What a chatbot summary drops.

This is the call that made us build Narrera the way it is built.

A founder we know ran a thirty-minute customer interview through a popular meeting assistant and got back a tidy list: five pain points, three feature ideas, a sentiment score. It looked like research. It took thirty seconds to read.

Then we listened to the recording with the summary beside it.

What was there that was not said

Two of the five pain points did not appear in the call. They were plausible for the industry, and they were the kind of thing the founder had typed into the prompt as context. The model had helpfully included them.

What was said that was not there

At minute fourteen the interviewee said: "Yesterday we had four no-shows before ten, and two of them had already been rung." Dated, specific, a number, a failed workaround. It is the strongest sentence in the call. It was not in the summary, because the summary's job was to be tidy, and a story is not tidy.

What we decided

• Findings are quotes. The text on the screen is the text that was said, linked to the second it was said.
• Every finding has a score, and the score is about the kind of claim, not about how interesting it sounds.
• Nothing is dropped silently. Weak claims are shown, dialled down, with the question that produced them.
• Results count people, not quotes, and never open before you have heard from enough of them.

A computer should not decide what counts and keep no record of what it dropped. That is the sentence in the box on our home page, and it is the reason the box exists.