Last updated: 23 September 2026
Quick Answer: Build a voice of customer program by linking score drops, complaints and cancellations to customer interviews, then giving every finding an owner and a deadline. Surveys size a problem and interviews explain it. Reach matters as much as questions: in the EU, a third of consumers who hit a problem in 2022 never complained.
A voice of customer program is the standing process a company uses to collect, explain and act on what customers say.
In the consumer survey behind the European Commission's Consumer Conditions Scoreboard 2023, a quarter of consumers said they had hit a problem with a trader in their own country in the previous 12 months, one they felt gave them legitimate cause to complain. One third of them did nothing about it.
The top reasons were that complaining would take too long (49%), the sums involved were too small (42%), or a satisfactory fix seemed unlikely (36%). The survey asked every adult, not only people who complained, so it counts the customers who stayed silent.
So a voice of customer program built on what customers volunteer hears mostly from people who still believe speaking up works. More surveys will not fix that. What fixes it is a program that treats every score and complaint as a trigger for a conversation, goes looking for the customers who stayed quiet, and holds a named person accountable for what each conversation found.
What Does a Voice of Customer Program Collect Beyond NPS?
It collects three kinds of signal:
- Solicited scores such as NPS and CSAT say where to look and how many customers feel something.
- Unsolicited signals such as complaints, tickets and reviews surface problems nobody thought to ask about.
- Customer interviews explain both, in the customer's own sequence of events.
Many programs stop at the first two.
Unsolicited signals cover ground that internal monitoring misses. When Gillespie and Reader coded 1,110 health care complaints from across England, 23% of the problems involved major or catastrophic harm.
The complaints also exposed blind spots that staff self-reporting rarely captures: trouble at admission and discharge, many small systemic failures, and staff failing to listen to patients who raised concerns. Health care supplies the evidence because complaints there have been coded at national scale with a tested tool, something few consumer businesses publish.
Scores have the opposite weakness: they merge customers who reason differently. In a June 2026 study, 571 German respondents took both a standardized survey and an AI-led interview on migration policy. Wuttke, Kreuter and colleagues found that people at similar points on the scale could hold quite different mental models of the issue.
Two customers who both give a 6 may be drifting away for unrelated reasons, and a program that stops at the score treats them as one problem.
Where Each Signal Source Fits
The table is sorted alphabetically by source.
| Source | What it tells you | Who it misses | Better choice when |
|---|---|---|---|
| Complaints, tickets and reviews | Problems customers cared enough to raise, in their words | Customers who gave up without saying anything | You need early warning on issues no survey asks about |
| Customer interviews | The sequence of events behind a score or complaint, and what would have changed it | Customers the invitation channel fails to reach | A metric moved and nobody can say why |
| Operational data (orders, returns, cancellations) | What customers did, at full coverage | Motive: it records the exit, never the reason | You need to size a problem before explaining it |
| Relationship and transactional surveys | How many customers feel what, and the trend | Nonresponders, and any reason the questionnaire did not list | You need a stable trend line or a score by store, product or team |
Five Setup Decisions for a VoC Program, in Order
Best practice for a voice of customer program comes down to five choices made before the first interview: the decisions it serves, the listening posts, the interview triggers, recruitment and consent, and who owns each loop.
- Name the decisions the program serves. Renewal risk, onboarding drop-off and a pending price change need different customers and different questions. A listening source that feeds no decision gets switched off.
- Inventory the listening posts. List every survey, complaint queue, ticket system, review feed and cancellation flow, who runs it, and where its data lands.
- Set interview triggers. A trigger is an event that earns a conversation: a detractor score, a closed complaint, a cancellation, a failed onboarding step, a key account 90 days from renewal. Give each trigger a monthly quota and a contact window, typically within a week of the event, while memory is fresh.
- Choose recruitment and consent. Recruit from your own customer records, because the program exists to hear these customers, not a sample of people like them. Record channel consent per customer; calls and texts to US mobile numbers carry their own rules, set out in TCPA rules for AI phone research.
- Assign loop owners before fielding. One person answers individual customers, and one owns each fix. Write both down now, because nobody volunteers for either after the report lands.
How often to run waves, and how to keep them digestible, is covered in continuous customer research tools and operating cadence.
How Do Customer Interviews Run at Scale Inside the Program?
As a standing monthly wave fed by triggers, not as a one-off study. Triggers fill quotas, the conversations run concurrently on an AI moderator, and coding keeps pace with fieldwork. Scale here means covering every trigger and segment each wave, not interviewing every customer.
A Worked Monthly Wave
An illustrative design for a subscription brand with 40,000 active customers, not a benchmark:
- Detractors (scores 0 to 6): 60 interviews.
- Passives (7 and 8): 30 interviews, to catch problems before they turn into detractor scores.
- Promoters (9 and 10): 20 interviews, as the comparison group.
- Cancellations in the past 14 days: 50 interviews.
