StackPilot / Occupation guides / Manual 023

Occupation manual 023 · review-only workflow

AI can organize claim evidence. Claim authority stays human.

Use AI to preserve approved source labels, flag unknowns, and prepare reviewer questions. Do not let it investigate, decide coverage or liability, assess fraud or damages, calculate a reserve or payment, settle, deny, communicate, or update a claim system.

For: claims professionals and reviewersOutput: evidence & review packetReviewed: 2026-07-29
Command brief

What this guide is for

O*NET places claim review, investigation, claim-record work, settlement work, and communication inside this occupation. Those are consequential professional functions, not safe defaults for a general-purpose model. This field manual gives AI a smaller clerk role: make the evidence gaps, source versions, and human authority visible before any authorized person decides what happens next.

Use this manual when: a claims team needs a bounded, review-only packet of approved source labels, unknowns, and questions. Do not use it when: someone wants an investigation, a coverage/liability/fraud/damages opinion, a calculation, a status update, a claimant message, a record update, or a payment.
Mission outcome

Make one review packet, not a synthetic claim move

Output 01Source/version/owner register with stated permitted use.
Output 02Unknown-and-conflict log that never becomes a claim conclusion.
Output 03Named-reviewer receipt ending at needs_review.
  1. Bound the scenario and authority.

    Use a fictional or organization-approved scenario. Name the claims authority owner and the jurisdiction, line/product, policy/plan, privacy, records, and legal/compliance reviewers needed. Stop for a live claim file, private information, or a requested decision/action.

  2. Inventory only approved source labels.

    AI may record the source label, owner, date/version, and stated permitted use. It must not determine whether a source establishes coverage, liability, causation, fraud, damages, a reserve, or a payment.

  3. Make every gap explicit.

    When an authority, version, source, or approval is missing or conflicts, write unknown. Do not fill the gap with model memory, an assumed deadline, or a plausible claim narrative.

  4. Draft questions, never findings.

    Prepare neutral questions for the named claims, legal/compliance, jurisdiction/licensing, privacy/security, records, and model/vendor-risk reviewers. Do not draft a settlement, denial, payment, notice, or claimant communication.

  5. Hand off and stop.

    Create the local packet and allow the named reviewer to approve, revise, reject, or escalate it. The packet never authorizes an investigation, determination, communication, claim-system change, or payment.

Safe AI roles

Give AI a clerk job, never a claims-professional job

Source clerk

Organizes approved source labels, owners, dates, versions, and stated uses.

Gap finder

Flags missing, stale, duplicate, unsupported, or conflicting material without resolving it.

Question scribe

Drafts neutral reviewer questions tied to named human authority.

Authority boundary

Safe preparation vs. approval-required work

AI may prepareAuthorized human must decide or do
Inventory organization-approved source labels and metadataAccess a claim file or choose/apply a source to a claim
Flag missing, stale, duplicate, unsupported, or conflicting supportInvestigate, assess people/evidence, or determine coverage, liability, causation, fraud, damages, or status
Produce an unknown log and neutral reviewer questionsSet a reserve; calculate valuation, benefit, payment, settlement, or denial; interpret policy/legal requirements
Assemble a local review-only receiptCommunicate externally, update a claim system, pay, report, or take any external action
Hard stop: If the next request involves live claim data, a person, coverage, liability, causation, fraud, damage, a reserve, money, a deadline, a notice, a system record, or an external party, stop the AI workflow and route it to the authorized human under the current procedure.

NAIC materials provide regulatory context, not a universal jurisdiction or carrier rule. State adoption, product/line, carrier procedure, license/authority, and the controlling source must be verified before any non-fictional use.

Failure modes

Five ways a preparation tool becomes an unsafe claims tool

  1. Calling a summary an investigation. A clean source list is not evidence weighing, fact finding, or claim handling.
  2. Turning an unknown into a conclusion. A missing source/version/authority cannot become coverage, liability, fraud, damages, reserve, or payment through model confidence.
  3. Borrowing a rule from the wrong place. A NAIC model, a different line, an old policy edition, or another jurisdiction is never a controlling live claim rule without authorized confirmation.
  4. Adding private data “to help.” Do not paste claim, policy, medical, financial, image, payment, credential, or system data into a general-purpose tool without a separately authorized process.
  5. Letting a draft become an action. The packet ends at needs_review; it cannot investigate, decide, calculate, message, update, pay, or send.
Runnable local artifact

Claim Evidence & Review Desk

Enter only fictional or organization-approved, minimum-necessary text. This form stays in the browser and creates 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

Use the proposed Claims & Policy Service Evidence Desk for intake/service boundaries, the Auto Damage Appraiser Desk for vehicle evidence, and the Insurance Underwriter Desk for pre-decision underwriting evidence. These routes are proposals only until separately approved and integrated.