Source documents and working tools.

Documents we have read in full and cite by section, each linked to the issuer's own copy. The last column of the first table is the one that matters to an agency and the one most coverage leaves out.

AI instruments by jurisdiction

Every state insurance department, swept for an AI bulletin, circular letter, regulation or guidance. Six were read in full for Finding 001 and are marked. The rest were retrieved and not yet read: the instrument name, date and verb are as the department's own page states them, and nothing here is cited as a finding until it has been read.

Tier: retrieved, not read in full, except the six marked. "Reaches producers" is "Yes" only where agents, producers, licensees or representatives are expressly addressed. Re-swept monthly; a row that changes gets a corrections entry if it was ever cited.

All links retrieved 16 September 2026; the ISO forms link to the Big "I" summary because the form text is not public. Six states reviewed. Other obligations may arise under privacy and security law, carrier appointments, or delegated authority agreements. Absence from this schedule is not a finding that nothing applies.

Pre-send verification checklist

Eight tests to run on an AI-assisted proposal before it leaves the agency. Every threshold is ours, labeled as ours, and is not a regulatory requirement.

1. Source-document reconciliation.Every premium, limit, deductible and named insured in the output traced to a page in the carrier quote packet. Any untraceable dollar figure: stop.
2. Coverage and exclusion comparison.Statements checked against actual policy and endorsement wording, with form number and edition recorded. Wording unavailable: request it and label the point unverified. Silence is not coverage.
3. Version comparison.Renewal proposal diffed against the expiring program; every changed limit, deductible or exclusion flagged. An unflagged change: stop.
4. Missing-data check.Required fields absent from the packet are listed as "not provided." Any inferred value presented as quoted: stop.
5. Repeatability.Same packet run again, compared field by field. A material numeric difference: escalate. A repeat run detects instability, not correctness.
6. Silent-omission check.The reverse direction: walk the packet and confirm each coverage part appears in the proposal or is deliberately excluded. This is the failure the first test cannot catch.
7. Citation check.Every quoted policy language, statute or carrier statement verified against the document. Unverifiable: remove the assertion or stop.
8. Named sign-off and log.A licensed producer signs; tool, model version, template, reviewer, timestamp and defects are logged. Without a named reviewer the proposal does not go out.

Structure adapted from ASOP 56 §3.6 (reconciliation, repeatability, alternative-model comparison) and NIST AI 600-1 actions MP-2.3-001, MS-2.5-003, MG-2.2-001, GV-1.3-002 and MG-4.3-002. If you want a number that means anything: defects per 100 proposals, per 1,000 traced fields, share of proposals with at least one stop, each with its numerator, denominator, period and compiler.

Agency AI policy clause outline

Eight clauses, each with its published basis and its tier. An editorial outline, not a ready-to-sign policy; your counsel and E&O adviser set scope, retention and responsibilities.

Editable versions of these tools, and three more, are in the working resources: blank worksheets you can fill in, not forms that submit anything.

Questions that end a demo early

What does the accuracy figure count, and what is the denominator? Are human corrections included in it? Who is live on this, and for how long? What broke?

A real deployment has a broken story. A pilot doesn't.