← Field notesAI + Human ReviewPublished

§Sheet B-AI-S

Why AI should support reviewers, not replace them

Permit decisions belong to people. Here are the design rules that let AI assist without taking the call: human approval, citations, confidence, override, and audit.

Author
PermitOS Team
Read time
6 min read
Published
Date not set

Where the line sits

PermitOS is decision-support. The line is bright: PermitOS reads, retrieves, and surfaces. People decide.

That line is not a marketing position. It is a product constraint we use when designing every feature.

Human final approval

No PermitOS recommendation results in an issued permit. Every approval is human-issued by the authority having jurisdiction. The product surface for reviewers is intentionally built around 'confirm, dismiss, or request more info' — not 'approve'.

Source citations

Every flag carries a source. The source points to a specific document, page, or rule reference — not 'the model thinks'. If a recommendation does not have a verifiable source, it does not ship as a recommendation.

Confidence levels

Recommendations carry a confidence value and a severity. Reviewers see both at a glance. Low-confidence flags are surfaced as 'needs human review,' not as facts.

Reviewer override

Override is first-class, not a hidden option. Confirm, dismiss, or request more information are equally weighted controls. The product never makes 'agree with the recommendation' the path of least resistance.

Audit trails

Every recommendation and every reviewer action is recorded with timestamp, actor, and context. The point is not surveillance. The point is that the public record reflects who decided what, and on what basis.

TagstrustreviewerauditAI

§Get started

Run a precheck on your next package.

Catch the preventable issues before formal review.