AI Voice Agent for Surveys: How to Modernise CATI in Opinion Research
Content

An AI voice agent for surveys is a virtual interviewer that makes the fieldwork calls, administers the questionnaire —with filters, grids and branching—, records the answers in structured fields and leaves the data ready for analysis, so the research team can focus on design, quality control and the cases that genuinely need a person. For an opinion research institute, it is a way to absorb demand peaks, depend less on interviewer availability and keep the quality standard of CATI.
This guide covers the bottlenecks of telephone fieldwork, what an agent needs to do, how Spoki Voice fits in and when to complement CATI with WhatsApp surveys.
Why AI-powered CATI matters to opinion research now
Institutes still run a substantial share of their fieldwork by phone when the questionnaire is long, quotas are tight or the sample cannot be reached online alone. The model with human interviewers works, but cost per interview weighs on margins, capacity is rigid —if a client needs fieldwork “right now”, the team may be booked two weeks ahead— and refusals at the introduction cost time. On top of that, high-volume outbound calls end up flagged as spam, so numbers have to be rotated and each study needs its own number, so that a call-back lands in the right questionnaire.
Quality is non-negotiable: the agent has to follow the questionnaire, capture usable data and sound natural when the respondent asks for a question to be repeated or interrupts. Some institutes have tried other automated voices and did not take them into production because they did not clear that bar.
What an AI voice agent for surveys needs to do
Conversation quality. It has to handle repetitions and interruptions without leaving long silences while it fills in a field. Low latency helps, and Spoki Voice answers with low latency and noise cancellation. That does not replace a test with a real questionnaire.
Fidelity to the questionnaire and the data. It must run a complete web questionnaire, not just a short list of closed questions. A typical study has filters, grids, branches and durations of close to a quarter of an hour. Answers have to land in structured, traceable fields that can be exported in several formats, including JSON. Without that fidelity, the call may sound good and the data file still be useless.
Languages, numbers and retries. Many studies run in more than one language, including those for public administrations. The agent should start in the language of the study and switch during the same call if the respondent answers in another one; it is worth checking real coverage for each language (recognition, model and voice). Fieldwork should not depend on the office phone system: numbers are bought and assigned in the platform, retries are scheduled and a call-back reaches the same agent and the same study, with number rotation when needed.
Agent disclosure and data protection. The person on the line has to know from the first moment that they are speaking with a virtual agent. It is an ethical requirement in the industry and, since 2 August 2026, an obligation under Article 50 of the EU AI Act, unless it is obvious. The notice is configured in the greeting. Recordings, transcripts and answers must comply with the GDPR: data minimisation, retention periods, restricted access and, when the study requires it, EU-based infrastructure and a data processing agreement.
How Spoki fits an opinion research institute
Spoki is the agentic conversational platform. It brings voice, WhatsApp, SMS and email together in a single thread per contact.
In a fieldwork flow, the typical fit looks like this:
- You start from the real questionnaire and a briefing, the same kind of material you would give a human interviewer: study objective, instructions, frequently asked questions and, if it adds context, results from a previous wave.
- The agent conducts the interview by phone, follows filters and branches, asks for repetitions when needed and fills in the structured fields while it talks. Short pauses cover the work happening behind the scenes, much like an interviewer clicking “next” and waiting for the screen to load.
- Unanswered calls do not consume voice usage. Spoki Voice is pay-as-you-go, with no fixed volume fee. You can set up automatic top-up so a wave is not cut short for lack of credit; if the balance runs out mid-call, the interview may end. Numbers are managed in the platform: the first one is included and additional ones have a monthly fee. Calls can be masked so the respondent receives them from the client’s number. For up-to-date figures, see Spoki pricing.
- The data stays in contact fields (each question can be a column), completed surveys are tagged and the data is exported in several formats —CSV or JSON, among others— or connected to the institute’s tools. Calls can be listened to and transcribed from the platform.
- If the study calls for it, the conversation continues on WhatsApp, SMS or email —a pre-call notice, an invitation, a reminder or a recontact— without losing the history.
The sensible approach is not to replace all human fieldwork at once, but to run a proof of concept with a real survey: interaction quality, completion, record accuracy and cost.
WhatsApp surveys and mixed-mode designs
The phone is still the channel for many opinion studies, but it does not have to be the only one. A mixed-mode design can reserve AI-powered CATI for the long questionnaire or the hardest-to-reach profiles, and use WhatsApp, SMS or email for the rest of the journey.
A WhatsApp notice before the call —the number that will call and the purpose of the study, sent only to people who have agreed to be contacted— can make it easier for the person to pick up when the call comes. The same channel then works for reminders, short questionnaires, recontact when a quota is missing or confirmation that the answers have been recorded.
In Spoki, the notice and the call sit in the same thread. They do not replace the CATI questionnaire or the sampling rules: they complement them.
