Academy

AI that survives an audit

When an intelligent system makes a decision, sooner or later someone asks: 'why?' A system that can't answer fails an audit. We believe auditability has to be in the design from day one, not a later add-on.

The Felesh team 4 min

Any system that makes decisions sooner or later faces one question: “why was this decision made?” That question may come from an auditor, a regulator, or a colleague trying to understand what happened. A system that cannot answer — whose decisions are a black box — fails at that moment. We believe auditability shouldn’t be a later add-on; it should sit at the heart of the design from day one.

Recorded reasoning

The cornerstone of an auditable system is a simple principle: every action should leave a trail of what it was based on. We call this recorded reasoning. When a decision is made — whether by a human or an agent — it’s linked to the information that justified it. This “based on what” link is exactly what later answers the question “why.”

A decision linked to its basisThree pieces of information feed one decision. A customer message from 14 Oct says the crack was seen on day 9, photo IMG-2218 shows a crack across the glass, and warranty clause 4.2 covers 30 days from purchase. Each is linked to the decision that claim #6602 moves to Review, which is stored with the note: based on 3 items. When someone later asks why, the links answer.what it was based onthe decisioncustomer message, 14 Octcrack seen on day 9photo IMG-2218crack across the glasswarranty clause 4.230 days from purchaseclaim #6602moves to Reviewbased on 3 itemsstored with itask why later, and the links answer
Every decision is stored with the information that justified it, so the link already exists when someone asks why.

A glass box, not a black box

The key difference is that the system holds not just the outcomes, but the reasoning too. A black box only says “this decision was made”; a glass box says “this decision was made, on the basis of this information, by this path.” When an auditor later asks why a matter advanced to the next stage, the answer is in the system itself — not in someone’s guess, but in the recorded trail.

Black box and glass box, same questionAn auditor asks why claim #6602 advanced to Review. A black box answers only that claim #6602 moved to Review, and the rest is someone’s guess. A glass box answers that claim #6602 moved to Review on the basis of the 14 Oct message, photo IMG-2218 and warranty clause 4.2, by the path intake, photo check, rule check. The answer sits in the system itself.auditor: why did claim #6602 advance to Review?black box#6602 to Reviewthat is all it saysbasis: not keptanswer: someone guessesglass box#6602 to Reviewon the basis ofmessage 14 Oct, IMG-2218,warranty clause 4.2by pathintake, photo, rule checkthe answer sits in the system, not in a guess
A black box keeps only the outcome; a glass box keeps the outcome, its basis and its path, so the answer is in the system.

Transparent authority

Auditability isn’t only about individual decisions; it’s also about who is allowed to do what. In our design, that authority is transparent too: you can see on what basis an agent earned permission for a task, what corrections a human reviewer applied to its work, and the history of its accepted and rejected proposals. No decision is ownerless, and no authority is hidden.

Who is allowed to do what, on the recordThree kinds of entry, each with an example. Permission: the agent may approve claims under 500K, granted 11 Sep by Arash on his review. Corrections: on claim #6571 the payout went from 900K to 640K, fixed by Arash on 28 Sep with the reason noted. Proposals: reject #6590 was accepted, approve #6602 was rejected by Arash. No decision is ownerless and no authority is hidden.entryexample from the recordpermissionmay approve claims under 500Kgranted 11 Sep by Arash, on his reviewcorrectionsclaim #6571: payout 900K to 640KArash, 28 Sep, reason notedproposals#6590 reject: accepted#6602 approve: rejected by Arashno decision is ownerless, no authority is hidden
Authority leaves a trail too: each grant, correction and proposal is one entry with a name and a date on it.

Why this matters from day one

The easy temptation is to build the system first and add auditability later. But that almost always fails, because auditability is an architectural property, not a closing report. If the system doesn’t keep the reasoning trail from the start, you can’t build it from nothing later. That’s why we treat it as a foundational principle, not a secondary requirement.

Kept from day one or added laterTwo timelines from January to June, each with a decision every month. On the first, reasoning was kept from January, so when the audit asks why in June, every decision answers. On the second, reasoning was added only in June: the five earlier decisions show a question mark, because a trail that was never kept cannot be rebuilt.reasoning kept from day onereasoning added in JuneJanFebMarAprMayJunaudit: why?audit: why?every decision answers?????a trail never kept cannot be rebuilt
Reasoning kept from the first decision answers the audit; added in June, it has nothing to rebuild the earlier months from.

Putting it together

We believe trust in an intelligent system comes from its transparency, and transparency comes from design. A system that links each decision to its reason, makes its authority visible, and holds its reasoning — not just its outcome — is one that can stand up to the hardest questions. And in a world where AI makes ever more important decisions, it’s this capacity to answer that draws the line between a trustworthy tool and a risk.