AI Telephone Surveying: Methods & Insights

• AI Telephone Surveying is an automated survey mode that uses automatic speech recognition, large language models, and text-to-speech to conduct dynamic, voice-based interviews.
• These systems integrate precise survey control with adaptive conversational techniques, enabling clarifications, branching, and randomization to ensure methodological rigor.
• Empirical studies show that adaptive AI telephone surveys improve participation and response quality, especially in large-scale and inclusion-focused deployments.

Searching arXiv for the cited telephone surveying and conversational survey papers to ground the article in recent work. AI telephone surveying is the administration of surveys through automated voice agents that combine automatic speech recognition (ASR), LLMs, and text-to-speech (TTS) to conduct structured or semi-structured interviews over telephone channels. In contrast to conventional interactive voice response (IVR) systems, these agents can support clarifications, branching, interruptions, and limited conversational adaptation while preserving survey control requirements such as exact wording, skip logic, and randomized question order when the instrument demands it. Recent work situates AI telephone surveying at the intersection of survey methodology, conversational AI, speech technology, and human-centered design, with applications spanning quantitative panel surveys, open-ended elicitation, adaptive conversational assessment, and inclusive data collection for low-literacy populations (Leybzon et al., 23 Jul 2025, Lang et al., 27 Feb 2025, Tang et al., 31 Oct 2025, Maurya, 16 Jun 2026).

1. Concept and scope

AI telephone surveying denotes a voice-first survey mode in which an automated agent places or receives calls, asks survey questions, captures spoken answers, and advances through the instrument in real time. A common architectural description is ASR as “ears,” an LLM as “brains,” and TTS as “mouth,” enabling the system to ask questions, hear answers, and respond dynamically without relying entirely on preset scripts (Tirumala et al., 1 Sep 2025). This distinguishes the mode from IVR, which primarily plays pre-recorded prompts and records inputs with fixed logic and limited clarification capacity (Tirumala et al., 1 Sep 2025).

Current systems vary in ambition. Some are optimized for quantitative rigor, emphasizing exact wording, answer-order randomization, skip logic, and termination rules in long instruments administered to panel samples (Leybzon et al., 23 Jul 2025). Others emphasize scalable deployment under real-world conditions, integrating STT, LLM, and TTS for open-ended and closed-ended questioning at thousands-of-calls scale (Lang et al., 27 Feb 2025). A distinct line of work focuses on adaptive conversational surveys, where follow-up strategy is modified within the call according to estimated response quality (Tang et al., 31 Oct 2025). Another strand centers on inclusion and participation among low-literacy and marginalized populations through voice-first, culturally aligned, value-sensitive design (Maurya, 16 Jun 2026).

This suggests that “AI telephone surveying” is not a single method but a family of implementations occupying different points on a spectrum between strict instrument execution and adaptive interviewing. A plausible implication is that the term is best understood as a mode of survey administration rather than as a single survey methodology.

2. Technical architecture and conversational control

Most reported systems share a real-time pipeline of STT or ASR, LLM-based dialogue control, and TTS output (Lang et al., 27 Feb 2025, Leybzon et al., 23 Jul 2025, Tirumala et al., 1 Sep 2025). In one large-scale deployment, the core loop was described as STT $\rightarrow$ LLM $\rightarrow$ TTS, with direct outbound calls and WebRTC-based web calls, real-time branching, idle prompts, and silence timeouts (Lang et al., 27 Feb 2025). A quantitative-survey system for the SSRS Opinion Panel similarly integrated ASR, LLM, and TTS, with the agent reading questions exactly as scripted, handling interruptions and pauses, and clarifying ambiguous responses such as “liberal” when the intended categories were “somewhat liberal” and “very liberal” (Leybzon et al., 23 Jul 2025).

