Share article
„All a customer has to do is say what happened, and artificial intelligence asks for the details on its own. It promises an end to empty “Why?” boxes and answers that finally explain what lies behind the score. In experiments, however, almost half of the people asked to answer by voice chose to close the survey instead. So where does voice help feedback, and where does it quietly damage it?“
The Paradox of the Most Natural Interface
Companies are asking their customers more questions than ever, yet they are learning surprisingly little. A typical satisfaction survey reduces an entire experience to a single number, most often NPS (Net Promoter Score, the willingness to recommend a company on a scale from 0 to 10) or CSAT (Customer Satisfaction Score, satisfaction with a specific interaction). It then adds an empty text box asking “Why?”.
Yet that box carries the information leadership needs most, namely the cause. And it is precisely the box respondents fill in least willingly, especially on mobile, where answers to open-ended questions are demonstrably shorter than on a computer.
Dissatisfaction with this state of affairs is not new. In 2022, Gartner predicted that by 2025, 60% of organizations with VoC programs (Voice of the Customer, the systematic collection and analysis of customer feedback) would supplement traditional surveys by analyzing voice and text interactions with customers. According to Gartner, customer service departments regard surveys as a less valuable source than other methods.
Generative AI now offers a third path between the questionnaire and call analytics. That path is a survey that conducts a conversation. Respondents speak or type in their own words, the system understands them and asks follow-up questions wherever an answer is vague. On paper, it is the ideal combination of a questionnaire’s scalability and the depth of a qualitative interview.
The data, however, reveal an uncomfortable paradox. Speech is the most natural way people share experiences, and yet respondents in controlled experiments repeatedly refuse to use it. The real question, therefore, is not whether AI understands the spoken word. It understands it better and better. The question is rather under what conditions customers are willing to speak, and when their own words tell a company more than the points they give on a rating scale.
How a Survey That Asks Follow-Up Questions Works
The common label covers two technologies that can be combined.
The first is the conversational survey. Instead of a static form, the respondent is guided by a chatbot built on an LLM (Large Language Model, such as GPT). It asks an open-ended question, evaluates the answer and follows up based on its content. It can ask for details (“What exactly held you up?”), steer the respondent back to the topic if they have drifted, or offer its own interpretation and ask for confirmation (“Do I understand correctly that this is mainly about the shipping cost?”).
The last option is particularly interesting for CX (Customer Experience). During the conversation, the model assigns the answer directly to a category. Coding that analysts today carry out weeks after data collection thus happens in real time and is checked by the customer themselves.
The second technology is voice input. In dictation, ASR (Automatic Speech Recognition) converts speech into text directly on the screen. Respondents see the transcript, can correct it, and the company stores only the text. In recording, an audio track is stored and transcribed afterwards. The recording also carries tone, pace and intonation, information that is missing from written text.
Combining the two technologies creates a voice AI interviewer that conducts a conversation much like a person. Early academic pilots suggest it can match human interviewers in data quality. So far, however, only on small samples in controlled settings.
What Changes for the Respondent
The most visible change is the shift from filling in a form to having a conversation. As early as 2020, before today’s language models arrived, a team from the University of Illinois compared roughly 600 respondents. Half completed a standard online questionnaire, half answered a chatbot. The chatbot generated significantly higher engagement, and the answers were more informative, more specific and clearer (Xiao et al., 2020).
The second change is time. In a Spanish experiment by Melanie Revilla and Mick Couper (2026), answering an open-ended question by voice recording took a median of 27 seconds, compared with 61 seconds when typing. For a customer on a mobile phone, that is a significant difference. It applies, however, only when the customer is in a situation where speaking is possible.
Depth Versus Processability
The data are consistent on this point. Spoken answers are longer, have a richer vocabulary and are different in nature. In a series of smartphone experiments, German methodologists described written answers as deliberate and considered, and spoken ones as intuitive and spontaneous (Gavras et al., 2022). For CX, this implies a practical difference. Written text shows what customers want to tell a company, while the spoken word is closer to how they actually experienced the situation.
