What Is CATI and What Is Replacing It (2026)

Last updated: 19 August 2026

Computer assisted telephone interviewing, or CATI, is a survey method in which a human interviewer calls a respondent and reads questions from a script displayed on screen, entering answers as the conversation goes. The software handles routing, quotas and validation while the person handles the conversation.

It is being replaced not because telephones stopped working, but because of who stopped answering them. Pew Research Center's own telephone response rates fell to 6% by 2018, down from roughly 9% a few years earlier. A method that once reached a cross-section of a country was, within a decade, reaching whoever still picked up an unknown number.

This piece covers what CATI is, why it declined, what is actually replacing it, and where the evidence on AI-administered calling currently stands.

What Is Computer Assisted Telephone Interviewing?

CATI is a telephone survey run by a human interviewer working from an on-screen script, with software controlling question order, skip logic and quotas. It replaced paper-and-pencil telephone interviewing in the 1970s and 1980s and remains a recognized mode in professional practice.

The defining feature is the split of labor. The computer enforces the instrument. The person conducts the conversation, handles clarification requests, and judges when an answer is complete.

That split is why CATI has held on longer than its response rates suggest it should have. For questionnaires that need branching logic and consistent administration across hundreds of respondents, it works, and the standards bodies still treat it as a legitimate mode. The AAPOR standards address telephone administration directly, and the ESOMAR code and guidelines govern consent and participant welfare regardless of how the call is placed.

Why Did CATI Decline?

Because coverage collapsed while the method stayed the same. Caller ID, mobile substitution, spam filtering and sheer call volume changed who was willing to answer, and the people who remained willing stopped resembling the wider population.

Pew's figure is the cleanest public marker: 6% by 2018. The industry did not fail on craft. Interviewer training, questionnaire design and dialing technology all kept improving through the same period. The mechanics of that collapse are traced in why phone survey response rates collapsed.

What broke was the sampling frame. A 6% response rate is not fatal on its own, since low response does not automatically mean bias. It becomes fatal when the people who answer differ systematically from the people who do not, on exactly the variables a study is trying to measure.

Cost followed coverage. Reaching a completed interview at a 6% response rate means placing many times more calls, and every one of those is a paid interviewer minute. The economics of a human-dialed telephone study got worse in direct proportion to the coverage problem.

What Is Replacing It?

Four things, and they are not interchangeable. Most teams that have moved off CATI have not moved to one successor, they have split a single study across two or three.

Read the last two columns together. Traditional CATI is still the right answer for a regulated study, a very small specialist sample, or a topic where a respondent may need a person on the line. It is the wrong answer mainly when the study needs volume and the budget assumed 1990s response rates.

The vendor landscape has split along the same line. Voxco and IdSurvey remain the established CATI platforms built around human interviewer workflows, with call scheduling, supervisor monitoring and quota management as first-class features. UserCall, Voiceter and VeriCATI are the newer AI-native entrants, where the voice agent conducts the interview and the software handles transcription and coding as fieldwork runs. The two groups solve different problems, and a team replacing a human phone room is buying from the second group whether or not it uses the word CATI.

How Does an AI Phone Interview Actually Work?

An automated voice agent places the call, states that the interviewer is automated, captures consent, then administers the questionnaire conversationally. It handles branching and asks for clarification when an answer is too short. Responses are transcribed and coded as the fieldwork runs rather than afterward.

The operational differences from CATI are concrete rather than atmospheric. There is no interviewer roster to recruit, train and schedule, so a study is not capped by headcount. Calls run in parallel. Because there is no shift pattern, the constraint moves from staffing to compliance windows and respondent availability.

Compliance is where the detail lives, and it is jurisdiction-specific. In the United States the governing instrument is the Telephone Consumer Protection Act, and the FCC confirmed in February 2024 that an AI-generated voice counts as an artificial voice under it, so the call needs prior express consent. Recording consent is separate again and set by state law. Alchemic's AI phone research runs on that pattern, with consent captured on the first turn and every call recorded and searchable. The full rule map is set out in TCPA rules for AI phone research calls.

Who Can a Phone Study Reach That Others Cannot?

Anyone with a working number, including people on feature phones, shared devices and unreliable connections. That is a materially wider frame than any browser-based method, and in several large markets it is the difference between a national sample and an urban one.

The ITU's Facts and Figures 2025 reports mobile broadband coverage as nearly universal while quality and affordability gaps persist, and counts 2.2 billion people still offline, most of them in low and middle income countries. DataReportal's Digital 2026 report counts more than 6 billion people online, which sounds like the problem is solved until you separate being online from being reachable for a scheduled thirty-minute video session.

