AI Explainability: Showing Why a Recommendation Was Made

The full executive answer

I can show you, on this screen, four things about every recommendation. First, its sources: every insight lists the exact documents it drew on — we call them grounded sources — and the named frameworks it cited, and those citations are checked against our library of 74 real frameworks before anything is saved. If the AI invents a framework name, the system strips it out automatically. Second, its numbers: a validator called numeric provenance checks that every figure in the prose exists somewhere in the data we gave the model. The AI is allowed to phrase facts — it is never allowed to introduce them.

Third, its confidence: every board-grade output carries a confidence band — a stated range, not false precision — and that band is adjusted by our Evidence Map on the server, outside the model. If independent sources contradict a company’s self-assessment, the band is widened downward and the conflict is stamped on the output. The "how much of this is independently corroborated" percentage is computed by our code, never by the AI, so it can never be invented. Fourth, its second opinion: on the highest-stakes modules a second, independent model from a different provider analyses the same data, and material disagreements are shown to you, never hidden.

To be precise about the limits: this is traceability and justification, not a mathematical proof of the model’s internal reasoning — nobody in the industry can open up a large language model neuron by neuron. What we give a board is what NIST calls explainability at the level a decision-maker needs: what data went in, what sources were used, how confident the system is, where it disagrees with itself, and a named human who remains accountable for the decision.