Requested action
The exact consequential action the agent proposed or attempted, including target and material parameters.
Agentic AI Execution Control
An AI agent executed without authority. What must you prove now?
When an agent reaches a consequential action without valid authority, a chat transcript or policy statement is not enough. The record needs to establish what the agent attempted, what authority existed, where enforcement should have occurred, whether executable commitment formed, how the action reached consequence, and whether the sequence can be replayed.
The incident question is not merely what the model said.
A consequential agent can move through tools, APIs, payment rails, infrastructure, workflows, or delegated services. The critical governance question is whether valid authority existed at the point where the action became executable - and what control actually prevented or permitted commitment.
Evidence an investigation should preserve
Authority state
The human, institutional, credential, policy, or delegated authority that existed - or did not exist - at the relevant moment.
Execution route
The tools, services, APIs, or alternate paths through which the action moved toward consequence.
Commit formation
Whether a valid executable commitment formed, when it formed, and what evidence supported that formation.
Control response
Whether the system allowed, held, denied, escalated, bypassed, or failed to evaluate the action.
Outcome and replay
What actually happened in reality and whether the full chain can be independently reconstructed.
Executable proof examples
Authority is only part of the execution question
The registry contains executable cases that isolate two closely related failure boundaries. TA14-EA-000013 shows a previously valid approval losing present standing before execution. TA14-EA-000020 shows a runtime action materially diverging from the action fixed at commit. Together they show why an investigation must distinguish authority, commitment, and execution correspondence instead of treating an authorized process as proof that every resulting action was authorized.
TA14-EA-000013 - Authority expires before execution
Present-authority revalidation produces HOLD and prevents the pending consequence from proceeding on expired standing.
Inspect the artifact ->
TA14-EA-000020 - Executed action differs from commit
Execution correspondence fails when the runtime invocation materially differs from the committed action.
Inspect the artifact ->
Logging the incident is not the same as governing execution.
Observability can show that an event occurred. Governance has to answer whether the action was supportable, whether authority was admissible, whether commitment should have formed, and whether the execution path was actually controlled. Those are different claims and should be examined separately.
TA-14 can examine one bounded execution claim without requiring the system under review to become TA-14. Native architecture and evidence remain native; unsupported findings and bypass conditions remain valid examination results.
Frequently asked questions
What evidence matters after an AI agent executes without authority?
Preserve the requested action, authority state, evidence basis, commitment state, execution route, control response, actual outcome, and enough lineage to reconstruct the sequence.
Is an audit log enough to prove an AI action was authorized?
No. A log may show that an event occurred. Authorization requires evidence that the relevant authority existed and remained applicable at the consequence boundary.
What if the agent was authorized but executed a different action?
That is a separate correspondence problem. The runtime action must be compared with the action actually fixed at commitment; process-level authorization alone does not establish that a divergent execution was authorized.
Related execution-evidence problems
Did authority change first?
Examine whether a previously valid authorization lost standing before execution.
Did the evidence become stale?
Determine whether changed evidence should have triggered revalidation before consequence.
Compare against a blocking claim
See what evidence would be needed to prove an action actually stopped instead.
Commercial examination
Need to establish what actually failed?
Use the $249 Execution Evidence Snapshot for one narrow evidence question. Use the $750+ Execution Claim Review when the incident requires authority analysis, alternate-route examination, changed conditions, commitment analysis, or replay. Payment buys the examination - not a favorable conclusion.