AI vs human interviewers: what actually differs on a survey call

The question research teams ask about AI phone surveys is usually framed as "is it as good as a human interviewer?" That's the wrong frame. The honest one is: what changes when the interviewer is a model instead of a person — because some of those changes help your data, and some hurt it.

The interviewer effect cuts both ways

People are polite. Faced with a warm human voice asking "how satisfied are you?", respondents nudge their answers upward — they don't want to disappoint the person on the line. This "interviewer effect" is well documented, and it's visible in the regulator's own tenant-satisfaction data, where staff-administered modes tend to return higher scores than self-completion.

An AI interviewer largely removes that social pressure. That's genuinely useful if you want a truer read — but it also removes the warmth that keeps some respondents on the call. Neither is strictly better; they're different biases, and the right choice depends on what you're measuring.

<rect x="20" y="110" width="680" height="40" fill="#fff" stroke="#E2E8F0"/>
<text x="34" y="134" font-weight="600" fill="#0A1628">Consistency</text>
<text x="276" y="134">Varies by person &amp; fatigue</text>
<text x="497" y="134">Identical wording every call</text>

<rect x="20" y="150" width="680" height="40" fill="#F8FAFC" stroke="#E2E8F0"/>
<text x="34" y="174" font-weight="600" fill="#0A1628">Cost &amp; scale</text>
<text x="276" y="174">£15–£30 per complete</text>
<text x="497" y="174">Concurrent; a fraction of that</text>

<rect x="20" y="190" width="680" height="40" fill="#fff" stroke="#E2E8F0"/>
<text x="34" y="214" font-weight="600" fill="#0A1628">Probing &amp; empathy</text>
<text x="276" y="214">Strong — reads the room</text>
<text x="497" y="214">Improving, still narrower</text>

<rect x="20" y="230" width="680" height="40" fill="#F8FAFC" stroke="#E2E8F0"/>
<text x="34" y="254" font-weight="600" fill="#0A1628">Accents &amp; reach</text>
<text x="276" y="254">Adapts naturally</text>
<text x="497" y="254">Good, but must be checked</text>

Where AI clearly helps

Consistency. Every respondent hears the exact same question, in the same order, in the same tone. For regulated measures where wording is prescribed, that standardisation is a real quality gain — no interviewer improvising, no drift across a long shift.

Cost and speed. Human telephone interviewing runs roughly £15–£30 per completed response once quotas are met; AI-phone fieldwork is a fraction of that, and calls run concurrently rather than one interviewer at a time. A large sample can be fielded in days.

Where a human still wins

Be honest about the limits. A skilled interviewer reads confusion and rephrases on the fly, handles a distressed or sensitive respondent with genuine judgement, and probes an open-ended answer in a way current AI does less well. For deep qualitative work, or emotionally charged subjects, a person is still the right call.

The test that survives either way

Whichever interviewer you choose, a survey is only as good as its sample. The AI must still reach a representative spread of people — every accent and age group, not just the digitally confident — or it inherits the same bias as an online-only survey. Ask any provider the same questions the regulator cares about: achieved sample against the groups that matter, how representativeness is assessed, and how the data is weighted.

For the full methodology, see Can an AI run a survey people finish? and the complete TSM guide.

🎧 Listen to the survey briefing above.