What is Decision Provenance?

Ask a leadership team what it decided last year and someone will find the answer. Ask why, which evidence carried the day, who overruled whom and whether the call turned out to be right, and the trail usually ends at an undated slide deck and the memory of whoever was in the room. The decision survives. The reasoning does not.

Decision Provenance is the name for what is missing. Decision Provenance is the unbroken, traceable chain from evidence to outcome: which facts, at which version and in which quality state, fed which recommendation, who decided, what was overridden, and what actually happened. It does for judgement what a chain of custody does for physical evidence. It is not extra paperwork around the artefact; it is the thing that makes the artefact trustworthy at all.

The chain, link by link

Provenance is only as strong as its weakest link, and there are six:

• The facts, versioned. Which versions of which facts were on the table, not what the numbers were later corrected to.
• The quality, labelled. The evidence state each fact carried: measured, modelled, inferred, stated or unmeasured.
• The evaluation, frozen. The Scenario Run that priced the options is immutable and timestamped, so the moment of reasoning cannot be quietly revised.
• The recommendation, explained. Which option ranked first, and which constraint bound it.
• The choice, attributed. Who decided, and where the human overruled the recommendation: who, when and against what evidence.
• The outcome, scored. What happened, compared with what was recommended and what was decided.

Remove any link and the rest becomes decoration. A recommendation without fact versions is an opinion with formatting. A recorded choice without a scored outcome is a diary. Together they form a chain a stranger can walk, in either direction, years later.

Auditable confidence versus asserted confidence

Every organisation expresses confidence in its decisions; the question is what stands behind the expression. Asserted confidence is a number typed into a slide, and its credibility rests entirely on the seniority of the person asserting it. Auditable confidence decomposes on demand: this dimension is weak because its facts are merely stated; this date binds because the contract clause clears last; this recommendation shifted between March and June because one fact was re-versioned. When confidence is composed from labelled evidence rather than declared, interrogation replaces persuasion, and a junior analyst with a better fact can legitimately move a number a senior director asserted.

One concrete example

Illustrative, with no customer implied. Eighteen months from now, a new chief operating officer inherits a portfolio and asks a simple question about the largest account: why did we agree to those terms? In most firms that question launches an archaeology project: old email threads, a departed sponsor, three conflicting versions of a deck. With provenance it is a lookup. The renewal was evaluated on a known date against known fact versions, two of which were only stated by the client. The recommendation was to renegotiate scope. The commercial director overrode it to protect the relationship, and said so at the time. The outcome, scored a year later, came in below the modelled case. Nothing needs reconstructing and nobody needs blaming; the next negotiation simply starts from the record instead of from rank.

Provenance and the expectations on regulated AI

Regulation is arriving at the same demand from another direction. The EU AI Act sets expectations of transparency, traceability and human oversight for high-risk AI systems, and the practical test behind those words is provenance shaped: show which inputs produced which output, and show where a human decided. An organisation that cannot answer will experience explainability as an expensive retrofit; one with provenance experiences it as a property it already has. None of this is legal advice, and whether specific obligations apply to a specific system is a question for counsel. But the engineering posture is a choice available today, and it is the one ONX makes for itself: candidate matching in Hiring is treated as high-risk under the strict reading, with human review gates, and every recommendation on the platform is kept explainable and challengeable, because a conclusion you cannot interrogate is not one you should be asked to trust.

Provenance is the difference between “trust us” and “check for yourself”. The wider discipline of treating decisions as governed, traceable objects is set out in Enterprise Decision Intelligence.