influx of complaints lowers the average score without teaching anything new, and masks the signal that could have emerged within those complaints.
It cannot be verified. An "anger" score of 0.62 points to no specific passage that can be listened back to in order to form an opinion. Yet step 4 of the method, confirmation, is precisely what distinguishes a signal from an artefact.
It gets irony, humour and registers wrong. "Great, another outage" is not a compliment. A customer who speaks loudly is not necessarily dissatisfied. Differences in register between regions, age groups and cultures produce false positives that are not randomly distributed.
It becomes hard to defend as soon as it concerns an employee. Since February 2025, the European AI Act has prohibited systems that infer the emotions of a person in the workplace from biometric data such as the voice, except for medical or safety reasons. An advisor is an employee at their workstation. Even when the analysis concerns only the customer, the question arises as soon as an indicator of this kind is used to judge how the call was handled.
Raisetalk has made a strict choice on this front: capture expressed signals and observed behaviours, what was said and done during the exchange, never an assumed internal state. Each item captured is linked to a specific passage of the conversation, with its justification, and can be dismissed after verification. It is a constraint, and it is also what makes a weak signal defensible: you can show where it comes from.
How Raisetalk helps detect them
The four steps of the method rely on features available to all users.
Ask the question in natural language. Discovery, on the Understand page, queries all analysed conversations the way you would ask an analyst: "What topics do customers raise when calling back about the same request?", "Which questions do customers reword several times?". The answer is a summary accompanied by its sources: the conversation excerpts it is based on, each with a link to the full exchange. That is what makes the confirmation step possible. The mechanics are detailed in our article on querying calls in natural language.
Ask the same question every week. A Discovery answer can be scheduled by email. The question is not frozen along with its answer: it is asked again with each send, on the conversations of the elapsed period. It is literally weak signal monitoring: same question, period after period, and you see what appears that was not there the week before.
Break down by team, by tag, by grid. The Understand page filters, period, teams, tags, criteria, apply to questions as well as to reports. A signal can thus be compared from one team to another or from one contact reason to another, which addresses the second precaution of step 3.
Keep track of a signal once it is named. When a signal has been identified and phrased in observable terms, it can become a criterion in your grid. An alert rule triggered on that criterion then keeps its history without emailing anyone, as an in-app notification. Every occurrence is recorded and can be consulted, which gives the curve that step 3 needs.
Bring the signal to the steering committee. At the level of an entire contact centre, the management report presents the signals expressed by customers team by team, as a periodic reading rather than an isolated alert.
Start with a single question. Rather than looking for every weak signal at once, pick one of the four forms, for example repeat calls for the same reason, ask the question on the last month, listen to ten occurrences, and schedule it every week. After a month, you will know whether the signal is rising, and you will have learned to read it.
What detection does not do for you
A weak signal shows where to look. It says neither why, nor what to do. The rise in repeat calls about an announced letter may come from a print provider, a change in mailing rules or an advisor promising something that is not planned. Only a discussion with the teams concerned can decide between these hypotheses.
This is also why detection must remain upstream of any decision concerning an individual. A signal that concentrates on one team opens a discussion with its manager, it does not constitute an evaluation. For individual decisions, the reference remains the evaluation grid and the score it produces.
Finally, weak signals only concern customers who call. Those who leave without a word leave none. This is the limit shared by any voice of the customer built from calls, and it is for them that surveys and digital journey tracking keep their usefulness.
Key terms
- Weak signal: a rare, recurring and growing fact, found in a small share of conversations and whose frequency increases from one period to the next
- Strong signal: a massive and sudden phenomenon, visible without any detection tool, such as a spike in calls after an outage
- Expressed signal: an element explicitly stated by the customer during the conversation, such as a mention of a competitor or a request for written confirmation
- Observed behaviour: an element noted in the flow of the exchange, such as repeated rewording or a missing expected step
- Sequence break: a departure from the usual flow of a conversation type, through a missing step or an unexpected one
- Signal rate: the share of conversations within a given scope in which the signal appears, over a given period
Frequently asked questions
How do you detect weak signals in customer conversations?
In four steps: analyse every conversation rather than a sample, phrase each signal as an observable fact, track its frequency period after period against a stable scope, then confirm by listening to a dozen occurrences before acting. The most reliable forms are a recurring phrase, the same reason for a repeat call, a reworded request and a break in the expected flow of the exchange.
What is the difference between a weak signal and an alert?
An alert reacts to a conversation: it is triggered when a specific event occurs, for example a cancellation threat. A weak signal is read across a set of conversations and over time: what matters is the growth of a rare fact, not an isolated occurrence. The two complement each other, and an alert can be used to keep the history of a signal once it has been identified.
Do you need to analyse customer sentiment to detect weak signals?
It is not necessary, and it is often counterproductive. A sentiment score varies with the nature of the calls, points to no verifiable passage and gets irony wrong. The most useful weak signals are detected in what the customer says and does: a question that keeps coming back, a repeat call for the same reason, a reworded request.
How many conversations do you need to detect a weak signal?
Enough for a fact present in less than 1% of exchanges to appear several dozen times per period. This is why a sample of a few hundred calls listened to is not enough, whereas analysing every call of a mid-sized contact centre is.
Who should act on weak signals within the company?
Detection falls to the quality team or customer relationship analysis. Interpretation, on the other hand, happens with the teams who know the process concerned: product, billing, marketing or operations. A signal that stays within the contact centre without reaching the team able to act on its cause has served no purpose.
Getting started
Weak signals are already in your calls. The only question is whether someone can see them before they become spikes. To check on your own conversations, try Raisetalk now.