A confidence label is a required tag, such as verified, inferred, or unverifiable, attached to every claim an AI research agent returns, so a person or a downstream rule can tell a retrieved fact from a plausible-sounding guess before it reaches a prospect.
Every claim a research agent returns should carry one of a small set of labels, and the label should never be optional. One version of the framework uses three: VERIFIED for a claim backed by a specific quote from an admissible source, INFERRED for something reasoned from adjacent facts that still needs a person to check the reasoning, and UNVERIFIABLE for a claim with no admissible source behind it at all. A related version of the same idea labels claims verified, inferred, or open, with 'open' marking anything that only shows up on a source the subject fully controls, like their own About page, until something independent confirms it.
A confidence field sitting unused in a JSON schema doesn't do much work on its own. Nothing forces an agent to fill it in accurately, and nothing downstream enforces what happens when a label comes back low, unless the routing rule was decided before the agent ever ran. UNVERIFIABLE claims should never reach personalization copy or a live sequence. INFERRED claims go to a review queue, not straight into copy. Only VERIFIED claims move into a properly configured send unreviewed, and even those are worth spot-checking on the highest-value accounts.
Require the label as a mandatory field in the prompt, not an optional one, and decide the routing rule for each label before the agent runs a single record. Deciding what to do with a low-confidence claim after it's already sitting in a draft email is deciding it too late.
Teams treat a confidence score as if it does the same job as a confidence label. A bare number is something a person has to take on faith. A label paired with the actual quote it's based on is inspectable: anyone can check the source in seconds instead of re-researching the whole claim from scratch.
Tell us how your motion runs today. We'll show you what we'd engineer.
Contact us