A modular formulation appears in research on conversational agents for surveys and interviews. There, the architecture includes engineered prompts, specialized knowledge bases, session variables $x = (x_1,\dots,x_K)$, and a Process Manager that orchestrates question generation, sufficiency checks, clarifications, branching, and multilingual switching under configurable parameters $\theta$ (Yu et al., 2024). The paper formalizes the first question as

$$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$

followed by a sufficiency check

$$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$

and, when needed, a clarification question

$q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$

Structured variables are then extracted as

$$x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$$

to reduce token load and support downstream branching (Yu et al., 2024).

A reinforcement-learning variant, AURA, uses a lighter tabular control mechanism rather than free-form LLM policy generation. It scores each response with a four-dimensional LSDE metric and selects among five follow-up question types using an $\epsilon$-greedy policy, updating expected values within the call (Tang et al., 31 Oct 2025). The five action types are specification, elaboration, topic probe, validation, and continuation, each intended to move different dimensions of the quality signal (Tang et al., 31 Oct 2025).

For telephone deployment, additional control layers recur across papers: silence timeouts, idle prompts, option validation, limited clarification loops, and deterministic consent gates (Lang et al., 27 Feb 2025, Leybzon et al., 23 Jul 2025, Maurya, 16 Jun 2026). Reported telephony stacks include Vapi AI with a Node.js server in one value-sensitive phone-survey system (Maurya, 16 Jun 2026), BlandAI plus GPT-4o and REDCap in a biomedical survey pipeline (Kaiyrbekov et al., 2 Apr 2025), and WebRTC plus direct outbound telephony in a large-scale multilingual deployment (Lang et al., 27 Feb 2025).

3. Measurement, response quality, and evaluation criteria

Evaluation in AI telephone surveying is heterogeneous because different projects optimize different outcomes. Quantitative-survey deployments emphasize cooperation, completion, break-off, and respondent satisfaction (Leybzon et al., 23 Jul 2025). Large-scale operational studies report fully completed and partially completed interview rates, conversation structure, and call duration (Lang et al., 27 Feb 2025). End-to-end biomedical pipelines evaluate transcript word error rate (WER) and downstream response extraction accuracy (Kaiyrbekov et al., 2 Apr 2025). Conversational and adaptive systems measure response quality more directly through content-based metrics (Tang et al., 31 Oct 2025, Xiao et al., 2019).

The standard ASR error measure is

$\mathrm{WER} = \frac{S + D + I}{N},$

where $\rightarrow$0 denotes substitutions, $\rightarrow$1 deletions, $\rightarrow$2 insertions, and $\rightarrow$3 the reference word count (Tirumala et al., 1 Sep 2025, Kaiyrbekov et al., 2 Apr 2025). In a biomedical phone-survey study, average per-line WER was $\rightarrow$4, with $\rightarrow$5 for native speakers and $\rightarrow$6 for non-native speakers, while GPT-4o extracted structured survey responses with an average accuracy of about $\rightarrow$7 from those imperfect transcripts (Kaiyrbekov et al., 2 Apr 2025). The absence of a clear correlation between WER and extraction accuracy in that study suggests that structured answer inference can remain robust when the ontology is constrained and schema validation is applied (Kaiyrbekov et al., 2 Apr 2025).

AURA introduces a turn-level response-quality model based on LSDE: Length, Self-disclosure, Emotion, and Specificity (Tang et al., 31 Oct 2025). The components are normalized to $\rightarrow$8 and aggregated as

$\rightarrow$9

Length is capped using the empirical $x = (x_1,\dots,x_K)$0 words from 467 prior responses,

$x = (x_1,\dots,x_K)$1

self-disclosure uses first-person pronoun count with empirical cap $x = (x_1,\dots,x_K)$2,

$x = (x_1,\dots,x_K)$3

emotion is the magnitude of VADER compound sentiment,

$x = (x_1,\dots,x_K)$4

and specificity is

$x = (x_1,\dots,x_K)$5

where $x = (x_1,\dots,x_K)$6 sums entity, temporal, and spatial indicators (Tang et al., 31 Oct 2025). Immediate reward is then

$x = (x_1,\dots,x_K)$7

Open-ended conversational-survey research predating recent voice systems used Gricean Maxims to operationalize informativeness, relevance, specificity, and clarity, and defined a Response Quality Index

$x = (x_1,\dots,x_K)$8

Informativeness was estimated as

$x = (x_1,\dots,x_K)$9

with $\theta$0 denoting word frequency (Xiao et al., 2019). This work was not telephone-native, but later syntheses explicitly map these quality measures to ASR+LLM+TTS telephone systems (Xiao et al., 2019).

