Your agents submit claims, contact patients, move money, and change records. At renewal, nobody can produce anything a carrier can price. You have a log of what the agent did. You don't have the authorization, the evidence, the failure model, or the named human. We produce that, in two weeks, as an independent party.
Delegated action scales in both directions. An agent with access to records, payments, and workflows can extend expertise across thousands of transactions; the same reach lets one flaw repeat before anyone sees the pattern. Three audiences now ask the same question and get different answers: the carrier at renewal, the board at the audit committee, the regulator when something goes wrong.
The market has already moved. Carriers are excluding AI by default and insuring by exception, and the exception goes to the insured who can show the paper.
Generative-AI exclusion records in force across 49 states by mid-2026, after ISO's CG 40 47 exclusion became available for general liability policies in January.
Insurance Business, Sep 2026 →Major carriers filed to adopt the ISO forms or their own equivalents by April 2026; one introduced an "absolute" AI exclusion across D&O, E&O, and fiduciary lines.
Same source →Underwriting topics now surfacing on applications: AI policies, approved use cases, testing, human review, access controls, monitoring, incident response, vendor oversight.
National Law Review, Jul 2026 →Five sections, each answering the question an underwriter actually asks, each backed by evidence rather than assertion. Delivered with a gap register and a readiness grade — one letter per section and one overall, with the sentence a CFO can repeat in the boardroom.
The enumerated actions, the systems and data it may touch, the dollar or clinical limits per action and per day, what is explicitly out of scope, and the person who signed.
Pre-deployment evaluation against the scope, in-production monitoring that compares each action to what was permitted, exception and override rates, sample size and period.
Three worst-credible-failure scenarios with reach, cost, and detection time; transaction volume at peak; downstream systems it writes to.
The kill switch and who can pull it, limits enforced in code rather than policy, the actions that require a human before execution, and a named accountable owner with a response-time commitment.
A change log for model, tools, permissions, and workflow; the events that trigger re-evaluation; the date of the last refresh and who signed off. A packet written against last quarter's agent is already stale.
The Pressure Test Delegation family. Every engagement produces the same packet, graded the same way, so a broker who has seen one knows how to read the next.
One named agent deployment, surveyed end to end.
Every consequential agent in the organization, in one exhibit.
The assessment has to evolve, so the relationship does too.
One agent, its business function, its current version, and the accountable human. A 45-minute scoping call settles it.
Documents, reports, query results, signed schedules — not descriptions of them. Where nothing exists, the row says so. The template is honest or it is useless.
The reviewer assigns grades and ranks every gap by how much a carrier would care and how long the fix takes. The two roles stay separate by design.
Your broker, your audit committee, your regulator. One packet pre-answers three audiences. A Standing Surveyor engagement keeps it current.
Delaware's June 2026 committee draft would let an AI agent run a company inside a 30-month regulatory sandbox. The liability shield for the humans behind it depends on three conditions: adequate capitalization, a maintained activity log, and disclosure of the company's autonomous status. We ran the packet against a hypothetical company formed under that draft.
| Section | Grade | Why |
|---|---|---|
| 1 · Authorization scope | C | "Designated to manage the business" is a grant of office, not a bounded scope of action. |
| 2 · Performance evidence | C | The activity log records what happened; nothing compares it to what was permitted. |
| 3 · Blast radius | D | "Adequately capitalized" is a number with no failure scenario behind it. |
| 4 · Containment | B | Sandbox administrator can shut it down; the member is named. External shutdown is last resort, not first. |
| 5 · Change cadence | D | No provision for re-evaluation when the model, tools, or permissions change. |
Capital + log + disclosure gets an agent through the door. Authorization + attestation + a named human gets it through underwriting. The distance between "legal to form" and "possible to insure" is a packet, not a platform.
Request the full worked exampleHealth systems and payers first; fintech and professional services second. You own the renewal, the audit committee, or the regulator relationship — and your organization has agents in revenue cycle, prior authorization, scheduling, or patient outreach.
Your incentive is to make the client insurable at a good rate. The packet is the submission form that doesn't exist yet for agent risk — standardized, graded, and shaped the way an underwriter reads.
Affirmative AI coverage needs a defensible baseline. The packet maps an insured's agent authority, evidence, failure model, and containment onto the questions your form already asks.
Preventive medicine is the discipline of understanding exposure, finding preventable harm, intervening, and learning from what happens. Underwriting is that logic with a price attached. Dr. Plummer brings clinical and public-health training, hands-on work on AI authorization frameworks — what a system may do, under what conditions, with what evidence — and a fiduciary-minded view of risk.
He writes publicly on AI accountability through The Governed Agency. This work is conducted in a personal, independent capacity and does not represent any employer or agency.
A 20-minute call is enough to tell whether a Position Read fits your renewal timeline. Brokers and underwriters who want to react to the packet format are welcome — a "we just exclude agents" is as useful to us as a yes.