AI telephone surveys for research agencies in India
Bring your approved questionnaire and your respondent list. AiSewak conducts supported telephone interviews in Indian languages and returns structured responses for your research team. We have fielded no study yet — so this page describes the offer and the division of responsibility, and contains no completion rate, no turnaround time and no client.
- 0
- Survey studies AiSewak has fielded and published. This is a new offer, not a track record.
- 0
- Respondents on an AiSewak panel. The sample is yours; we source no one.
- 10:00–21:00
- IST bands left on by default in TRAI's TCCCPR preference register
- 13 May 2027
- When the DPDP Act's consent and penalty sections commence. They are not in force today.
A research agency buying telephone fieldwork is buying a narrow, well-defined thing: interviews conducted against an instrument the agency wrote and approved, with a respondent list the agency supplied, returned as responses the agency can code, check and defend. Everything else — sampling frame, weighting, the claim the study makes — stays where it belongs, with the agency. AiSewak sells the fieldwork half and takes no position on the rest.
On the question agencies ask first and vendors usually blur: the agent conducts the interview itself. There is no human interviewer on the line. A person is involved before the call, approving the exact wording, order, skip logic and refusal handling, and after it, reviewing transcripts and auditing coded answers. That is what "supported" means here — supervised, not interviewer-assisted. If your protocol requires a trained human asking the questions, this is not that service, and you should say so early rather than discover it at pilot.
It serves market-research agencies, social-research organisations, government programme evaluators and customer-experience agencies. What follows is written as a scope document rather than a pitch, because the useful thing to publish with zero deployments behind you is precisely what you will and will not do.
Who does what, in order
Six steps. The agency owns four of them, which is the point.
- 1
You approve the instrument
The questionnaire is yours. We convert it into a call script and hand it back for sign-off on exact wording, question order, scale wording, skip logic, permitted probes and the refusal script. Nothing enters the field that you have not read in the form it will be spoken. Where translation is involved, you approve the translation and the back-translation, not a summary of them.
- 2
You supply the sample
The respondent list is yours. AiSewak operates no panel, sources no respondents and will not offer to. We need the numbers, a language field per record and — if you are managing quotas — whatever quota fields you track, because we will report against them rather than enforce them. Do not send respondent data before there is a signed scope covering it.
- 3
We configure and you listen
Language selection, voice, pacing, the AI disclosure at the top of the call, the emergency and distress break-outs, and the attempt and callback rules. Then a small set of test calls you sit through or read as transcripts. That is the go/no-go gate, and it exists so the first thing you see is not a data file.
- 4
Fieldwork runs inside a window you set
Calling hours, the number of attempts per record, how a requested callback is honoured and when a record is retired. Every attempt is logged with its outcome, including the ones that never connected — those rows are the denominator everything else is read against.
- 5
Quality checks run during, not after
A reviewed sample of transcripts while the field is open, checked against the approved instrument for wording drift and skip-logic correctness. A problem found in week one is a re-run of a few hundred records; the same problem found at delivery is the study.
- 6
You get responses, transcripts and dispositions
Coded responses, the verbatim transcript behind each one, a disposition code per record in the sample, and recordings where the respondent consented on the call. Formats are agreed in the pilot scope — this page names none, because we ship no export product and will not imply one.
What a telephone survey settles, and what it does not
A telephone survey is not automatically representative, and no delivery method makes it so. Representativeness is a property of the sampling frame, the response pattern and the weighting — all three of which sit on your side of this arrangement. A list of mobile numbers is a frame with known exclusions, and running it faster or in more languages does not close them.
Non-response is not random, and refusals at the AI disclosure are the specific version of that problem you should plan for. Some proportion of people will decline as soon as they hear the call is automated. That proportion is unknown to us, we will not guess it, and it is one of the things your pilot exists to measure with its own denominator.
Whether people answer an automated interviewer the same way they answer a human is an open methodological question, not a settled one. It has a direction of effect on sensitive items that any experienced fieldwork director will already be thinking about. We are not the right party to resolve it for you, and a vendor who tells you it makes no difference is selling rather than measuring — put a split against your existing method into the pilot if the study depends on it.
So the honest boundary is this: AiSewak is accountable for whether the approved question was asked as written, whether the branching followed the instrument, and whether the coded answer matches the transcript. Your agency remains accountable for whether the resulting dataset supports the claim the report makes.
Question types a phone interview can actually carry
Everything here is a constraint of the telephone as a medium, before it is a constraint of this system.
