The Audit Trail Is Not the Decision Trail.
Logs tell you what happened. A decision trail has to explain why.
An organisation can have excellent technical logging and weak Decision Traceability. The difference shows up when someone asks a question the log was never designed to answer.
Reconstructs activity.
Who logged in. Which model responded. What transaction occurred. Which workflow fired. Which user clicked approve.
Reconstructs judgement.
What evidence existed. Which assumptions mattered. What AI contributed. What the human decided. What was challenged. Why the choice made sense.
Disclosure, mitigation and assurance answer different questions.
Research from the MIT AI Risk Initiative and MIT FutureTech examines how major companies publicly acknowledge AI risks and describe their responses. One reported concern is the limited connection between disclosed risks and the mitigations intended to address them.
Risk disclosure
What exposure does management acknowledge? This establishes visibility, not control effectiveness.
Board challenge: What material risks remain uncertain?
Mitigation evidence
Which measures address each exposure, and what testing shows they work under relevant conditions?
Board challenge: Can the risk-to-control link be demonstrated?
Decision assurance
Who chose the response, on which evidence, with what authority, challenge and residual-risk acceptance?
Board challenge: Could an independent reviewer reconstruct the judgement?
Research: MIT AI Risk Initiative and MIT FutureTech, Unmitigated? How Are the World's Largest Companies Responding to AI Risks? See MIT AI Risk Disclosure Tracker. The decision-assurance interpretation above is The Decision Layer™ analysis, not a conclusion attributed to MIT.
The output can be preserved while the decision context disappears.
The provider, model version, or configuration changed after the original decision.
Replaying the workflow today may produce a different answer, so current evidence cannot safely substitute for historical evidence.
Do we preserve the model and configuration context needed to reconstruct the original decision?
Technical context is not organisational context.
Prompts and outputs are logged, but the risk appetite, approval threshold, or business rule in force at the time is not linked to them.
The organisation can reproduce the interaction without explaining why the output was acceptable.
Can we connect the technical record to the authority and policy that governed the decision?
The action survives. The reasoning does not.
A reviewer overrode or accepted an AI recommendation, but only the final action was retained.
The organisation cannot distinguish meaningful human judgement from ceremonial approval.
What evidence shows what the human actually considered?
The decision crosses systems. The evidence does not.
An external AI service supplies a score, recommendation or agent action, but the underlying evidence remains with the supplier.
Internal logs begin after the most important reasoning has already occurred.
What traceability rights are contractual, and what evidence disappears at the API boundary?
Each system has a log. Nobody has the story.
Recommendation, approval, and execution now happen in different systems owned by different teams.
Local logs remain intact while end-to-end accountability becomes fragmented.
Can one decision identifier connect evidence, authority, action and outcome?
Pick one decision from six months ago.
- Who initiated it?
- What evidence was available?
- Which model, configuration and policy context applied?
- What did the AI contribute?
- What did the human contribute?
- What challenge occurred?
- Who had authority?
- What would have caused the decision to be reopened?
- Which material AI risk and business consequence prompted the decision?
- Which mitigation was selected, and what evidence showed it addressed that risk?
- Who accepted the residual risk, and where is that acceptance recorded?
If the answer requires days of archaeology across email, meeting notes, logs and people's memories, the organisation may have logging. It may not have Decision Traceability.
Select one consequential AI-enabled decision from the last six months. Ask management to produce the risk-to-mitigation mapping, the contemporaneous decision record, and outcome verification. Give an independent reviewer the evidence pack without the original decision-maker in the room. Record missing links, accountable owners, and deadlines for remediation.
A risk disclosure shows what management acknowledges. Mitigation evidence shows what was done. A decision trail shows why the choice was defensible at the time.
Next: Board Case Study, 17 November. The Evidence Did Not Survive.