What Your AI Cannot Prove

Your pilot went well. That was never the question.

In a regulated industry, an AI system is acceptable only when it can prove what it did: which source it used, what stopped it from acting, why it decided as it did eleven months ago, when the person who has to answer for it is sitting across from an examiner and the engineer who built it has moved on.

This is the executive companion to the Knowlytix AI Engineering Series — written for the boards, executives and compliance leaders who decide whether an AI system is built, approved and defended. It doesn’t ask you to read the architecture. It gives you the questions that separate a system you can sign off on from one that merely demonstrates well, and tells you what a credible answer sounds like.

The book carries one running example throughout: a governed complaint-handling assistant at a retail bank. The technical volumes build and test that system; this one picks it up at the decision to approve it. Where a scenario is illustrative rather than reported, it says so, and regulatory positions are stated as of mid-2026.

Written by Al Nugroho, who has spent two decades in second-line roles at large financial institutions — model risk, operational risk and compliance — on the side of the table that asks for the evidence.