Retail AI users want data disclosure and an off switch before better recommendations

ThoughtSpot’s new YouGov-backed survey offers a useful warning for teams putting analytical agents in front of customers: better answers alone may not repair trust if users cannot see what data was used or turn the system off.

The survey covered 4,833 adults—2,617 in the US and 2,216 in the UK—and was fielded online from June 16 to June 23, 2026. ThoughtSpot says the results were weighted to represent adults in both countries. The headline result is stark: only 13% said they trusted AI-powered product or brand recommendations, while 47% distrusted them and 34% were neutral. Just 1.4% named AI recommendations as their most trusted source for purchase decisions.

Control outranked accuracy

The most actionable result is the ordering of proposed trust builders. Transparency about which data is used ranked first at 34.0%, followed by an option to opt out of AI personalization at 31.4% and a clear explanation of why an item was recommended at 30.6%. Human support alongside AI reached 29.8%. Consistently accurate recommendations ranked lower, at 21.2%.

That does not mean accuracy is optional. The same report says 66.2% experienced AI misunderstanding their needs at least rarely, and 59% were unlikely to continue shopping with a retailer if AI recommended products outside their price range. A significantly late delivery would make 53% unlikely to return. The distinction is that accuracy is only one part of the trust contract.

The privacy baseline is equally difficult: 74% agreed retailers already collect too much personal data. Cross-site or cross-app tracking was called intrusive by 57.6%, while 48.8% objected to real-time location use. Even knowingly provided profile or preference data was considered intrusive by 27.8%.

The limit—and the deployment lesson

This is a consumer retail survey, not an evaluation of enterprise text-to-SQL systems. It does not measure query correctness, semantic-layer coverage or whether showing generated SQL improves error detection. Its findings should not be presented as proof that disclosure controls increase NL2SQL accuracy.

It does, however, identify three interface requirements worth testing whenever conversational analytics produces customer-facing recommendations or actions: show which governed data informed the answer, provide a clear opt-out or human route, and explain the recommendation in plain language. Teams should evaluate those controls separately from model accuracy rather than assuming a better benchmark score will carry the entire trust burden.

The neutral 34% may be the practical opportunity. ThoughtSpot’s own framing is that trust is won through transparency and control as product features, not policy-page language. For data-agent builders, that turns provenance and refusal paths from governance documentation into visible parts of the user experience.