AI governance
AI that cites its sources, and shows where a person decides.
Each AI feature answers for a stated jurisdiction and carries a confidence tier. Drafts, Site Scan suggestions and determinative outputs wait for a person. Two parts of the product act before a person reviews them, and this page names both.
The frame
How the AI features work
The frame is part of the product, not a policy on top of it. Where a feature works differently, this page says so.
The jurisdiction it answers for
Features tied to a work site (Site Scan, photo analysis, voice capture, inspection intake and incident forms) use the jurisdiction of the physical work: the incident case first, then the establishment, then the workspace's headquarters, then federal requirements.
Chat and document analysis use the workspace's headquarters jurisdiction.
State-specific depth varies. Where a state has no authored guidance yet, answers cite federal requirements, and Site Scan says so on its findings and in its report.
Citations checked against regulatory text
Chat answers cite the regulatory text they were retrieved from, and their citations are checked against it.
Site Scan and photo findings check proposed citations before they are saved; a citation that cannot be verified is flagged where it appears, not hidden.
A confidence tier
Each AI feature class is assigned one of three confidence tiers: informational, advisory or determinative.
The tier is recorded with each decision in the AI decision log.
A record of what was decided
Advisory and determinative outputs and the reviewer's decision are recorded in a hash-chained AI decision log.
The boundary
What waits for a person
Site Scan findings
SafeGora suggests possible hazards with the standards they relate to; a supervisor accepts or rejects each one, and accepted findings become corrective actions.
Drafts
SafeGora drafts JHA steps, root-cause options, meeting agendas and summaries. Drafts wait in a review queue until a person edits and approves them.
Determinative outputs
Outputs in the determinative tier are held in a review queue and can't be accepted until a person approves them. The decision is recorded.
Filings
The AI doesn't submit filings. Sending OSHA records or first reports to an agency is a person's act.
The exceptions
Where the product acts first
Two parts of the product act before a person reviews them.
Defect Snap equipment checks
When a photo check finds an equipment defect at or above the lock threshold (severity 3 by default), SafeGora places a provisional do-not-operate hold and leaves the detection pending. That hold is not a human-confirmed detection. A competent person confirms or overrides it later. Detections below the threshold stay proposals and wait for a person.
Changes made by the assistant and agents
Write-class tools produce proposals that a person approves. They do not auto-commit. The only exception is the named digest/notify tools of an automated run, which write notification and outbox rows only, are idempotent, and are capped per run. Starting the run is not a decision. Corrective-action tools commit only after an explicit approver, and the record names who approved. An assistant can propose a verification but cannot record one.
Keep exploring