Conversational follow-up works in a similar way. An experiment by NORC at the University of Chicago with 1,800 respondents showed that AI follow-up questions substantially increased the share of specific and well-reasoned answers (Barari et al., 2026). Across the sample as a whole, however, they did not surface new topics. A probe therefore deepens what an individual says, but it may not reveal what is missing from the sample. It also comes at a cost, because the interview took an average of six minutes instead of three.
The same study uncovered another trap. When the AI offered its interpretation for confirmation, respondents agreed more often than was warranted. For one of the questions, 81% of people confirmed the proposed category, while independent human coders agreed with the AI in only 72% of cases. People tend to acquiesce, and the question “Do I understand you correctly?” is therefore not, by itself, a reliable check.
The Price of Richness
The richer the answer, the harder it is to process. Spoken answers contain more repetition and are more cluttered. The main risk, though, is transcription. Research from Cornell University found that roughly 1% of transcripts produced by OpenAI’s Whisper in 2023 contained entire fabricated sentences that were never spoken in the recording, and more than a third of these contained harmful content (Koenecke et al., 2024). According to the lead author, newer versions of the model make fewer such errors. For a VoC program, however, even a small percentage is a warning sign, because an invented sentence in a complaint can end up in a management report.
What is often overlooked is comparability over time. A conversational survey is adaptive, and each respondent receives slightly different questions. For finding root causes, that is an advantage; for tracking a metric’s trend, it is a problem. A sensible design therefore keeps the core metric closed-ended and deploys the conversational layer on the “why” question.
When Voice and Conversation Make Sense, and When They Don’t
Depth is paid for in respondents. In German experiments, researchers randomly assigned who would answer by voice and who by typing. In the voice group, 45% of people abandoned the survey early, compared with only 13% in the typing group. In a follow-up study, the figures were 51% versus 24%. Voice, in other words, tended to discourage respondents. Their own choices point in the same direction. When participants in the Dutch LISS panel could choose how to answer, 94% chose typing. Voice appeals mainly to younger, tech-savvy users.
The most telling evidence comes from the experiment by Revilla and Couper. Even though respondents always had a text alternative available, merely offering voice reduced the overall response rate. The most common reason people did not use voice was not technology. It was their surroundings, typically the presence of other people, followed by difficulty expressing themselves orally.
Moreover, voice is not necessarily better even where respondents are able to speak. In text-based interviews on mobile phones, people report more precise numbers and disclose sensitive information more openly than in voice interviews, regardless of whether the questions are asked by a person or an automated system (Schober et al., 2015).
Where It Makes Sense
Voice and conversational follow-up deliver the greatest value for “what happened” and “why” questions, where a company is looking for a cause rather than a number. A typical case is the follow-up question after a low score, where every additional detail shortens the path to a fix. The benefit grows with the number of open-ended questions. With each additional question, answers get shorter, but not when they are recorded by voice.
It is essential to offer voice as an option, not a requirement. Prompting respondents to dictate also works better than prompting them to record. The same restraint applies to text-based AI probes. NORC allowed at most one probe per question and recommends following up only where answers tend to be vague, particularly among mobile users.
Where It Doesn’t
The conversational layer is not suited to metrics that are tracked over time or benchmarked against competitors, because adaptive questions undermine comparability. Nor is it suited to precise figures, such as how many times a customer called support, because spoken answers tend to be rounded. Caution is also warranted on sensitive topics, whether finances, health or EX (Employee Experience), where an employee is evaluating their own manager. Typing leads to more candid answers there, and voice runs into the problem of surroundings, because in an open-plan office very few people will assess their boss out loud.
Finally, if a company needs a representative picture of its entire customer base, the high dropout rate for voice will skew it toward younger and more tech-savvy customers. Depth gained at the expense of representativeness is a bad deal.
What to Consider Before Deployment
Language and Accent
Speech recognition errors are not randomly distributed. When Stanford researchers tested five commercial systems from Amazon, Apple, Google, IBM and Microsoft, the systems made an average of 35 errors per hundred words for African American speakers and 19 for white speakers (Koenecke et al., 2020). The measure used was WER (Word Error Rate, the share of incorrectly transcribed words). The systems have improved significantly since then, but the principle still holds. Errors accumulate among groups whose speech is less represented in the training data.