The variety of metrics reveals a methodological split. Some studies treat the telephone AI as a survey administrator whose success is judged by cooperation and correct execution; others treat it as an interviewer whose success is judged by conversational yield. This suggests that evaluations should be aligned to survey purpose rather than standardized prematurely across all AI telephone systems.

4. Empirical findings across deployment settings

Recent evidence supports the feasibility of AI telephone surveying, but also shows clear performance differences by task type, population, and instrument design. In quantitative panel surveying, an AI interviewer deployed to the SSRS Opinion Panel administered real omnibus instruments, including a 123-question survey of approximately 30 minutes with skip logic, branching, early termination, question randomization, and answer-order randomization (Leybzon et al., 23 Jul 2025). In Wave 2 of that study, 70 of 104 adults answered the calls; of those 70, 30 completed the survey, producing a completion rate labeled COOP1 in the paper of $\theta$1. Of the 70 who answered, 29 hung up during the introduction, leaving 41 who started at least one question, which the authors reported as a continuation rate labeled COOP2 of $\theta$2. Among those who started, completion was $\theta$3. Among completions, $\theta$4 gave a neutral or positive in-survey experience rating, and $\theta$5 selected the highest rating (Leybzon et al., 23 Jul 2025).

That same study reports targeted engineering improvements between Wave 1 and Wave 2: proactive probing for ambiguous responses, ASR/comprehension improvements, elimination of TTS stuttering, reduced latency, and prompts for mutual-silence deadlocks (Leybzon et al., 23 Jul 2025). Intro completion increased from $\theta$6 to $\theta$7, and survey completion among non-HUDI starters rose from $\theta$8 to $\theta$9 (Leybzon et al., 23 Jul 2025). The embedded comparison of a 123-question long instrument and a 46-question short instrument found more positive follow-up evaluations in the short arm, with manual audits indicating no substantive behavioral differences aside from length (Leybzon et al., 23 Jul 2025).

At larger deployment scale, an LLM-based telephone survey system was tested in the United States and Peru (Lang et al., 27 Feb 2025). The Peru deployment included 2,739 total call attempts. For Peru direct outbound calls $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$0, fully completed interviews were $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$1, yielding RR1 $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$2, and partially completed interviews at $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$3 completion were $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$4, yielding RR2 $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$5. Fully completed interviews had mean duration of about $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$6, median $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$7, and range about $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$8 to $$q_{1} = \mathrm{LM}^{a}(u_{1};\,\theta_{0},\,P_{0}),$$9. Human-led administration of the same questionnaire would typically take about $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$0 minutes, which the paper interprets as evidence of less probing depth by the AI (Lang et al., 27 Feb 2025).

The same large-scale study characterizes completed conversations structurally: mean $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$1 turns, user–AI turn ratio $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$2, mean $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$3 AI turns, mean $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$4 participant turns, $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$5 AI questions, and $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$6 words per AI turn (Lang et al., 27 Feb 2025). Participant turns averaged $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$7 words overall and $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$8 words on open-ended items (Lang et al., 27 Feb 2025). These numbers indicate that the systems can sustain balanced two-party exchange over a full interview, but also that qualitative elaboration remains limited relative to human interviewing.

