During a high-stakes demo to a research team at a major buy-side firm, KnowYourCompany.ai's system was asked about the pricing outlook for a mid-cap healthcare company. Rather than answering with incomplete data, it responded: "Insufficient context to form a reliable view."
The senior analyst's reaction was the moment that stuck with me. She said: "This is the first AI tool that's been honest with me." Competing platforms had confidently answered the same question, but their answers contradicted her own research.
When she verified those confident answers from other tools, she found them directionally wrong on regulatory pricing impact and cost headwinds. She put it plainly: "I can work with 'I don't know.' I can't work with confidently wrong."
Building in an acknowledgment of uncertainty was contentious internally. Half the product team wanted the system to always generate a response, arguing users expect answers, and that coverage percentage — the share of queries that return a substantive result — is a competitive advantage in demos.
The other half pointed at the stakes. In equity research, a confident wrong answer has a measurable cost: analysts spend real time verifying outputs, and portfolio managers allocate capital based on research summaries that might be flawed.
One analyst put the tradeoff simply: she'd rather get no answer than spend forty-five minutes validating an AI output she isn't sure she can trust — often longer than doing the research by hand.
Three mechanisms drive the confidence assessment:
Source sufficiency scoring checks whether the system actually has access to the data sources a given query type requires before it generates anything.
Temporal validation checks that the available data is current enough to support a conclusion — the common failure mode is retrieving something relevant-looking but stale.
Conflict detection surfaces contradictory signals instead of silently resolving them. When the evidence is mixed, the system says so instead of picking a side.
Looking at early user interactions after launch:
The part that mattered most: when the system did give a confident answer, users trusted it significantly more, precisely because they'd seen it decline when it should. One analyst said it best: "When your system tells me something, I actually believe it. Because I've seen it tell me when it doesn't know."
The incentives in AI product design reward comprehensive, always-available answers. But in domains where the decisions are consequential — healthcare, legal, engineering, policy, finance — trust is built through demonstrated honesty, not through coverage.
Building "I don't know" as a real feature goes against that grain. It's also, in our experience, the thing that made people believe the rest of what the system says.
Originally published on the KnowYourCompany.ai blog.
Thanks for reading. More writing →