The agent that learned to recommend nothing
We tuned a skincare concierge to recommend nothing in roughly a third of conversations. That is when the trust scores moved.
Most teams ship an AI agent and judge it on conversion. Did the user buy. Did the ticket close. Did the funnel move. Reasonable on paper. Wrong in practice for anything that involves expertise.
We built a skincare concierge for a founder with thirty years of formulation work. The first batch of conversations felt like sales calls because every exchange ended with a product. The founder caught it in the review queue and was firm: real consultations sometimes end with "you don't need another product right now, give the one you have eight weeks."
We tuned the agent to recommend nothing in roughly a third of conversations. That is when the trust scores moved. Customers started telling the agent things they would never tell a salesperson. Bad reactions. Breakouts before weddings. Twenty years of trying products that didn't work.
The agent's job stopped being "sell well." It became "listen first, then earn the recommendation."
Two takeaways for anyone building agents in expert domains:
The metric is wrong if it only points one way. If the agent can only succeed by selling, it cannot tell the truth. Add a second metric that rewards the right kind of pause.
The expert in the loop is not optional. Without the founder reviewing the first hundred conversations, we would have shipped a faster sales rep, not a concierge. The agent learned her judgment by being corrected on the cases that mattered.
The honest version: AI will not replace the founder's taste. It will ensure every customer gets a thoughtful answer at 10pm on a Tuesday, and a warm hand-off to the human when it matters.