Approval Is Not Evidence.
Three evidence decisions need attention before assurance becomes archaeology.
Governance often preserves the approval and loses the decision. That distinction matters most when an outcome is challenged months later and the organisation has to explain not only who signed, but why signing was reasonable.
Define minimum decision evidence.
Specify what must exist before a consequential AI-enabled decision is accepted: material facts, assumptions, alternatives, challenge, authority and rationale.
Separate approval from rationale.
A workflow timestamp is evidence of approval. It is not evidence of judgement. Preserve the reasoning that made the approval defensible.
Decide what must survive.
Evidence scattered across email, model logs, meeting notes and vendor portals is evidence in theory. In practice, it is archaeology.
A defensible decision is not one with perfect information. It is one whose evidence, assumptions and reasoning can still be understood after the room has emptied.
Most organisations preserve activity better than judgement.
AI systems generate more data than most governance teams could ever read. That can create a false sense of evidential abundance. Logs multiply. Dashboards thicken. Approvals acquire timestamps. Yet the material context of the decision can remain surprisingly fragile.
Decision Evidence is narrower and more demanding. It asks what a reasonable outsider would need in order to understand why the decision made sense given what was known at the time.
| What often survives | What often disappears |
|---|---|
| Approval timestamp | Why the approver believed the evidence was sufficient |
| Model output | Which assumptions and thresholds shaped reliance on it |
| Meeting invitation | What challenge materially changed the discussion |
| Final recommendation | Which alternatives were rejected and why |
| System log | What the human understood, questioned or accepted |
The approval survives. The rationale does not.
The organisation can prove that the decision passed governance, but not why the evidence was judged sufficient.
The output survives. The context does not.
The recommendation remains available after the model version, prompt, policy threshold or source data have changed.
The vendor owns the evidence.
Critical reasoning depends on logs or documentation that sit outside the organisation and may not be retained on the same timetable.
The challenge happened, but left no trace.
A strong question changes the decision. Minutes record the outcome, not the objection that improved it.
The evidence is everywhere.
Nothing is missing individually. The problem is that reconstruction requires six systems, three inboxes and somebody who still remembers the meeting.
Before accepting a consequential decision, ask whether six things would survive it.
| Evidence element | What should be recoverable later |
|---|---|
| Material facts | The information actually available when the decision was made |
| Assumptions | The conditions believed to be true, including uncertainty |
| Alternatives | Material options considered and why they were not selected |
| Challenge | The objections, tests or contrary views that mattered |
| Authority | Who had the right to decide, approve, stop or override |
| Reasoning | Why the final decision was reasonable in its original context |
The question is not whether every decision needs a dossier. It is whether the evidence requirement scales with consequence. A low-impact workflow may need a compact receipt. A material decision may need a reconstructable decision trail.
- For our most consequential AI-enabled decisions, what evidence must exist before approval?
- Can we distinguish evidence of approval from evidence supporting the decision?
- Which parts of the decision trail depend on a supplier retaining records for us?
- Could we explain why a decision was reasonable if the outcome was poor?
- What evidence is deliberately not retained, and why?
- Who decides when the evidence base has changed enough to reopen the decision?
| Action | Test of completion | Owner |
|---|---|---|
| Define a minimum evidence standard | Material AI-enabled decisions have a documented evidence requirement proportionate to consequence | CRO / AI Governance |
| Test one past decision | A reasonable outsider can reconstruct the evidence and reasoning without relying on memory | CAE / Assurance |
| Review supplier evidence dependency | Contracts and retention arrangements support the organisation's reconstruction needs | CIO / Procurement |
| Create reopening triggers | Material changes in model, data, policy, supplier or operating context force reconsideration | Executive owner |
An approval tells us somebody said yes. Decision Evidence tells us why yes was reasonable at that moment.
Next: Decision Signals, 10 November. The Audit Trail Is Not the Decision Trail.