Is CATI the same as an AI interviewer?

No; CATI commonly supports a human interviewer with software.

Does random dialing guarantee representative results?

No; coverage, selection within contacted units, and nonresponse still matter.

Can telephone and web responses be combined?

Yes, with an analysis that considers design and possible mode differences.

AI Glossary / Statistics and probability / Intermediate

Telephone survey

Definition

A telephone survey collects information from respondents through a telephone interview or an automated telephone interaction. It is a mode of data collection, not a sampling design by itself. Its validity depends on who can be reached, how respondents are selected, how questions are administered, and how the collected responses are analyzed.

What it means

Using the phone can let an interviewer clarify instructions and guide respondents through a questionnaire. It can also exclude people who are not covered by the contact frame, cannot be reached, or cannot participate comfortably in that mode.
A telephone survey may use a customer list, an address-based sample with telephone follow-up, or a telephone-number sampling approach. These designs imply different selection probabilities and coverage limitations. Calling many numbers does not automatically produce a representative sample.

How it works

Define the target population and unit of analysis. A phone number may belong to a household, individual, shared account, or business switchboard. The person who answers is not necessarily the intended or randomly selected respondent.
Choose a sampling frame and selection process. Check whether relevant people lack listed numbers, have multiple numbers, or share numbers. Multiplicity and incomplete coverage can change the probability of inclusion and the interpretation of responses.
Create a questionnaire suited to listening. Long response lists, complex tables, and visually dependent tasks may be difficult over the phone. Use clear wording and test the order, length, and routing with realistic respondents.
In computer-assisted telephone interviewing, often abbreviated CATI, software presents questions, controls routing, and records responses. This can improve consistency, but incorrect programmed logic can also propagate errors across interviews. Test the instrument before collection.
Specify contact and follow-up procedures. Calls at only one time of day can systematically miss some groups. Track attempts and dispositions, distinguishing no contact, refusal, ineligibility, partial completion, and completed interview.
Train interviewers to follow the script and use neutral clarification. Their tone, prompting, and presence can affect answers, especially for sensitive questions. Automated systems introduce different usability and measurement considerations rather than eliminating all mode effects.
Document the treatment of unknown eligibility and missing answers. A reported response rate depends on its denominator and formal definition; completed interviews divided by every dial attempt is not a universally meaningful response-rate calculation.
Analyze the results with the actual sampling design and justified adjustments. Nonresponse weighting can help only to the extent supported by auxiliary information and assumptions. A telephone mode does not itself justify a simple-random-sample margin of error.
If AI assists transcription or coding, validate those outputs across relevant accents, languages, and response types. Preserve uncertainty and the original meaning of answers rather than treating an automated transcript as error-free measurement. Follow the project's authorized privacy and recording procedures.

Three practical examples

Customer research: interview a selected sample about their recent service experience.
Official statistics: conduct telephone follow-up within a broader survey design.
AI product feedback: gather structured user experiences while checking interviewer and transcription effects.

Think of it this way

Think of the telephone as the doorway through which the survey reaches people. A convenient doorway can improve conversation, but it does not tell you whether everyone had a chance to enter or whether those who entered resemble those who did not.

Worked example

Consider an original simplified survey with 800 selected eligible customers, each associated with one valid contact number. Suppose eligibility is known for all 800, and the collection ends with 280 completed interviews and 520 eligible nonresponses. There are no partial interviews in this toy setup.
The simple completed-interview share among selected eligible customers is 280/800 = 35%. Among the 280 respondents, 168 report using an AI feature, so the respondent usage proportion is 168/280 = 60%.
The 35% figure describes participation under the stated simplified denominator. The 60% figure describes reported usage among respondents. Neither number by itself establishes usage among all 800 selected customers or the full customer population.
If nonrespondents are less likely to use the feature, the respondent percentage can overstate population use. Repeated call attempts to the same customer do not increase the number of independently selected units. More elaborate real surveys require explicit disposition and response-rate rules rather than importing this simplified calculation unchanged.

Illustrative case study

Fictional case: Willow Support runs a telephone survey about an AI assistant. Calls initially occur only during office hours, and many working customers are never reached. Analysts broaden contact times, track reasons for nonresponse, and compare respondent characteristics with the sampling frame. They also audit a sample of automated transcripts against the interviews. The final report identifies the collection mode and response limitations rather than presenting the respondent percentage as an unquestioned population fact.

Common misconceptions and mistakes

Do not confuse telephone collection with random sampling. The number of dial attempts is not the number of respondents. A completed-interview percentage is not necessarily a standardized response rate. Interviewers can introduce measurement effects even when they follow a structured instrument. Multiple telephone numbers and shared numbers can complicate selection. Automated interviewing does not remove coverage or nonresponse problems.

Business use

Use telephone surveys when spoken interaction and interviewer support fit the population and questions. Compare reach, respondent burden, measurement quality, and cost with other modes. Retain contact dispositions and design information so the resulting estimates can be evaluated and reproduced rather than reduced to a count of completed calls.

Check your understanding

What does 35% measure in the example?

Answer: Completed interviews among the 800 known eligible selected customers.

What does 60% measure?

Answer: Reported feature use among respondents.

Would extra calls to the same customers increase independent sample size?

Answer: No.

Related terms

Nonresponse
Coverage error
Survey mode
CATI
sampling frame
interviewer effect
response rate
mixed-mode survey

Learn next

Study survey contact dispositions and questionnaire design for spoken administration. Then compare mode effects and selection probabilities across telephone, web, and in-person collection.

Sources

U.S. Census Bureau, Developing CAPI and CATI Design Guidelines: https://www.census.gov/library/working-papers/2023/adrm/rsm2023-09.html U.S. Census Bureau, CPS Collecting Data: https://www.census.gov/programs-surveys/cps/technical-documentation/methodology/collecting-data.html Australian Bureau of Statistics, Types of error: https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/types-error The telephone-survey counts are original and simplified.

Previous: Teleoperation Next: Temperature