We believe AI in enterprise workflows must have unambiguous boundaries. We use AI exclusively as a structured verification lens, never as an unconstrained agent with write access to your production systems.
Customer data is never used to train, retrain, fine-tune, or improve any public or private artificial intelligence models. Our commercial API agreements with OpenAI strictly prohibit model training on customer payloads. Payloads sent to OpenAI API endpoints are retained temporarily (up to 30 days) solely for abuse monitoring before permanent deletion.
When connecting operational signals to task completion, XRAY Bridge uses a two-stage verification rail. Inbound facts from external tools are first evaluated against deterministic rules. When semantic evaluation is required, our Stage-2 AI Judge (powered by gpt-4o-mini in strict JSON schema mode) evaluates the factual evidence against the task description.
Autonomous completion (Autonomous Close) requires both a verified verdict from the Stage-2 Judge AND a confidence score of 90% or higher. If the AI model times out, returns malformed JSON, or experiences an upstream outage, the system fails closed: tasks are automatically downgraded to human review staging (Comment & Stage for Human Review) with aiJudgeVerdict: 'skipped'. An AI outage will never cause unauthorized task closures.
Before any agent payload, brief summary, or external message is delivered, our outbound policy engine scans the content for sensitive tokens. API keys (e.g., sk-..., ghp_..., xoxb-...) are redacted, payment cards are masked to the last 4 digits (e.g., **** **** **** 1234), and private communication channels (#hr-confidential, #legal) are excluded at the scope boundary.
By default, every new workspace operates in Review Staging (Comment & Stage). In this mode, verified facts and task recommendations are staged for human review and never auto-closed in external trackers.
Workspaces that opt into Autonomous Close are governed by strict safety limits: a daily auto-close cap (default 50 tasks/day), mandatory confidence floor (≥ 90%), and full 1-Click Rollback hygiene snapshots for every automated run.