How Confidence Should Work in AI-Assisted Decisions
A confidence percentage attached to an AI-generated recommendation is only useful if it means something specific. Too often it's presented as a bare number with no explanation — which makes it decoration, not information.
The problem with an unexplained confidence score
"73% confidence" tells you almost nothing on its own. Confident about what — that the recommendation is directionally correct, that all the relevant facts were available, that the numbers hold up under scrutiny? Without a stated basis, a confidence score can create false reassurance: a specific-sounding number that hasn't actually been earned by the evidence behind it.
What a meaningful confidence score requires
- A stated basis. Confidence should reflect how much of the decision rests on verified facts versus assumptions — not a vague sense of certainty.
- Visible missing information. If key facts are unavailable, confidence should be capped accordingly and the report should say what's missing, not just imply it through a lower number.
- Consistency across reports. The same evidence quality should produce roughly the same confidence level regardless of which decision it's attached to — otherwise the number isn't calibrated to anything.
- A clear read on what raises it. A useful confidence score comes with an implicit or explicit answer to "what would make this more certain" — which is itself actionable information.
What low confidence should actually communicate
Low confidence isn't a failure of the analysis — it's often the most valuable output, when it correctly signals that a decision shouldn't be made yet, or should be made with explicit awareness of what's unverified. A report that returns "35% confidence, gather more information" has done its job if that's genuinely where the evidence stands, even though it feels less satisfying than a confident yes or no.
This is why Ascendra ties confidence directly to known facts versus stated assumptions in every report, rather than presenting it as an unexplained score — see it applied in a real hiring decision where confidence stayed moderate specifically because key growth assumptions were unverified.