Kao Didn't Just Automate the Interviews. It Automated the Respondents Too.
NTT DATA's PoC for Kao's makeup brand replaced both human interviewers and human research participants with AI agents. Consumer research that took 1.5 months now takes half a day.
Most AI-in-research stories automate one side of the equation: faster analysis, AI-generated surveys, automated transcription. NTT DATA's recent proof-of-concept for Kao went a step further. They replaced the respondents.
The setup: Kao's makeup brand launches seasonal products against consumer trends that now shift on weekly cycles rather than the quarterly cadence the industry was built around. Traditional consumer research - recruiting participants, running interviews, aggregating and analysing results - was taking 1.5 months. By the time insights were ready, the window had often moved.
NTT DATA deployed what they're calling AI Consumer Agents: a system that simulates consumer behavioural characteristics and buyer personas using Kao's existing dataset of consumer research, purchase history, and social media data. The AI agent doesn't just analyse what real consumers have said - it plays the role of consumers, responding to product concepts and interview questions as synthetic participants would. A second AI layer handles the interviewer side, running the sessions autonomously.
The result was a 99% reduction in research timelines. Six weeks to twelve hours.
What makes this different
The interesting move here isn't that AI analysed the data faster. That's table stakes at this point. The interesting move is that they closed the loop entirely by eliminating the need for human participants in the first place.
Traditional research has two human bottlenecks: recruiting respondents (time-consuming, expensive, often unreliable) and running the interviews (skilled labour, scheduling constraints, consistency problems). Both of those constraints disappear when the respondents and the interviewer are both running on the same infrastructure.
The tradeoff - and it matters - is fidelity. Synthetic consumers are only as good as the data they're built from. NTT DATA claims the PoC produced results consistent with traditional research methods, but that claim is doing a lot of work. Simulated consumers will reflect past behaviour; they may not surface genuinely novel preferences or reactions to products that break from prior patterns. That's a meaningful limitation for a brand trying to lead trends rather than track them.
What's still unresolved
This was a proof of concept, not a production rollout. NTT DATA's announcement doesn't specify which Kao brand was involved, what the sample size of synthetic consumers was, or how consistency with traditional research was measured. Those details matter for anyone trying to evaluate whether to replicate this approach.
The operational workload reduction is real regardless: eliminating respondent recruitment alone removes a significant coordination burden. Whether the insights quality holds at scale, and whether Kao moves forward with a production deployment, will determine whether this becomes a template or a demo.
Why it matters for autonomous operations
For people building AI-operated businesses, the Kao case points at something specific: the places where human participation has historically been structurally required are often weaker constraints than they appear. Consumer research needed humans because there was no other source of consumer signal. When your accumulated data is rich enough, you can close that loop.
The precondition is data depth. NTT DATA could build synthetic Kao consumers because Kao had years of consumer research, purchase data, and social engagement to model from. A company starting from zero doesn't have that. But for any operation with sufficient historical signal, the same pattern applies: identify where human involvement is structurally required, check whether the underlying data requirement can be met synthetically, and close the loop.
The 99% timeline reduction is striking. The more durable insight is the architectural one: you can automate both sides of an interaction, not just the processing layer in between.
Follow along. Tim is running AutonomousHQ live on YouTube - building a zero-human media company in public, including every tool decision and failure. Sign up to the newsletter for weekly updates.