Retries, calling hours and data export
In real fieldwork, it is not enough for the agent to conduct one good interview: you have to orchestrate the attempts and get a usable data file.
- The automation fires when a contact enters a list. So you set everything up first (calling hours, retries, tags) and only then import the numbers. As soon as a contact enters the list, the call can go out.
- Several attempts, not a single try. If there is no answer or the call reaches voicemail, a second and third attempt are scheduled a few hours apart. If the person refuses the survey, they are not called again. Soft cases (call back later, can’t talk now) are tagged for recontact, but that has to be configured; “not reached” retries are not enough.
- Calling window. Retries must not land in the middle of the night. Business hours are set at account level and, if needed, at each step of the automation. Weekends are a judgement call: many studies do not call on Sundays.
- Batches, not the whole sample at once. A test with very few numbers may not yield a single completed survey, because phone response rates tend to be low. It makes sense to validate the questionnaire first by calling yourself —to check that the agent reads the sample variables correctly, such as the municipality, and asks for the data if it is missing— and then launch a pilot batch, review transcripts, tags and credit, and only then scale up.
- Completed-survey tag and export. Answers are stored in contact fields and exported in several formats, including JSON. A “survey completed” tag lets you export only finished interviews, not every attempt. In the platform you can listen to calls, read the transcript and see duration and status.
How to set up a proof of concept
- Choose a representative questionnaire and then a more complex one: with grids, filters or branches, which is where the limits show.
- Separate the script from the knowledge base and fine-tune it with short tests (unexpected answers, filters, sample variables) before launching fieldwork.
- Configure numbers, calling window, retries, the completion tag and the virtual-agent disclosure before importing the list. Create the empty list, link it to the automation and only then upload the contacts.
- Test by calling yourself first, to hear how it reads the questionnaire and the variables. Then launch a pilot batch, not the whole sample.
- Review on the same day transcripts, refusals, completed surveys, cost per completed interview and credit (with automatic top-up, if needed). If data quality holds, scale up study by study.
Frequently asked questions
What is an AI voice agent for surveys?
It is AI software that makes or answers fieldwork calls, administers the questionnaire, records answers in structured fields and leaves the data ready for analysis, with the option to continue the contact on WhatsApp, SMS or email.
Can a voice agent run CATI with filters, grids and branching?
It can, if it is configured from the real questionnaire and not from a simplified list. A useful test is one of the institute’s typical surveys and one of its most complex.
Do WhatsApp surveys replace CATI?
Not necessarily. WhatsApp works well for a pre-call notice, invitations, reminders, short questionnaires and recontacts. AI-powered CATI is still useful when the questionnaire is long, quotas are demanding or the profile responds better by phone.
Can you send a WhatsApp notice before the fieldwork call?
Yes. A prior message identifying the number and the study, sent only to people who have consented, can make it easier for them to pick up. In Spoki, the notice and the call sit in the same thread.
Are people who don’t answer, or who refuse the survey, called again?
No-answers and voicemail usually get a second and third attempt, within the study’s calling window and after a configurable interval. Anyone who refuses should not be called again. If they ask to be called later, that recontact has to be scheduled separately.
Do you have to say the caller is a virtual agent?
Yes. The institute is responsible for disclosing it at the start, in line with the industry’s ethical standards and Article 50 of the EU AI Act, applicable since 2 August 2026. The notice goes in the greeting.
Are calls that nobody picks up charged?
No. Under Spoki Voice’s pay-as-you-go model, unanswered calls do not generate voice minutes. It is worth calculating the cost per completed interview, not just the theoretical cost per minute.
Do you need your own phone system or WhatsApp to get started?
No. Numbers are managed in the platform. WhatsApp, SMS and email are optional channels in the same thread.
How long does it take to test it with a real survey?
The starting point is the questionnaire link and a briefing. According to Spoki, a standard Spoki Voice agent is up and running in 1–3 days; more complex integrations can take 1–2 weeks.
Conclusion
For many opinion studies, the bottleneck is not questionnaire design but the capacity to run quality fieldwork when the client needs it, without driving up cost or loosening data standards. An AI voice agent for surveys does not replace the institute’s team: it takes over the repeatable interviews, records the data in a structured way and leaves people in charge of quality control, quotas and the difficult cases.
The way to check is a test with your own questionnaire, your own briefing and your own agent-disclosure rules. From there you can grow and complement AI-powered CATI with WhatsApp, SMS or email surveys in the same thread.
📞 Get a call from Spoki Voice: the best way to judge an AI voice agent for surveys is to hear it. Leave your number and the Spoki Voice AI agent will call you in about 60 seconds so you can test the conversation yourself, or write to Spoki to plan a fieldwork proof of concept. Get a call from Spoki Voice.