In healthcare-oriented phone surveying, a study of 40 surveys with 8 participants using a BlandAI phone agent and GPT-4o transcript analysis found average WER $$s_{1} = \mathrm{LM}^{s}\big(u'\,(r_{1},\,q_{1}) \,\big|\, \theta_{0},\,P_{0}\big) \in \{0,1\},$$9 and structured response extraction accuracy near $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$0, with total cost \$q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$10.75 per survey (Kaiyrbekov et al., 2 Apr 2025). Participants reported occasional interruptions, variable tone, and some roboticness, but generally judged the interaction engaging and comprehensible (Kaiyrbekov et al., 2 Apr 2025).

Telephone delivery has also been studied in inclusion-focused contexts. Among 315 adult married women without undergraduate degrees in rural Uttar Pradesh and Bihar, completion by modality improved from paper-based $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$2 and web-based $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$3 to voice (web) $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$4, voice (phone) $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$5, value-sensitive conversational AI by phone $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$6, and layered conversational AI by phone $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$7 (Maurya, 16 Jun 2026). Between-modality differences were significant, with Kruskal–Wallis $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$8, $q^{f} = \mathrm{LM}^{c}(r_{1},\,q_{1},\,u').$9, effect size $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$0, while the pairwise difference between value-sensitive convAI and layered convAI was not significant $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$1 (Maurya, 16 Jun 2026). The strongest evidence therefore concerns the shift from text-heavy modes to voice-first interaction, and from standard phone voice to value-sensitive conversational design (Maurya, 16 Jun 2026).

Finally, adaptive conversational surveying shows that within-session policy adaptation can measurably change response quality. AURA, initialized from 96 prior campus-climate conversations and evaluated on 80 simulated calls, achieved a $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$2 mean gain in response quality and significantly outperformed non-adaptive baselines with $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$3 and $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$4 (Tang et al., 31 Oct 2025). The system reduced specification prompts by $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$5 and increased validation behavior $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$6, suggesting that learned prompt selection can rebalance follow-up style in ways that better fit respondent engagement state (Tang et al., 31 Oct 2025).

5. Design paradigms: rigor, adaptivity, and value-sensitive voice interaction

Three major design paradigms can be identified in the literature.

The first is quantitative-rigor design, in which the agent’s primary responsibility is faithful survey execution. The SSRS-based AI interviewer was explicitly designed to preserve exact wording, question-order randomization, answer-order randomization, skip logic, and early termination logic (Leybzon et al., 23 Jul 2025). The large-scale Peru deployment likewise used deterministic consent and disclosure scripts, range checks for structured responses such as NPS $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$7–$x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$8, and conditional branching (Lang et al., 27 Feb 2025). The defining property of this paradigm is that conversational flexibility is permitted only insofar as it does not alter the instrument.

The second is adaptive conversational design, in which follow-up policy is optimized online. AURA represents the clearest example. It operationalizes five prompt types, discretizes respondent engagement into five states based on $x_{k} = \mathrm{LM}^{s}(y \mid u_{k}),\quad k = 1,2,\dots,K,$9 and $\epsilon$0, and uses an $\epsilon$1-greedy policy with best-performing fixed $\epsilon$2 over 10–15 turns (Tang et al., 31 Oct 2025). State assignment is defined by thresholds on current quality and improvement:

$\epsilon$3

with low if $\epsilon$4, medium if $\epsilon$5, high if $\epsilon$6, and improving if $\epsilon$7 (Tang et al., 31 Oct 2025). Expected values are initialized from priors:

$\epsilon$8

and updated session-locally by

$\epsilon$9

with $\mathrm{WER} = \frac{S + D + I}{N},$0 (Tang et al., 31 Oct 2025). This design prioritizes learned adaptivity over static script fidelity.

The third is value-sensitive and culturally aligned design, most explicitly demonstrated in low-literacy phone surveys in India (Maurya, 16 Jun 2026). That work incorporated respectful salutations, explicit consent framing, reminders that any question may be skipped or the survey discontinued at any time, slower pacing, locally appropriate tone and dialect variations, gender-matched voices, and active-listening backchanneling such as “Hmm,” “Ji,” and “Samajh gayi” (Maurya, 16 Jun 2026). The operational flow remained validator-driven, with GPT-4o-mini prompt logic determining whether spoken responses matched predefined options, but conversational phrasing was adapted to reduce perceived authority and social pressure (Maurya, 16 Jun 2026).

