Last updated: 4 September 2026
An automated research call in the United States sits under three separate rule sets at once, and clearing one does not clear the others. Federal telecom law governs whether the call may be placed at all. Federal consumer protection law governs whether what you are doing counts as telemarketing. State law governs whether you may record it.
Most teams discover this late. The fieldwork is designed, the vendor is chosen, and only at launch does someone ask whether the dialing schedule, the consent script and the recording policy are vendor preferences or legal requirements. They are requirements, and getting them wrong produces a legal problem rather than a data quality problem.
This is a practitioner's map of what applies, not legal advice. The specifics turn on how your study is structured and who you are calling, so the citations below are to primary sources and the final read belongs with counsel.
Does the TCPA Apply to an AI Research Call?
Yes. The Federal Communications Commission settled the central question in a Declaratory Ruling adopted on 2 February 2024 and released on 8 February, confirming that the Telephone Consumer Protection Act applies to AI technologies that generate human voices. The ruling carries the reference FCC-24-17 in docket 23-362, and its language is unambiguous: the restrictions on "artificial or prerecorded voice" encompass current AI technologies that generate human voices, so calls using them require the prior express consent of the called party. An AI moderator speaking in a synthesized human voice is an artificial voice for the purposes of the statute, and the Commission made the same point again in ruling that AI-generated voices in robocalls are illegal where the underlying call is unlawful.
That ruling matters more than any vendor feature. It means an AI phone interview is not a new category awaiting regulation. It is an existing regulated category, and the fact that a human never speaks does not soften the obligations. The governing text sits at 47 U.S. Code section 227, with the operative delivery restrictions at 47 CFR 64.1200.
The practical consequence is that "we use AI" is never an answer to a compliance question. It is the reason the question is being asked.
What Consent Does an AI Voice Call Need?
The answer splits on two axes: what kind of line you are calling, and whether the call is telemarketing.
For calls to wireless numbers, an artificial or prerecorded voice requires prior express consent, and the FCC's own consumer guidance on unwanted robocalls and texts states the general rule plainly. Most US consumer research today is reaching mobile phones, so this is the binding case rather than the exception.
If the call is telemarketing, the standard rises to prior express written consent. A genuine research interview that sells nothing is not telemarketing merely because it is automated, but that is a conclusion about the content of your call, not a status you can assert.
Two things follow for study design. Consent has to exist before the call is placed, which makes it a sampling and recruitment question rather than a script question. And the consent you hold has to cover this call: permission captured for one purpose does not automatically extend to an unrelated study months later.
Is Survey Research Exempt From the Do Not Call Registry?
Partly, and the distinction is the one most often got wrong.
The Do Not Call Registry and the Telemarketing Sales Rule govern telemarketing, which turns on inducing the purchase of goods or services. Pure survey and opinion research generally falls outside that definition, which is why research calls are not treated the same way as sales calls under those rules.
That exemption is narrower than it sounds, for two reasons.
First, it says nothing about the TCPA. A research call using an artificial voice to a cell phone still needs consent under section 227 whether or not the Do Not Call Registry applies to it. Teams routinely conflate the two and conclude that research is exempt from everything.
Second, the exemption evaporates the moment the call does sales work. The FTC's compliance guide to the Telemarketing Sales Rule is the reference for where that boundary sits, and the Commission reaffirmed its protections against AI-enabled telemarketing fraud in 2024.
When Does a Survey Stop Being Research?
The industry has names for this because it has been a problem for decades. Selling under the guise of research is sugging; fundraising under the guise of research is frugging. Both convert a research call into something else, and the legal exposure follows the substance rather than the label.
The test is not what the script is called internally. It is whether the call is trying to produce a sale, a lead or a donation. Lead generation dressed as a survey is telemarketing that has mislabeled itself, and a screener that routes qualified respondents to a sales team is the clearest version of this.
Professional codes address the same boundary independently of the statutes. The ESOMAR code and guidelines and the Insights Association both treat the separation of research from selling as fundamental rather than advisory, and clients increasingly enforce it through contract whatever the regulator does.
What Must a Respondent Be Told, and When?
| Obligation | What it requires in practice | Where it comes from | Failure mode |
|---|---|---|---|
| Consent before dialing | Prior express consent on file for artificial-voice calls to wireless numbers | TCPA, 47 USC 227 and 47 CFR 64.1200 | Calls placed with no lawful basis |
| Higher consent if selling | Prior express written consent where the call is telemarketing | TCPA and the Telemarketing Sales Rule | A research exemption claimed for a sales call |
| Automated-agent disclosure | State that the interviewer is not human, before substantive questions | Professional codes and emerging practice | Respondent misled about who is asking |
| Recording consent | One-party or all-party consent depending on the state | State wiretap and eavesdropping statutes | An unusable recording, and potential liability |
| Identification of the caller | Who is calling and on whose behalf | FCC rules and research codes | Call reads as a scam and the response rate collapses |
| Retention and deletion | A stated period, applied to derived artifacts as well as audio | State privacy laws such as CCPA and CPRA | Indefinite storage by default |
Read the last column as the audit list. Every one of these fails quietly rather than loudly, which is why none of them surfaces until somebody asks.
What About Recording the Call?
This is where a single national policy stops working, because recording consent is state law rather than federal.