Single-choice closed questions
The workhorse. Options read in the approved order, repeated once on request, with the order fixed or rotated as your instrument specifies. This is the question type that travels best over a phone call in any language.
Rating and agreement scales
Workable, with the scale read out and available on request. Long scales are a memory test rather than a measurement over audio, so if your instrument uses an eleven-point scale it is worth deciding deliberately whether it survives the mode or should be collapsed — and recording that decision in the method note.
Numeric and date answers
Counts, ages, amounts, durations. Read back for confirmation where the value matters, because a misheard number is silent in a way a misheard category is not.
Open-ended responses
Captured as the respondent said them and transcribed verbatim. Coding happens afterwards against your frame, and the transcript segment stays attached to the code so you can check it. Open-ends are where an automated interview is least like a form and most like an interview.
Skip logic and branching
Conditional routing, screener terminations and per-branch wording, all taken from the approved instrument and verified against it during fieldwork. Skip-logic correctness is one of the two things we check every reviewed transcript for.
Grids, rankings and long multi-response
The hard ones. A battery of twelve items rated on the same scale, or a full rank ordering, is difficult over audio regardless of who is asking. We will tell you which parts of your instrument we think will struggle before the pilot rather than deliver weak data and let you find out.
How you audit a coded answer against the transcript
This is the question that separates a fieldwork supplier from a black box, and it has one correct answer: no coded value is delivered without the transcript segment it was derived from. If a vendor cannot show you that link for any row you pick, the coding is unauditable and the dataset cannot be defended at review. We write the link into the pilot scope as a deliverable rather than describing it as a feature.
Practically, an audit runs in three directions. Pick a coded row and read the segment behind it — does the code follow from what was said. Pick a transcript with an unusual answer and check where it landed — did an open-end that did not fit the frame get forced into a nearby code. Pick a record that terminated at a screener and confirm the branch that fired was the one your instrument specifies.
Where the respondent consented to recording, the audio sits behind the transcript for the same checks. Where they did not, the transcript is the record, and the consent status is a field you can filter on rather than an assumption you have to make.
Your own back-checks remain yours. If your protocol re-contacts a share of completes to verify the interview happened, run it — we will not substitute our review for your validation, and a supplier that offers to is offering to mark its own work.
Refusals, callbacks and calls that never connect
A completion rate without its denominator is a marketing number, so the sample rather than the connected calls is what everything is reported against. Every record in the list you supplied comes back with an outcome: never connected, unreachable number, connected and refused at the disclosure, refused part-way through, partial or dropped, callback requested, screened out, completed. Those categories are agreed with you before fieldwork so they match how your agency already reports.
A refusal is final. The agent acknowledges it, ends the call, and the record is suppressed rather than re-attempted under a different wave. A callback is honoured at the time the respondent named, in the language they were speaking, and a respondent who asks not to be called again is recorded as such permanently — that flag belongs to the person, not to the study.
Attempts are capped at a number you set and every one is logged with its timestamp and outcome, including the silent failures. Repeated attempts into a number that never answers inflate cost and nothing else, and they are the easiest thing in telephone fieldwork to leave running unnoticed.
We publish no completion rate anywhere on this site, because we do not have one. The first real figure will come from a pilot with an agency, with that agency's denominator, and it will be reported as a figure from one study rather than as a capability.
The rules the call itself has to obey
Every agent discloses that it is an automated AI service in the first sentence, in the language of the call, before a question is asked. That is a fixed property of every AiSewak deployment rather than a survey-specific setting, and it is also the point at which some respondents will decline — which is the honest cost of doing it properly.
On calling hours: TRAI's TCCCPR preference register leaves the 10:00 to 21:00 IST bands on by default, with the bands either side switched off for every subscriber regardless of registered preference, and individual subscribers may have blocked more. A scrub immediately before each wave, retained against that wave, is the only defensible practice — a list bought last month is not a scrub.
Do not assume a research exemption. Which class a study call falls into under the TCCCPR turns on the purpose of the call, and purpose can be inferred from the content rather than from what anyone intended. A survey commissioned by a government body and routed through the DLT platform, an agency's own commercial research, and a customer-satisfaction call placed on behalf of the brand that sold the product are three different situations. Settle yours with telecom counsel before the first wave rather than after the first complaint.
On data: the DPDP Act 2023 is only partly in force. Its consent, notice, retention and penalty sections commence on 13 May 2027 under G.S.R. 843(E), so the obligations most procurement checklists quote are a scoping question today and a compliance question then. Nothing on this page is legal advice and nothing here asserts that a given deployment is compliant. The sourced version of all of this, with the provisions and dates, is on the compliance reference page linked below.