These paradigms are not mutually exclusive. A plausible implication is that future systems will combine quantitative-rigor constraints for core instrument integrity, adaptive control for follow-up selection, and value-sensitive surface realization for participation and trust.

6. Speech technology constraints and human factors

The feasibility of AI telephone surveying is inseparable from the limitations of conversational telephone speech technology. A foundational ASR benchmark on English conversational telephone speech reported a record $\mathrm{WER} = \frac{S + D + I}{N},$1 WER on the Switchboard subset of the Hub5 2000 evaluation set through fusion of recurrent nets with maxout activations, very deep convolutional nets with $\mathrm{WER} = \frac{S + D + I}{N},$2 kernels, and bidirectional LSTMs operating on fMLLR and i-vector features, plus advanced language-model rescoring (Saon et al., 2016). That result establishes the upper end of narrowband telephone recognition under benchmark conditions, but later work on AI interviewers notes that real-time streaming transcription can be materially worse in deployed systems (Tirumala et al., 1 Sep 2025).

A position paper on AI voice interviewers summarizes current evidence by arguing that these systems already exceed IVR capabilities for both quantitative and qualitative data collection, while emphasizing three practical limits: real-time transcription error rates, limited emotion detection, and uneven follow-up quality (Tirumala et al., 1 Sep 2025). It notes that “English word error rates hover around 5%” for state-of-the-art ASR, but that “real-time/streaming transcription error rates can be significantly higher (~10.9% on average)” (Tirumala et al., 1 Sep 2025). The paper also identifies degraded ASR performance for heavily accented speech, limited support for code-switching, and unquantified risk from background noise and device variability (Tirumala et al., 1 Sep 2025).

Emotion handling is repeatedly identified as a blind spot. When audio is reduced to text, paralinguistic information is lost unless parallel speech-emotion-recognition modules are run; existing systems yield mixed performance (Tirumala et al., 1 Sep 2025). AURA’s telephone adaptation proposal therefore extends its text-based emotion component by combining lexical sentiment with acoustic features such as pitch, energy, jitter, shimmer, speech rate, and pause length in a fused score

$\mathrm{WER} = \frac{S + D + I}{N},$3

with $\mathrm{WER} = \frac{S + D + I}{N},$4 tuned offline (Tang et al., 31 Oct 2025). That formulation remains prospective rather than field-validated, but it captures the direction of current research (Tang et al., 31 Oct 2025).

Turn-taking and latency are also central. The large Peru deployment relied on fixed maximum AI turn length, idle prompts, and silence timeouts (Lang et al., 27 Feb 2025). The SSRS system added prompts to overcome mutual-silence deadlocks and reported smoother interaction after latency and voice-stability fixes (Leybzon et al., 23 Jul 2025). The fitness-for-purpose review recommends end-to-end turn latency targets of at most $\mathrm{WER} = \frac{S + D + I}{N},$5–$\mathrm{WER} = \frac{S + D + I}{N},$6 ms for natural flow and a hard cap of $\mathrm{WER} = \frac{S + D + I}{N},$7 s, while noting that respondents may have lower tolerance for latency when they know they are speaking with an AI (Tirumala et al., 1 Sep 2025). Not all implementation papers report these metrics, however, which limits cross-study comparability.

Human factors do not reduce to technical fidelity. In the low-literacy India study, slower pacing, skip reminders, culturally aligned voices, and familiar conversational rhythms were linked to improved completion (Maurya, 16 Jun 2026). In the SSRS pilots, shorter instruments were associated with more positive judgments of naturalness and understanding (Leybzon et al., 23 Jul 2025). In the biomedical study, some participants felt unable to revise earlier answers once the AI had moved on, even though they might have attempted revisions with a human interviewer (Kaiyrbekov et al., 2 Apr 2025). These observations suggest that perceived agency, repair affordances, and local social norms are methodological variables, not merely interface preferences.