Most states follow a one-party consent rule, where the party recording the conversation may consent to it. A significant minority require all-party consent, and calls that cross state lines can implicate the stricter rule at either end. That variation is the reason a compliant operation captures recording consent explicitly on every call rather than maintaining a state-by-state matrix and hoping the routing is right.
The safe operating posture is the one that also produces better research: disclose that the call is recorded, capture agreement on the first turn, and make the recording obvious rather than incidental. A respondent who knows they are recorded and agreed to it is not a compliance risk, and a recording obtained otherwise may be unusable as evidence for the finding it supports.
State privacy statutes then govern what happens next. Under California's regime and the state laws modeled on it, a voice recording and a phone number are both personal information, which brings purpose limitation, retention limits and deletion rights into scope after the interview is over.
Who Can a Compliant Phone Study Actually Reach?
Anyone with a working number whose consent is properly held, including people on basic handsets and unreliable connections. That frame is materially wider than any browser-based method, and compliance is what keeps it usable rather than what narrows it.
Pew Research Center's mobile technology fact sheet reports 16 percent of US adults as smartphone-only internet users, rising to 34 percent in households under $30,000 a year against 4 percent above $100,000. The ITU's Facts and Figures 2025 reports mobile broadband coverage as nearly universal while counting 2.2 billion people still offline. A voice call needs a signal and a few minutes. A scheduled video session needs considerably more.
Alchemic runs AI phone interviews on that pattern, capturing consent on the first turn, with every call recorded and searchable. It publishes 57+ languages including Spanish, Hindi, Tamil, Arabic and Indonesian, built and tested with native speakers rather than routed through a translation layer. Recruitment runs as managed fieldwork or bring your own, across fourteen markets that include the USA and the UK.
Knowing which consent language a population actually understands, and what incentive framing reads as coercive locally, is operational knowledge rather than a software feature. That is why it usually sits with a team running the study rather than with the buyer. How that reach compares against the alternatives is covered in AI phone interviews against IVR and voice bots.
What Changes When the Study Crosses Borders?
Each additional country brings a complete new rule set rather than a variation on the last one. A four-country study inherits four consent regimes, four calling-window rules and four preference registers.
This is a scoping problem, not an operations problem, and it is cheapest to solve before the questionnaire is approved. Three questions settle most of it. Which markets are in scope and what does each require for automated outreach? Can one consent script be adapted, or does each market need its own? And do the permitted calling hours leave enough time to hit the quota?
Where a market is genuinely unfamiliar, compliance gaps and sampling gaps tend to appear together, which is why teams usually treat entering a new market as fieldwork rather than as a software purchase.
Does the Same Apply to Messaging Research?
Partly, and the differences matter. Messaging platforms carry their own rules for business-initiated contact, separate from telecom law, and consent captured for a phone call does not automatically extend to a chat thread.
Privacy obligations do carry across unchanged, because a transcript is personal information in whatever format it arrives. What changes is the rule governing first contact, which determines whether a registry or a platform policy applies to the opening message.
Asynchronous formats also move the consent moment. A respondent answering an interview inside WhatsApp across several hours has more opportunity to reconsider than someone on a live call, which argues for making withdrawal genuinely easy rather than nominally available. The mechanics of that are covered in recruiting and consenting respondents on WhatsApp.
Whatever the channel, the transparency duty attaches to the AI-moderated interview itself rather than to the pipe it travels down.
Where Compliance Gets Genuinely Hard
A checklist that implies the problem is solved is worse than an honest list of what stays difficult.
- Rules differ by state, and recording is the sharpest edge. All-party consent states and interstate calls between them are the most common trap in an otherwise clean operation.
- Consent quality is not visible on a dashboard. A completed consent step records that the step completed, not that anything was understood.
- Retention is easy to state and hard to enforce. Deleting audio is trivial. Deleting every transcript, embedding and derived artifact is a systems question.
- Disclosure wording has no settled standard. Professional bodies are actively revising guidance for AI-moderated methods, so a script that was defensible last year is worth re-reading.
- Distress has no automated handler. A voice agent cannot recognize that a respondent should be routed to a person, which constrains which topics belong on an automated call at all.
- Compliance does not fix coverage. A perfectly lawful study can still reach the wrong slice of a market, and that failure never appears in a completion report.
The AAPOR standards and ethics materials treat disclosure and data handling as components of survey quality rather than as administrative overhead, which is the right frame for all six.
How Do You Audit a Vendor Before Signing?
Ask for artifacts rather than assurances. The difference between a document a vendor can produce on the call and a claim they can make on the call is the whole audit.
- Ask for the disclosure script, verbatim, and the exact point in the flow where it plays.
- Ask how consent is evidenced for each number in the sample, and how close to dial time it is checked.
- Ask which states the recording policy assumes, and what happens on an interstate call.
- Ask for the retention period in writing, and what deletion covers beyond the audio file.
- Ask whether customer data trains any model, and accept only a yes or a no.
- Ask what happens if a respondent becomes distressed, and treat "nothing" as an answer about which topics you can field.
Record the answers in the method note. Consent procedures are among the few things in research that a client, a regulator or an ethics reviewer may ask about years later, and reconstructing them from memory is an uncomfortable exercise. Vendors that also run the interviews rather than licensing a tool are covered in platforms that run the study for you.