# Administration — AI Operations

![Administration — AI Operations](/docs-content/4e7faf8c14-05-ai-operations.png.webp)

**Navigation:** Administration → AI Operations  
**Access:** AI operations, model-governance, or deployment permission  
**Primary audience:** AI platform operators, model owners, evaluation owners, and authorized release approvers

## Purpose

AI Operations controls the quality and release state of the AI processing capabilities used by the application. It combines the model registry, analysis pipeline, evaluation gates, staged rollout, rollback, recent runs, and change audit. It is not intended for ordinary business users or general infrastructure incident management.

## Page controls

Period selects the operational window. Auto Refresh updates volatile metrics while preserving the user's selected context. Environment and capability scope are available from the relevant detail views. The top status reports overall AI-service availability.

## Model registry

The registry lists each deployed capability, version, release identifier, environment, rollout state, owner, and update time. States include Shadow, Canary, Production, Paused, Failed, and Retired. Selecting a row opens compatibility, evaluation, deployment history, supported languages or tasks, and affected product capabilities.

## Pipeline health

The pipeline view follows major processing stages such as speech recognition, speaker diarization, language analysis, and evaluation. Each stage reports latency, error rate, queue, throughput, version, and state. The flow communicates where a problem occurs; it does not imply that every source uses every stage.

## Capacity and queues

Queues show pending workload by processing class. GPU Utilization shows average and per-device utilization, memory, temperature, and history. These are capacity diagnostics; they should not be used as model-quality measures.

## Evaluation gates

The gate table compares current value with the approved threshold for transcript quality, diarization, precision, recall, hallucination, sensitive-data leakage, toxicity, or other qualified measures. Pass indicates the configured dataset and gate succeeded; it is not a universal guarantee for every real input.

## Canary rollout and rollback

The rollout panel shows traffic percentage, step, start time, next step, and comparison with the active baseline. **Pause rollout** stops further progression. **Review rollback** opens the exact release scope, evidence, approvals, and consequences before returning traffic to the prior qualified version.

## Recent runs and audit

Recent Model Runs lists run ID, pipeline, stage, status, latency, and usage. Audit Log records deployment and evaluation changes, actor, resource, details, and approval source.

## Actions

- Open model, pipeline, queue, evaluation, run, or audit details.
- Pause or resume an authorized canary.
- Approve the next rollout step when gates and review allow it.
- Review and execute rollback under the displayed change control.
- Quarantine a release or route a failed run for investigation.
- Compare current and baseline quality.

## Recommended procedure

Start with failed quality gates and active rollout anomalies, then inspect pipeline stage, affected inputs, and baseline comparison. Pause before changing multiple variables. Roll back the complete qualified release when necessary and preserve run and approval history.

## Important boundaries

Infrastructure health, storage, service incidents, and generic job recovery belong to [Platform Operations](/documentation/packages/bnd-aiops/screens/administration-platform-operations). Source authorization and mapping belong to [Sources & Integrations](/documentation/packages/bnd-evidence/screens/administration-sources-and-integrations). Package installation and business configuration are managed through the platform’s package and workspace administration controls.

## Related screens

[Platform Operations](/documentation/packages/bnd-aiops/screens/administration-platform-operations), [Sources & Integrations](/documentation/packages/bnd-evidence/screens/administration-sources-and-integrations), [Governance](/documentation/packages/bnd-org/screens/administration-governance), and [Audit & Security](/documentation/packages/bnd-aiops/screens/administration-audit-and-security).