Key facts
| Event classes | Authentication, key lifecycle, admin changes, inference metadata |
| Accountability | Approval and decision records stay in your case systems and are correlated with AI events |
| Export | SIEM platforms including Splunk, Sentinel, QRadar and Datadog |
| Sovereign residency | Region-locked planes: EU (Frankfurt), GCC (UAE), APAC (Singapore), US (Virginia) |
| Air-gapped logging | Equivalent events inside the perimeter; export mechanism for external review |
| Retention | Configurable prompt retention; records policy governs metadata retention |
| Identity | RBAC and SSO/SCIM distinguish staff and service identities |
| Status | Audit export is live; assistants and responses endpoints are coming soon |
TL;DR
- Log enough to explain a decision to an oversight body.
- Keep decision records in case systems and correlate them with AI events.
- Retain logs in the jurisdiction and classification level of the data.
- Design disconnected logging before deployment; hindsight is not an option.
- Publish what the system records and how citizens can challenge an outcome.
How it works, step by step
- Translate your accountability obligations into audit questions that a reviewer will actually ask.
- Enable audit export from the workspace and route events to your monitoring stack with a stable correlation ID.
- Record decisions and reviews in your case system with the same identifier, so AI activity and official action join cleanly.
- Choose prompt retention per service: metadata by default, content only where policy and notice permit.
- Pin log storage to the jurisdiction and classification level required for each service.
- For restricted networks, deploy an internal collector and define the export path for oversight and records requests.
- Run a reconstruction drill with your oversight or internal audit function and close the gaps it exposes.
Try it yourself
Open the EU AI Act compliance checker →
From accountability to event schema
Public accountability questions are practical: which model produced this draft, what information did it use, who reviewed it, and could the citizen have reached a person? Turn each question into an event. Model and prompt version, retrieved document references, reviewer identity and timestamp, and the escalation path taken.
Plugsky supplies platform-side events — authentication, key lifecycle, administrative changes and inference metadata — through audit export. Decision records belong in your case management system. Join them by correlation identifier so an oversight review follows a case rather than assembling fragments from separate systems.
Sovereignty and classification
Logs inherit the sensitivity of the data they describe, so treat them as records: store them in the region-locked plane matching the service, or inside the perimeter for restricted networks. In air-gapped estates, run an internal collector and define how events leave for review — through your controlled channel, under records policy.
- Classification: match log storage to the data class, not to convenience.
- Minimisation: prefer metadata; capture content only with a stated purpose and notice.
- Separation: those who administer the platform should not be the only custodians of its evidence.
Transparency that earns trust
Auditability is stronger when it is visible. Publish a plain-language notice describing what the assistant records, why, how long records are kept and how a citizen can request human review or challenge an outcome. That notice should match what the logs actually contain — which is another reason to keep capture deliberate rather than maximal.
Finally, rehearse oversight: a records request, a disputed decision and a suspected key compromise. Each drill exposes whether the pipeline joins AI events to official decisions and whether exports survive classification handling. Fix what the drills find before an external request forces the issue.
Honest comparison
| Oversight requirement | Plugsky | Typical AI platform | Building in-house |
|---|---|---|---|
| Platform audit | Auth, key and admin events exportable | Vendor logs | Your build |
| Sovereign storage | Region-locked planes plus disconnected options | Few region choices | Wherever you host |
| Restricted networks | Equivalent events inside the perimeter | Rarely supported | Your design |
| Correlation with case records | Request IDs join your case systems | Not addressed | Your integration |
| Retention control | Metadata long, content minimised | Often fixed | Fully yours |
| Time to evidence | Days to enable, drills to prove | Weeks | Quarters |
Frequently asked questions
What must be logged for accountability?
Who or what acted, which model answered, what information was retrieved, who reviewed the outcome, and any administrative change affecting the service. Those events let an oversight body reconstruct a decision.
Can logs stay in our jurisdiction?
Yes. Store events in the region-locked plane that matches the service, or inside the perimeter for restricted networks. Keep keys and logs in the same jurisdiction as the data.
How does logging work on classified networks?
An internal collector ingests the same event classes, with a controlled export path for review and records requests. This should be designed before deployment, not retrofitted.
Do we have to record prompt content?
Only where there is a stated purpose and appropriate notice. Metadata plus retrieval references usually answers audit questions with far less exposure.
How do we handle records requests?
Define retention in line with records policy, keep decision records in the case system, and rehearse an export exercise so requests are routine rather than a project.
What about the EU AI Act?
The Act emphasises logging and traceability for higher-risk systems. The event classes described here — model, data, review and control changes — map directly to that evidence need.
How do we reassure citizens?
Publish what is recorded and why, provide human escalation, and ensure the notice matches actual logging behaviour. Transparency that overstates or understates capture erodes trust either way.