Respondent data is yours throughout. It is processed for the study it was supplied for and nothing else, never reused to build a contact list, never folded into any other client's work, with retention and the deletion path named in the scope rather than left to a default.
What does not exist yet, stated plainly
Read this section before the rest, if you only read one.
No completed study
Zero fielded surveys, zero published completion rates, zero turnaround times, zero cost comparisons against your current supplier. Every number an agency would reasonably want from a fieldwork vendor is a number we do not have, and the first ones will come from a pilot with a denominator attached.
No respondent panel
We hold no respondents, recruit none and buy none. If your study needs sample sourced, that is a different supplier and we will say so rather than subcontract it quietly.
No SPSS export, API or dashboard product
None of these is built. Delivery formats are agreed in the pilot scope and the scope is what binds. A vendor listing exports it has not implemented is the specific failure this section exists to avoid repeating.
No quota management module
We can report against the quota fields in your sample. Enforcing quotas mid-field — pausing cells, rebalancing, closing strata — is not implemented, so quota control stays with your team and your existing tooling.
Branding is the script, not a portal
The agent introduces itself and names who the study is for, exactly as your approved script says. That is the whole of what "agency branding" means here. There is no white-label respondent portal and no co-branded dashboard.
The outreach dashboard is a demonstration
Our CivicReach citizen-outreach dashboard shows what campaign-level reporting looks like. The figures on it are demo data — they are not results from a completed study and must not be read as any agency's fieldwork.
Read the page →AI telephone interviewing, IVR and interviewer-led CATI
Three different methods that all reach a respondent by phone. Choosing between them is a methodological decision, not a procurement preference.
| IVR / press-1 survey | Interviewer-led CATI | AI telephone interview | |
|---|---|---|---|
| Who asks the questions | A recording. The respondent answers with the keypad. | A trained human interviewer, following the instrument. | The agent itself, speaking. No human is on the line. |
| What an open-ended answer can be | Nothing, unless the answer is recorded for transcription later. | Whatever the respondent says; the interviewer records or codes it. | Whatever the respondent says; transcribed verbatim and coded afterwards. |
| Probing and clarification | None. A misunderstood question is a wrong or missing answer. | The interviewer may probe and clarify within the approved protocol. | Scripted probes only — whatever was written into the instrument and approved. |
| Variance between interviews | None. Every respondent hears the identical recording. | Real, and managed through training, monitoring and back-checks. | None between calls, but one shared failure mode: a flaw in the script is in every interview. |
| Auditing a single answer | The keypress log. | The interviewer's entry, the recording where one exists, and re-contact back-checks. | The transcript segment the coded answer was derived from. |
| Does it make the sample representative | No. | No. | No. |
None of these columns is a claim about which method produces better data, and none contains a measured figure — AiSewak has fielded no study, so it has nothing to compare. Representativeness is a property of the frame and the weighting, never of the delivery method. The last row is identical across all three on purpose.
Call a live agent before you decide
These are running agents, not recordings. Open one, press call and speak to it in Hindi or English — the same stack that runs the deployments described above.
A structured survey call, end to end
A Hindi agent that discloses it is automated, asks three questions including an open-ended one, honours a refusal and files structured fields. Closest thing here to a survey instrument running.
Open the demo →Yojana Didi — open-ended Hindi
Unscripted Hindi conversation against a knowledge base. Useful for hearing how the stack handles an answer it did not expect.
Open the demo →Yojana Baisa — regional variants
Marwari and Mewari handling, for the question of whether an instrument survives the dialect your sample actually speaks.
Open the demo →Go deeper
AI for surveys: voice survey agents
The longer write-up of the method — where automated interviewing fits, and where it does not.
From call transcripts to structured insight
How open-ended audio becomes coded, checkable data, and what gets lost at each step.
Indian languages and the Bhashini stack
What multilingual delivery involves once you go past a language count.
DPDP, privacy and data handling
Retention, residency, access control and deletion — the parts a research contract has to name.
Related pages: What is actually published about Indic speech accuracy · Outbound calling rules, with sources · Languages and dialects · What drives the cost of a deployment · Government voice automation
Frequently asked questions
Does the AI conduct the interview, or does it assist a human interviewer?
It conducts the interview. The agent speaks to the respondent and asks the approved questions itself; there is no human interviewer on the line. People are involved before the call, approving the exact wording, order, skip logic and refusal script, and after it, reviewing transcripts and auditing coded answers against them. If your protocol requires a trained human asking the questions, this is interviewer-led CATI and it is a different service from a different supplier.
