An AI-agent investigation starts an evidence clock, not a verdict

The Associated Press reported on 30 September that the U.S. Federal Trade Commission had opened an investigation into OpenAI, Anthropic and other artificial-intelligence companies over possible consumer dangers. AP said an FTC spokesperson confirmed the inquiry but declined to give details. Reuters separately reported that the agency planned formal demands for information and testimony. Neither account supplies the demands, the legal theory or a finding of harm.

That distinction matters. An investigation is a process for obtaining evidence, not a verdict. It should not be used to claim that a named developer broke the law or that every agentic deployment is unsafe. It does change the operating question for buyers and builders: could the organisation reconstruct what an agent was authorised to do, what it actually did, what controls fired and what people decided next?

Freeze the evidence map

Preserve a bounded record for every material agent incident and near miss. At minimum, keep the model and tool versions, system instructions, delegated credentials, permission changes, external destinations, tool-call sequence, safety interventions, human approvals, customer impact assessment and notification decision. Hash exports and record retention clocks so a later reconstruction can distinguish the original record from a summary written after the event.

Map custody as well as content. A vendor may hold model telemetry, an evaluator may hold sandbox logs, a cloud provider may hold network evidence and the deploying organisation may hold business context. Name an owner for each record and document lawful access before a demand or dispute arrives. Do not collect unrelated personal data merely because more logging feels safer.

Test the explanation before it is needed

Run one reconstruction exercise from alert to board decision. Ask a reviewer who did not handle the incident to determine the agent's authority, the boundary it crossed, the evidence supporting impact, the containment step and the reason for any customer notice. Record which questions cannot be answered and who must close each gap.

The strongest counterargument is that exhaustive logging can expose secrets, personal data and security techniques. That is real. Evidence preservation therefore needs field-level minimisation, access control, encryption, deletion rules and a legal hold that applies only when justified. A complete indiscriminate data lake is not the answer.

Add a simple control matrix to the exercise. For each authority path, identify the business purpose, data class, normal approver, emergency approver, revocation mechanism and evidence owner. Test whether revocation reaches cached credentials, queued tasks and delegated sub-agents, not just the visible user account. Compare the written boundary with one actual log sequence. If they differ, preserve both and record the remediation rather than editing the history into apparent compliance.

Notification deserves a separate decision record. Document who assessed consumer harm, which facts were known at the time, what uncertainty remained, which contractual or regulatory duties were considered and when the decision will be revisited. This is not an instruction to notify prematurely. It is a way to show that silence, disclosure and timing were deliberate decisions grounded in evidence rather than gaps in ownership.

The Skills Intelligence Role Dictionary can help assign incident commander, system owner, legal, privacy, security and customer-communication responsibilities. It cannot determine liability. Required legal and domain review should interpret any actual request from the FTC; this draft only turns the public investigation into an evidence-readiness test.

The immediate decision is to start a 72-hour evidence-readiness sprint: inventory agent authority paths, preserve one representative incident record and run a blind reconstruction. Escalate material gaps, but do not label them violations without the underlying legal and factual analysis.