7. Ethics, privacy, limitations, and research directions

Ethical and methodological controversies around AI telephone surveying concern disclosure, consent, privacy, data quality, and fairness. Explicit AI disclosure varies across studies. The SSRS pilot informed respondents in the introduction that they were speaking with an AI and could terminate (Leybzon et al., 23 Jul 2025). The Peru deployment used deterministic consent and disclosure scripts specifying identity, sponsor, purpose, duration, and data use (Lang et al., 27 Feb 2025). By contrast, the low-literacy India study notes that the AI nature of the system was not disclosed to participants and identifies this as a limitation, with future work planned to compare disclosed and non-disclosed conditions (Maurya, 16 Jun 2026).

Privacy architectures also differ. The modular conversational-agent framework routes participant utterances first to an on-premise LLM for identity screening before forwarding only non-sensitive content to online LLMs, with optional PII collection, anonymization, and reduced location granularity (Yu et al., 2024). AURA resets its expected-value table to priors at the start of each new call and explicitly retains no cross-user state, framing this as privacy-preserving session-local learning (Tang et al., 31 Oct 2025). Biomedical deployment work used fictitious personas to avoid storing PHI or PII in the study dataset, with recordings stored on the Bland platform and accessible only to the study team (Kaiyrbekov et al., 2 Apr 2025).

A persistent misconception is that higher completion implies higher data quality. The low-literacy India study explicitly cautions that it measured completion and retention only, not response validity, satisficing, or social desirability bias (Maurya, 16 Jun 2026). The large-scale Peru deployment likewise did not include a human-enumerator control arm and did not report formal agreement metrics, item nonresponse rates, or STT WER (Lang et al., 27 Feb 2025). Even the SSRS panel study, while methodologically focused, reports descriptive rates and experience data rather than formal error comparisons to human interviewers (Leybzon et al., 23 Jul 2025). Thus, current evidence is strongest for feasibility, participation, and structured-item handling, and weaker for causal claims about substantive measurement equivalence.

Fairness concerns arise primarily through speech recognition and interactional style. ASR accuracy may degrade for heavily accented speech (Tirumala et al., 1 Sep 2025), while prosodic emotion models may misread affect across accents or speech conditions (Tang et al., 31 Oct 2025). AURA therefore recommends calibration sets across demographics, differential-performance monitoring, and possibly contextual bandits conditioned on non-sensitive, consented context rather than protected attributes (Tang et al., 31 Oct 2025). Inclusion-focused work similarly recommends local stakeholder consultation to select voices, forms of address, and dialectal cues (Maurya, 16 Jun 2026).

Research directions in the literature converge on several fronts. One is richer state and policy modeling, including contextual bandits, Thompson sampling, LinUCB variants, turn count, recent action history, ASR confidence, and risk-sensitive action selection for short calls (Tang et al., 31 Oct 2025). Another is improved multilingual and dialectal support, especially under telephony noise (Yu et al., 2024, Lang et al., 27 Feb 2025). A third is stronger evaluation: randomized comparisons to human interviewers, measurement of response validity and social desirability, formal reporting of WER and latency, and subgroup analyses for representativeness and fairness (Tirumala et al., 1 Sep 2025, Maurya, 16 Jun 2026, Leybzon et al., 23 Jul 2025). A fourth is tighter integration of privacy gating, structured-variable storage, and human escalation protocols (Yu et al., 2024).

Taken together, the literature indicates that AI telephone surveying is already viable for structured quantitative data collection and increasingly capable in semi-structured interviewing, but remains constrained in emotion-sensitive, high-depth qualitative interviewing. The strongest present evidence supports its use where conversational flexibility improves over IVR yet full human probing is not indispensable (Tirumala et al., 1 Sep 2025). The open problem is not whether AI can administer surveys by phone, but under what design constraints, populations, and measurement goals it can do so without compromising methodological rigor.