Imagine you want to know how an ad really felt while people watched it, not what they remember a week later. Asking in the moment beats asking from memory, and it always did. What kept it off most projects was two costs. You now have an AI assistant. Hand it each cost in turn and watch which one it actually solves.
Illustrative panel. Outcomes reflect recorded research as of Aug 2026, not a live model callEvery photo, voice note and video a participant sends has to be watched, read and coded. For decades a trained person did that by hand, and the work never scaled.
A fixed schedule buzzes people at set times until they tire and stop answering. The fix would be a prompt a model writes on the spot from your live context, so it never annoys.
One cost fell. One did not. AI reads the captures at a scale a human team never could, and that half is real and buyable today. The prompt that writes itself is still a promise with no product behind it and no trial to back it.
So the honest state of the method is half-solved. Buy the reading. Stay sceptical of any tool that says it writes the ping for you, and ask to see the evidence it beats a fixed prompt.
A self-contained embed cannot call a model, so both outcomes are fixed from the record, not measured live. Cost 1 clears on peer-reviewed work reading 34,000 captures at scale (Guo et al., 2025) and on shipping vendor tools. Cost 2 stays stuck because its flagship system is an unvalidated preprint (Carmon et al., 2026) and no product both schedules a ping and authors it from context. Neither is a vendor accuracy claim.