- Complaints closed in the past 30 days: 40 interviews.
That is 200 conversations a month. Two rules keep it honest. Cap invitations at one per customer per quarter, since over-asking is how response rates erode, a pattern traced in survey fatigue and why response rates keep falling. And keep the core questions fixed across waves while one module rotates, so month three can be compared with month one.
At that volume, fieldwork speed decides whether the wave informs the month it describes. Alchemic runs a 200-interview qualitative study from brief to live dashboard in 3 days, 5 to 7 for complex designs, with recruitment as managed fieldwork or bring your own across 14 markets including the USA and the UK.
How much latitude any AI moderator has to chase an unexpected answer varies by tool, so read the first dozen transcripts from each new trigger before trusting its themes.
Which Channels Reach Detractors and Customers Who Never Complain?
The channels that ask the least of the customer. An email invitation reaches engaged customers. Quiet and unhappy ones may ignore it and still answer a call, a text or a messaging app, and some will talk but not type. Offer more than one channel inside the same study and let the customer choose.
The Commission's reasons for silence point the same way. "It would take too long" is an effort problem, which a lower-effort channel addresses directly. "A fix seemed unlikely" is a trust problem, which only a loop that visibly closes can address.
Mode also changes what you hear. In the same German study, voice interviews produced about twice as many words per respondent as text (a mean of 608.8 against 299.8), at 52.5 words a minute against 21.1. Dropout clustered among people assigned to voice, and those given a choice mostly picked text. Voice is richer for those who want it; forcing it loses the rest.
Alchemic interviews natively inside WhatsApp, with no link and no app, by text or voice note, and by outbound AI phone call, with consent captured on the first turn and every call recorded and searchable. The service publishes 57+ languages including Hindi, Tamil and Telugu. Web links still suit customers who live in their inbox.
What Does Closing the Loop Require in Practice?
Two loops with two clocks. The inner loop answers the individual customer, quickly, and only with their consent to act on what they said. The outer loop fixes the cause, more slowly, and needs an owner, a date and a re-measure.
The Inner Loop Needs a Clock
A public example of an inner loop at scale runs through the Consumer Financial Protection Bureau. Companies are expected to answer a complaint within 15 calendar days, or send an interim explanation and a final response within 60, and in 2025 they responded on time to 99.6% of more than 5.9 million complaints the Bureau forwarded.
That standard holds at millions of cases a year, and an internal program can set a shorter clock.
Keep research and recovery apart. At the end of each interview, ask whether the customer wants someone to follow up, and route only those who say yes to service.
The Outer Loop Needs an Owner
Every theme leaves the wave with a named owner, a decision it bears on and a date. Formatting findings for the person who decides, and owning them afterward, is covered in turning customer insights into decisions. Then tell the interviewed customers what changed, in the next wave's invitation if nowhere else.
The re-measure needs a record of what customers said before the fix. Alchemic's Insights Platform stitches interviews, surveys, support tickets and uploaded past research into one knowledge base, so a program manager can ask from Slack, MS Teams or WhatsApp whether a theme has come back since the change shipped, and the answer cites the respondents and quotes behind it.
How Do You Know the Program Is Working?
Measure what the program changes, not how much it collects. Four measures do most of the work: how fast triggers become interviews, how fast findings reach owners, how many findings ship a change, and whether the targeted metric moves in the next wave. Response volume on its own measures activity.
Set your own targets for each, then track them monthly:
- Coverage. Share of each trigger's quota filled, and share of interviews from customers who ignored the email invitation. If every interview comes from email, the quiet customers are still missing.
- Speed. Days from trigger event to interview, and days from wave close to a named owner.
- Action. Share of themes with an owner and a date, and share that shipped a change within the quarter.
- Effect. Whether the issue recurs, and whether the score moves, among the next wave's customers from the same trigger.
Where Do Voice of Customer Programs Stall?
Usually after the listening works. Programs stall when signals outrun anyone's capacity to act, when a few dozen interviews are read as a census, or when the loop to customers is promised and never closed. Four stall points recur:
- Dashboards without owners. A theme with no owner is a finding nobody will act on.
- Interviews read as a census. Forty conversations can name the reasons, not size them. Size with surveys and operational data.
- Over-contacting the same customers. Loyal customers get invited to everything and start ignoring all of it.
- Sensitive complaints routed to research. Legal, safety and distress cases need a human path. No system has been validated as a distress detector, so none should be relied on to notice.
When Another Approach Is the Better Choice
Interviews are not always the answer. Always-on measurement with case routing across every touchpoint is what enterprise CX suites such as Qualtrics and Medallia are built for; interviews diagnose, suites monitor. A dozen strategic B2B accounts are better served by account leads or a senior human interviewer than by a large automated wave. And a problem already visible in operational data, such as failed payments, needs a fix rather than research.
Once it has the client's brief, Alchemic designs and tailors the discussion guide to handle these risks before fielding, rather than leaving that work to the buyer.