For the Czech market, this translates into a concrete task. The methodological experiments cited in this article worked with German, Spanish and English. A company that wants to run a voice survey in Czech should therefore not rely on a vendor’s general benchmarks. It should measure the transcription error rate on its own recordings, separately for at-risk groups, namely older customers, regional speech and answers recorded in noisy environments. What matters is not the average but the gap between the best- and worst-recognized groups. Dictation has an advantage here, because customers see the transcript and can correct it before submitting.
Accuracy and Hallucinations
Risks arise at two levels. The transcript may contain fabricated sentences, as the Cornell University research showed. The language model may then misinterpret the answer or code it incorrectly. NORC tested its chatbot extensively before data collection and found only two off-topic probes during fieldwork. Such testing, however, is a prerequisite, not a given. Good practice therefore includes a pilot phase and regular audits in which a human analyst compares a sample of transcripts with the recordings and a sample of AI codes with their own assessment. Respondent confirmation cannot replace this check, because people tend to acquiesce.
Respondent Trust
Willingness to speak rests on trust. In the Spanish experiment, voice input was chosen more often by people who fully trusted the confidentiality of their answers and by people who use voice in everyday life. Gender, age and education had no significant effect. For survey design, this means that a clear explanation of what will happen to the recording is part of the methodology, not a legal formality. Dictation, where no audio is stored at all and the company receives only text, reduces privacy concerns. A voice recording is personal data under the GDPR (the EU General Data Protection Regulation) and requires an appropriate legal basis, notice to the respondent and a defined retention period.
Regulation Under the AI Act
Since 2 August 2026, the transparency obligations under Article 50 of the EU Artificial Intelligence Act (AI Act) have applied. Systems that interact directly with people must be designed so that people know they are dealing with AI, unless this is obvious from the context. Anyone deploying an emotion recognition system must inform the people affected.
For EX, Article 5(1)(f) is more significant. Since 2 February 2025, it has prohibited the use of AI to infer the emotions of people in the workplace, except for medical or safety reasons. The European Commission’s guidelines cite as a prohibited example a call center that uses voice recognition to track its employees’ anger. Tracking customers’ emotions does not fall under the ban. A voice pulse survey among employees that assessed their emotions from their tone of voice would thus fall into prohibited territory. The Act defines emotion recognition as inference based on biometric data, which includes voice. Any specific deployment should be reviewed by a lawyer before a vendor is selected.
Data Integrity
The final risk comes from the opposite direction. Sean Westwood of Dartmouth College (PNAS, 2025) built an autonomous synthetic respondent that, across 6,000 trials, passed 99.8% of standard attention checks. Fluent, coherent and content-rich answers, exactly what a conversational survey is looking for, can now be convincingly imitated by language models. Not even the voice channel can automatically be taken as proof that a human is responding. The safest approach remains to tie feedback to verified customers and specific transactions, rather than to open panels.
A Layer, Not a Replacement
Voice and conversational surveys are not the successor to the classic questionnaire. They are a new layer on top of it. Research so far shows that they produce longer, more spontaneous and more specific answers, reduce the time needed to respond and allow data to be coded during collection. At the same time, they lose some respondents, undermine comparability over time and introduce new risks in transcription and interpretation.
Methodologists therefore remain cautious. Based on the evidence so far, there is no reason yet to replace typing with speech in online surveys. The growing use of voice in everyday life and rapid technological improvement suggest, however, that this may change.
For CX leaders, this points to a measured approach. The core metric stays closed-ended and comparable. The conversational layer is deployed selectively on the “why” question, ideally for customers who gave a low score. Voice is offered as a choice alongside text, ideally in the form of dictation. The pilot measures not only the depth of answers but also dropout rates and transcription accuracy by segment. And in EX, emotion analysis from voice is not used at all.
The real question, therefore, is not whether to let customers speak. It is where their own words say more than the score they give, and whether the company is prepared to listen to those words carefully enough.
