AI-run focus groups arrive in political polling — and UK May elections are the test
TL;DR: French start-up Naratis is running qualitative political polls with AI agents conducting one-on-one interviews — one agent asks, others check the response is on-topic, probe for depth, and verify the respondent isn’t a bot. Founder Pierre Fontaine claims the method is “10 times faster, 10 times cheaper and 90% as accurate” than human-led qualitative work. The shift lands as global polling response rates have fallen from over 30% in the 1990s to below 5% today, eroding the economics of traditional research and pushing the industry to automate.
Industry leaders move carefully
Established firms are integrating AI but drawing harder lines around political work. Ipsos uses AI extensively in market research — including video analysis of self-recorded behaviour and “digital twin” experiments — but does not use AI-generated respondents in political surveys. OpinionWay chief executive Bruno Jeanbart told the BBC his firm “would never publish an opinion poll based on AI-generated data” because of trust concerns. The risks are well-rehearsed: hallucination, “common sense” responses that reflect what people typically say rather than what individual respondents think, and the deeper question of what is actually being measured when synthetic data substitutes real respondents.
For the UK, the timing matters. Scotland, Wales and English local elections are scheduled for May 2026, and the ICO has already flagged elections-period AI use as a supervisory priority. Naratis is the first vendor publicly running AI-led qualitative work for political clients; UK pollsters and campaigns will be watching whether its results in French polling cycles hold up against post-vote ground truth before adopting the technique.
Looking forward
The strategic split inside the polling industry is now between firms that use AI to listen better — analysing real-respondent video, social-media patterns, and scaling qualitative research to populations previously too small to study — and firms that go further by allowing synthetic respondents into the dataset. UK firms have so far stayed on the listen-better side of that line for political work, and that conservatism may hold given the reputational cost of a high-profile post-election miss. The 5% response-rate baseline is the structural pressure that eventually forces a decision, and the May results will be an early read on whether AI-run qualitative work can occupy the gap that paid-panel research no longer fills.