Skip to content

Human-in-the-Loop Governance

"Human in the loop" is easy to claim and easy to hollow out. A confirmation dialog that everyone clicks through is not a control.

This page describes how the human gate is actually enforced in OrviQ, and how to tell whether it is working in your tenant.


Three structural mechanisms

1. Permission separation

The permission to generate is always distinct from the permission to adopt.

GenerateAdopt
mapping.ai_proposemapping.review
expected_evidence.ai_suggestexpected_evidence.accept
adequacy.runadequacy.submit, then adequacy.review
inspection.extractinspection.canonicalize
inspection.ai_assistinspection.signoff
control_source.ingestcontrol_source.publish
library.smart_extractPublication of candidates
incident.ai_assistincident.manage or incident.investigate

Holding a generation permission without its counterpart means you can produce output and adopt none of it. That is a genuine separation, not a naming convention.

2. Non-authoritative entry states

AI-produced records always enter in a state that carries no authority:

ArtefactEntry state
AI-proposed mappingproposed
Extracted requirementCandidate, unpublished
Extracted controlCandidate, unpublished
Inspection observationCandidate, unconfirmed
Design adequacy checkDraft recommendation
Suggested expected evidenceSuggested, unaccepted
Advisory draft textNot a record at all

There is no code path by which an AI pass writes an approved, published or confirmed record.

3. Segregation of duties still applies

Triggering an AI pass does not exempt you from maker-checker. If you run an AI mapping proposal, you are the maker on every mapping it produced — and you cannot approve any of them.

This closes the obvious loophole: without it, a user could generate proposals and approve their own generated work, which is self-approval with extra steps.


The three recurring patterns

Every AI capability in OrviQ follows one of these.

Pattern A — retrieval only

The AI finds things. It creates nothing.

Control discovery retrieves ranked control candidates for an obligation. Nothing is written.

No gate is needed because nothing entered the system.

Pattern B — propose and approve

The AI creates a non-authoritative record. A different person makes it authoritative.

AI mapping proposal writes mappings in proposed. mapping.review approves them, and cannot be the proposer.

Pattern C — draft and author

The AI produces text. A person writes the record.

Advisory drafting across incidents, audit, inspections and resilience. The draft has no standing; the person authors the record.


Testing whether the gate holds in your tenant

Four checks worth running during evaluation or an internal audit of your OrviQ configuration.

1. Turn the AI entitlement off. Every governed workflow should continue to work. Registers, approvals, determinations, reports and reconstruction are unaffected. If something breaks, that would be a dependency worth knowing about.

2. Check for combined permission grants. Look for roles holding both mapping.ai_propose and mapping.review, or both expected_evidence.ai_suggest and expected_evidence.accept. The platform still blocks self-approval on individual records, but a role holding both sides of a gate weakens the design intent.

3. Look at first-time approval rates on AI-proposed records. A queue of AI-proposed mappings approved at 99% first-time is a queue nobody is reading. The worked example shows a realistic distribution: 36% approved as proposed, 29% adjusted, 21% rejected, 14% recorded as no-match.

4. Check whether anyone records negative conclusions. A crosswalk with no no_match records means nobody is recording that a suggestion was considered and rejected — which usually means suggestions are being approved or ignored rather than assessed.

Rubber-stamping is a configuration and culture problem, not a platform one

OrviQ can guarantee that a different person clicked approve. It cannot guarantee that they read anything.

The four checks above are how you find out whether your gate is real. The third one is the most revealing.


What the human gate is not

It is not a delay mechanism. The point is judgement, not friction. Where a decision genuinely needs only one pair of eyes, use a four-eye chain rather than adding stages nobody exercises.

It is not a liability transfer. A person approving AI output owns the decision fully. "The model suggested it" is not a defence, which is exactly why the provenance record exists alongside the approver's name rather than instead of it.

It is not optional for volume. An import or an AI pass producing 3,000 proposals produces 3,000 items awaiting review. If that is infeasible, generate less — see the crosswalk import guidance. An unreviewed backlog is honest; a bulk-approved one is not.


OrviQ Enterprise Governance, Risk & Compliance Platform