A voice call needs neither bandwidth nor a private room with good lighting. It needs a signal and a few minutes. For rural respondents, older respondents and lower-income households, that is often the only mode that reaches them at all.

Which Languages Can the Study Actually Run In?

Language is the second half of that reach, and the half most often skipped at design stage. A phone study is only as wide as the languages it can be conducted in. In multilingual markets, the respondents hardest to reach by link are also the least likely to be interviewed comfortably in English.

Alchemic publishes 57+ languages including Spanish, Hindi, Tamil, Bangla, Arabic and Indonesian. Managed fieldwork runs across fourteen markets, including the USA and the UK alongside South and Southeast Asia, the Gulf and Africa.

That combination matters most when a study crosses borders. A single instrument fielded across several countries has to hold the questionnaire constant while the language, the calling window and the incentive mechanism all change. That is a fieldwork problem rather than a software one, and it is worth scoping before entering an unfamiliar market.

What Does Phone Reach Cost You?

The tradeoff is real and worth stating plainly. A phone interview carries no visual stimulus, so packaging, creative and interface work cannot run this way. What phone buys is frame width, not richness.

Where a study needs both, mode-mixing is the usual answer: phone for the segments a link cannot reach, and messaging or browser for the rest. Interviews conducted inside WhatsApp sit between the two, asynchronous and without an app install, and a managed research program can run several modes against one instrument rather than treating each as a separate procurement.

What Does the Evidence on AI Calling Actually Say?

Early, encouraging and thin. The most substantial public test to date deployed an LLM-based telephone survey system across two populations, a pilot in the United States with 75 participants and a large-scale deployment in Peru with 2,739. The agent administered open and closed questions, handled clarifications and navigated branching logic without an interviewer roster.

That is one system, two countries, and a research paper rather than an industry norm. It is enough to establish that the method functions at scale. It is not enough to establish equivalence with human interviewing across topics, populations and question types.

Institutional work is now underway to close that gap. Gallup has begun publishing research on AI phone interviewing as a methodological question rather than a product launch, which is the kind of independent evidence the category has been missing.

Adjacent findings are worth reading alongside it. The Nielsen Norman Group's January 2026 test of AI interviewers, ten participants across two platforms, found the two platforms it tested followed the script rather than the insight, sticking to the guide and declining to reframe weak questions. That limitation is not specific to voice, but it applies to voice.

Mode effects also do not disappear because the interviewer is synthetic. Pew's work on mode of interview effects shows respondents present themselves more favorably to an interviewer, which is one reason removing the person changes answers on sensitive subjects rather than leaving them untouched. The size of that shift is topic-dependent, and Pew separately found few mode effects when asking about news consumption habits in its own 2017 analysis.

Where CATI and AI Calling Both Fall Short

Neither is a general-purpose replacement for the other, and the honest limits are easy to state.

• Response rates do not improve because the caller is automated. An unknown number is still an unknown number. Automation changes the cost of dialing, not the willingness to answer.
• No visual stimulus, in either mode. Concept boards, packaging and interfaces cannot be tested by voice. That is a hard constraint, not a roadmap item.
• Nonverbal signal is absent from AI calling entirely. A human interviewer hears hesitation and can decide to slow down. An automated agent has no equivalent judgment.
• Sensitive topics still need a person. Where distress is foreseeable, a system that cannot recognize distress should not be the only thing on the call.
• The evidence base is young. One large deployment and early institutional work is a starting point, not a settled literature. Treat vendor equivalence claims accordingly.
• Compliance is local and unforgiving. Calling windows, do-not-call registries and consent requirements differ by country, and getting them wrong is a legal problem rather than a data-quality one.

The Insights Association publishes complementary standards on disclosure and data quality.

How Do You Choose Between Them?

Start from who you need to reach, not from what the method costs. Coverage failures are invisible in a completion report, while cost overruns announce themselves immediately, which is why teams systematically over-weight the second.

Three questions settle most cases:

• Can your target segment be reached by a link at all? If a meaningful share cannot, a browser-only design has already decided your findings before fielding starts.
• Does the instrument need visual stimulus? If yes, voice is out for that section, whoever administers it.
• Could a respondent become distressed? If yes, route that portion to a human, regardless of budget.

Everything after those three is an economics question rather than a methods question. Whether an automated agent, a human interviewer or a broader AI-moderated design fits best depends on volume, languages and how much of the fieldwork you want to run yourself.