Mechanical preparation
Extraction, formatting, comparison, classification, and bookkeeping with source preservation.
Applied AI methodology
Automate proposals, not decisions.
A practical operating model for using AI to prepare work, surface evidence and uncertainty, and reduce administrative burden without transferring consequential authority to the system.
The distinction
Preparation can be automated. Accountability cannot.
Many operational workflows contain both routine mechanics and consequential judgment. Treating them as one undifferentiated automation problem creates false confidence and weak controls.
AI may
Read, compare, extract, organize, draft, and recommend.
These activities reduce clerical burden and help people focus on missing information, conflicting evidence, unclear ownership, and decisions that actually require judgment.
AI must show
Evidence, confidence, assumptions, and unresolved uncertainty.
A recommendation without traceable support is not decision-ready. Ambiguity is an output to manage, not something the system should conceal.
People retain
Authority over canon, exceptions, release, and consequential action.
The accountable reviewer decides what becomes authoritative, what needs escalation, what is accepted with qualification, and what should not proceed.
Four control tiers
A workflow should become more restrictive as consequences increase.
The design boundary is defined before automation is introduced, not after a failure.
Evidence-backed proposal
Recommendations carry rationale, confidence, source references, and visible uncertainty.
Reversible staging
Draft packets, review queues, and proposed updates remain reversible until confirmed.
Human authorization
Canon changes, approvals, submissions, eligibility, and other consequential actions require accountable review.
Operating pattern
Design the review experience, not only the AI step.
- Separate source evidence from generated interpretation.
- Represent missing information and conflicts explicitly.
- Make proposed changes easy to compare with the current state.
- Keep staging reversible until a person confirms release.
- Record what was accepted, rejected, or modified.
- Use the review record to improve rules and training over time.
Applied proof
The method is visible in the work.
Meeting intelligence, document intake, status reporting, request triage, knowledge assistance, and the CCSF AI Interview Coach all use the same principle: AI prepares useful work while evidence, exceptions, purpose, and authority remain visible.