Who supplies the sample and the questionnaire?
You do, both. The instrument is yours and nothing enters the field that you have not approved in the exact form it will be spoken. The respondent list is yours — AiSewak operates no panel, recruits no respondents and buys none. If your study needs sample sourced, you need a second supplier for that and we will say so rather than arrange it quietly.
What is your completion rate, and how long does fieldwork take?
We do not have either figure. AiSewak has fielded no survey study, so there is no completion rate, no turnaround time and no cost comparison to publish, and we would rather say that than quote an industry number as if it were ours. The first real figures will come from a pilot with an agency, reported with that agency's denominator and described as one study rather than as a capability.
Will the results be representative?
Not because of anything this service does. Representativeness comes from the sampling frame, the response pattern and the weighting, all of which stay with your agency. A list of mobile numbers is a frame with known exclusions, and running it faster or in more languages does not close them. AiSewak is accountable for whether the approved question was asked as written and whether the coded answer matches the transcript; your agency stays accountable for whether the dataset supports the claim the report makes.
How is this different from an IVR survey?
An IVR survey plays a recording and collects keypresses, so an open-ended answer is either impossible or an audio file somebody transcribes later, and a respondent who misunderstands a question has no way to ask. An AI telephone interview is spoken both ways: the agent asks, the respondent answers in their own words, the answer is transcribed verbatim and coded against your frame with the transcript segment kept alongside. The trade is that probing is limited to the probes written into the approved instrument.
How do I audit a coded answer?
By reading the transcript segment it came from. No coded value is delivered without it — that is written into the pilot scope as a deliverable, because coding you cannot trace is coding you cannot defend at review. Where the respondent consented to recording, the audio sits behind the transcript for the same check. Your own back-checks stay yours: if your protocol re-contacts a share of completes, run it, because a supplier offering to replace your validation with its own review is offering to mark its own work.
What exports and data formats do you provide?
Delivery formats are agreed in the pilot scope, and this page deliberately names none. There is no SPSS export, no API and no dashboard product implemented today, so listing one would be a promise against something that does not exist. What is delivered is the substance rather than the packaging: coded responses, the verbatim transcript behind each one, a disposition code for every record in the sample you supplied, and recordings where consent was given.
Are the respondent data and the results ours?
Yes. Respondent data is processed for the study it was supplied for and nothing else — never reused to build a contact list, never folded into another client's work. Retention and the deletion path are named in the scope rather than left to a default. Note that the DPDP Act's consent, notice, retention and penalty sections do not commence until 13 May 2027, so today these are contractual commitments rather than statutory ones, and nothing here is legal advice or an assertion that a given deployment is compliant.
When can the calls be made, and does a research call need consent?
TRAI's TCCCPR preference register leaves the 10:00 to 21:00 IST bands on by default and switches off the bands either side for every subscriber, whether or not they registered a preference; individual subscribers may have blocked more, which is why the scrub has to happen immediately before each wave. Do not assume a research exemption — the class a study call falls into turns on the purpose of the call, which can be inferred from its content. An evaluation commissioned by a government body, an agency's own commercial research and a satisfaction call placed for the brand that sold the product are three different situations. Settle yours with telecom counsel before the first wave.
Which Indian languages does the voice agent speak?
Hindi plus the 22 scheduled languages, and the regional variants that decide whether a call actually lands — Bhojpuri, Awadhi, Marwari, Mewari, Magahi, Santhali and others. Agents run on a mix of Indic-first models and the MeitY Bhashini stack, and a single agent can switch language mid-call when the person on the other end does.
How is pilot pricing calculated?
Bottom-up from the things that actually consume resources, then quoted for the specific study: connected telephony minutes, which follow interview length rather than list size; the number of languages, since each one is a separate script, voice, transcript review and round of failure analysis; instrument setup and the test-call round; and the human review sample during fieldwork. There is no rate card and no per-complete list price, because a ten-minute instrument in four languages and a three-minute screener in one are not the same product. The same cost structure is set out on the cost page linked above.
Request a survey feasibility review
Send the instrument, the languages and the rough sample size, and we will tell you which parts of it we think survive a telephone interview and which do not — including the parts where the answer is that you want a human interviewer. If it looks workable, the next step is a small pilot with your denominators and your back-checks. Discuss a survey pilot with us before sending any respondent data; there is no reason for a list to move before a scope exists.
Every AiSewak agent identifies itself as an AI at the start of the call, never asks for an OTP or a payment, and honours DND. Election deployments require ECI / state CEO registration as a political advertiser.