Clerk preparation
- preserve intake and index supplied records
- check completeness against an approved list
- compare versions and flag conflicts
- draft narrow requests and neutral status updates
- assemble decision and approval handoffs
Short answer: AI can help a clerk create an accurate, reconstructable preparation packet. It cannot interpret coverage, decide liability or benefits, investigate fraud, change a policy, settle or deny a claim, send a message, write an account, or release money.
Occupational description is not authorityProcessing applications, checking completeness, and handling requests does not grant permission to execute a policy change or decide a claim. Authority depends on role, license/appointment, carrier or plan, product, contract, state, and facts.
Keep every authority distinct
Specialist authority
- producer/agent: licensed transaction and explanation work
- adjuster/examiner: coverage, investigation, evaluation, and claim outcome
- underwriter/carrier/plan: risk, terms, pricing, issue/change/cancel decisions
- SIU/legal/medical/repair/regulator/payment: their qualified decisions and actions
Build the ten-part Claims & Policy Service Evidence Desk
Load jurisdiction and authority
Record carrier/administrator, state, line/product, program, insured/claimant and loss/service locations, procedure version, clerk authority, prohibited actions, specialist contacts, and payment authority.
Output: jurisdiction and authority card
Preserve the intake event
Keep exact sender words, timestamps, channel, supplied references and dates, requested action, attachments/provenance, asserted deadline, accessibility needs, privacy class, and separate clerk/AI summaries.
Output: source-faithful intake
Verify identity and disclosure scope
Use the carrier-approved procedure and record the claimed role, human/system verification receipt, representation documents, disclosure scope, account-action scope, and conflicts. AI never authenticates identity or capacity.
Output: identity and authority gate
Inventory evidence without evaluating it
Index source, date, document type, subject, files/pages, hash, privacy class, evidence state, duplicates, conflicts, and relation to the request. Do not infer authenticity, causation, coverage, liability, damage, medical necessity, or fraud.
Output: evidence inventory
Stack policy and document versions
Connect policies, declarations, certificates, endorsements, riders, applications, plan documents, notices, transaction records, procedures, and confirmed rules. Compare text, but do not choose or interpret the controlling version.
Output: version-controlled document stack
Draft the narrowest missing-item request
Use an approved checklist, authorized reviewer instruction, applicable form, or confirmed rule. Show what is needed, what is already present, why, secure return channel, minimum-necessary check, real deadline source, reviewer, and payload hash.
Output: approval-ready request draft
Keep status neutral and correspondence exact
Use only audit-log facts. Verify recipient authority, identifiers, dates, jurisdiction template, quotations against the exact version, notices, privacy, attachments, accessibility, deadlines, promises, and decision language.
Output: status draft and correspondence QA
Route security observations, not fraud conclusions
Record predefined anomalies neutrally and send the packet to the authorized security, SIU, antifraud, privacy, or supervisor queue. Changed payment instructions trigger a hold and known-channel verification.
Output: security/fraud observation packet
Control privacy and decision routing
Name data categories, purpose, applicable or unresolved regime, minimum needed, approved systems, roles, retention source, disclosure log, and security route. Match each exact question to the correct authorized role without an AI outcome recommendation.
Output: privacy card and decision handoff
Make the file reconstructable
Append receipt, access, extraction, verification, draft, correction, escalation, decision-reference, and approval events with actor, timestamp, source, hashes, reason, authority, and privacy class. Bind the exact payload and approval.
Output: audit trail and approval packet
Evidence inventory is not evidence evaluation
Use states such as received_unverified, source_verified, carrier_record, third_party_statement, qualified_expert_record, duplicate, conflicting, illegible_or_incomplete, restricted, and uncertain.
No automatic upgradeA photo does not prove cause or damage. An estimate does not establish covered repair scope. A medical bill does not establish necessity or benefits. A duplicate invoice flag does not establish fraud.
Model rules are not universal rules
NAIC models, charts, bulletins, and maps are research tools. They do not automatically become binding law in every state. Before importing a deadline, notice, claims practice, privacy rule, fraud procedure, or AI control, verify the current state/territory source, line, plan/program, carrier procedure, and effective date.
Useful design principles
Accurate policy facts, prompt and documented communication, nonduplicative requests, reconstructable files, clear AI governance, privacy controls, and authorized escalation.
Required scope checks
Property/casualty, life, accident/health, disability, workers’ compensation, ERISA plans, government programs, admitted/nonadmitted business, and state rules cannot be blended into one generic insurance workflow.
Decision words require a decision artifact
A status draft cannot use approved, denied, covered, liable, eligible, fraud, final, bound, cancelled, or paid unless an authorized decision artifact already exists and the exact communication is separately reviewed.
Privacy scope must be verified
HIPAA does not apply to every insurer or insurance record. ERISA, workers’ compensation, consumer reporting, state insurance privacy, breach, and program rules have different scopes. Where applicability is unresolved, use the most restrictive configured path and route to privacy/legal authority; AI does not decide the law.
Use this prompt for private preparation
Act as an insurance claims and policy-service preparation clerk, not a producer, adjuster, examiner, investigator, underwriter, carrier, plan administrator, lawyer, medical or repair expert, regulator, payer, sender, or account operator. Using only supplied approved records, create the jurisdiction/authority card, source-faithful intake, identity/authority gate, evidence inventory, document-version stack, missing-information draft, neutral status/correspondence QA, security observation route, privacy card, decision handoff, append-only audit events, and exact approval packet. Do not interpret coverage or policy application; decide liability, eligibility, benefits, necessity, causation, value, fraud, price, premium, or payment; authenticate identity; investigate; negotiate; settle; approve; deny; bind; change; cancel; send; file; pay; write an account; or call an external mutation API. Stop at approval_ready.
Final review before any carrier or plan workflow
Official source desk
BLS: Financial ClerksOccupation description without transaction or decision authority.BLS: Claims Adjusters and ExaminersSeparates claim evaluation, investigation, settlement, and payment work from clerk preparation.NAIC Model LawsOfficial model catalog; state enactment and amendment control.NAIC Model 900Claims-practice design cues with state-adoption limits.NAIC Model 902Property/casualty file reconstruction and communication controls.NAIC AI Model BulletinGovernance, validation, fairness, security, and documentation expectations; not universally self-executing.NAIC: Data Privacy and InsuranceState privacy landscape and model frameworks.NAIC: Insurance FraudState fraud bureau and reporting context; no clerk/AI finding authority.HHS: HIPAA Covered EntitiesHIPAA scope; it does not govern every insurance file.FTC: Start with SecurityData minimization, access, secure lifecycle, and vendor oversight.FBI: Business Email CompromiseKnown-channel verification for payment/account changes.
Test with synthetic storm-loss records first.
Use duplicate photos, an incomplete estimate, a deliberate endorsement conflict, and unanswered coverage/payment questions. The desk should inventory, draft, QA, and route—then refuse the decisions.