There's a paradox at the heart of phone research. Response rates have been falling for twenty years — Pew Research has tracked telephone survey response rates dropping from the mid-30s in the late 1990s into single digits — and yet whole sectors still survey primarily by phone. UK social housing is the clearest example: the phone remains the dominant mode for collecting Tenant Satisfaction Measures. Why keep dialling if fewer people pick up?
Because the alternative is worse in a way that matters more than raw response rate: representativeness.
The mode effect nobody can wish away
How you ask changes the answer you get. The regulator's own 2024/25 tenant-satisfaction data shows the same measure — overall satisfaction — landing very differently depending on survey mode:
That ~20-point spread isn't because postal respondents are objectively more satisfied. It's because each mode reaches a different slice of people. Online skews younger and more digitally confident; it quietly under-samples older, lower-income and digitally excluded tenants — exactly the groups whose experience a satisfaction programme is supposed to capture. Phone reaches across those groups far better, which is why regulated research clings to it even as pick-up rates fall.
So where does AI voice fit?
An AI-run phone survey sits inside the same "telephone" category — same reach, same ability to get a representative spread of people — but changes two things about the economics:
- Scale. Calls run concurrently rather than one interviewer at a time, so a large sample can be fielded in days, not weeks. That matters when a deadline is fixed.
- Consistency. Every respondent hears identical wording, order and tone. There's no interviewer having a bad afternoon, and no "interviewer effect" nudging people toward polite agreement.
But — and this is the honest part — AI voice inherits phone's hardest problem and adds one of its own. It must still reach a representative sample (an AI that only gets through to people comfortable talking to a machine is no better than online-only). And it must transcribe every accent fairly; if a strong regional accent transcribes less reliably, that community is under-counted before analysis even begins.
The test that actually matters
Whether the interviewer is a person or a model, a survey is only as good as the sample behind it. The right questions to ask any provider are the same ones the regulator cares about: what's your achieved sample against the groups that matter, how do you assess representativeness, and how do you weight when the sample skews? Speed and cost are real advantages — but they're only worth having if the data stays defensible.
For the full methodology and the rules, our guides go deeper:
🎧 We talk through the survey question in our audio briefing — listen here.