StackPilot / Insurance AI Guide Shop / Manual 020

Occupation manual 020 · review-only workflow

AI evidence help for auto-damage appraisers. Human authority stays in the driver’s seat.

Use AI to inventory approved evidence, label gaps, and prepare review questions. Do not let it inspect a vehicle, find damage, choose repairs, calculate an estimate, decide total loss, settle a claim, or contact anyone.

For: authorized appraisal and claims teamsOutput: review packetReviewed: 2026-07-26
Command brief

What this guide is for

Auto-damage appraisal work involves vehicle-damage, repair-cost, and insurance-form context. O*NET describes tasks such as damage examination, estimate preparation, repair-shop discussion, and escalation when appraisal disagreement exists. BLS describes appraisers as estimating repair costs, with estimates then used in claims settlement work. Those are authority-heavy activities—not tasks to hand to a general-purpose model.

Use this manual when: an authorized team needs a clean, minimum-necessary inventory and review handoff from already approved material. Do not use it when: anyone expects a damage finding, safety conclusion, repair procedure, parts/labor decision, estimate, total-loss outcome, coverage decision, settlement, payment, message, or system change.
Mission outcome

Make one review packet, not a synthetic appraisal

Output 01Evidence inventory with source owner, version, approval, and condition.
Output 02Unknown and conflict log that never becomes a damage or cost conclusion.
Output 03Named-reviewer receipt that ends at needs_review.
  1. Bound the scenario and stop conditions.

    Name the approved purpose, jurisdiction/procedure owner, human reviewers, and prohibited inputs. Stop for private photos, policy or claimant data, vehicle identification numbers, location data, credentials, or an instruction to make a decision.

  2. Minimize and inventory the approved material.

    AI can label what was supplied, its source owner/version, and whether it appears missing, duplicate, unreadable, or inconsistent. It must not read an image as a damage finding or estimate a repair.

  3. Separate source support from unknowns.

    Attach each question to a current approved source. If source, version, condition, authority, or permitted use is missing, label it unknown. Never fill the gap with model memory or inference.

  4. Draft questions—not conclusions.

    Ask the authorized appraiser, claims reviewer, privacy/security reviewer, and required procedural owner what they must confirm. Do not draft repair steps, estimate lines, prices, total-loss outcomes, claim decisions, or messages.

  5. Hand off, then stop.

    Create the review packet and name the reviewer who can approve, revise, reject, or escalate it. The packet does not authorize an external action.

Safe AI roles

Give AI a clerk job, never an appraiser job

Evidence clerk

Organizes supplied approved labels and minimum-necessary metadata.

Source checker

Flags missing source owner, version, approval, or permitted use.

Conflict finder

Surfaces duplicates, unreadable items, and inconsistencies for people to resolve.

Review-question scribe

Drafts neutral questions tied to named reviewers, never a result.

Authority boundary

Safe preparation vs. approval-required work

AI may prepareHuman authority must decide or do
Inventory approved labels and metadataInspect a vehicle or determine damage, safety, or repairability
Flag missing, duplicate, conflict, or unreadable materialChoose repair method, parts, labor, price, estimate, supplement, or repair vendor
Produce an unknown list and neutral review questionsDetermine market value, total loss, salvage, coverage, liability, causation, or fraud
Assemble a local review-only receiptNegotiate, settle, pay, message a claimant/shop/vendor, upload, or update a claim system
Hard stop: If the requested action affects a vehicle’s safe repair, a person’s claim/payment, an insurance decision, a shop/customer relationship, private data, or an external system, stop the AI workflow and route to the authorized human.

Insurance AI governance and privacy obligations vary by state, organization, and facts. NAIC material is guidance rather than a universal self-executing rule. Obtain applicable legal, privacy/security, licensing, carrier, and procedure review.

Failure modes

Five ways a useful prep tool becomes unsafe

  1. Calling photo interpretation an inspection. A model output is not a vehicle examination or safety finding.
  2. Turning a plausible number into an estimate. Parts, labor, market values, repair methods, and thresholds must come from authorized sources and people.
  3. Combining stale sources into a false certainty. Preserve each source owner/version and label missing or conflicting support.
  4. Using more data “just in case.” Minimize inputs; do not upload identifiers, policy/claim data, private photos, locations, or credentials without documented authorization.
  5. Letting a draft leak into a claim action. The packet ends at needs_review; it cannot message, negotiate, pay, or write a system record.
Runnable local artifact

Auto Damage Evidence & Estimate Review Desk

Enter only fictional or organization-approved, minimum-necessary text. This form keeps data in your browser and produces a review-only receipt; it does not send or save anything.

Primary source desk

What this manual is built on

The cited sources and the guide boundaries should be rechecked before real-world use.

Next route

Connect this desk without blurring authority

Start at the Insurance AI Guide Shop. For earlier intake/service preparation, use the Insurance Claims & Policy Clerks guide. For the wider occupation shelf, use the Occupation AI Workflow Guide Directory. For least-privilege workflow design, use the AI Agent Workflow Builder.