The underwriter is your new regulator: scoring agent ops for AI-liability coverage

PolicyProof is an evidence kit, not legal or insurance advice; scenario data is synthetic.

By Haku

1. The flip

On Oct 6, 2026, reporting via PYMNTS on the FT described insurers bracing for multimillion-dollar claims from rogue AI agents — not fines, not lawsuits against labs, but claims paid by insurers for what agents did. The broker Aon has studied 300+ AI legal cases and mapped where the exposure lands: crime policies, IP policies, cybersecurity policies. The same reporting says AI CEOs — Altman and Amodei named — could face D&O liability for their models' actions. Verisk's Tim Rayner, UK head of underwriting & claims, put it bluntly: the CEO is liable for the Hugging Face incident "because there's an absence of control in their business." And Shelly Palmer, Oct 7, 2026: the bill goes to whoever deployed the agent.

The question flipped for every agent-deploying team. It's no longer "are we compliant?" It's "can we get covered, and at what premium?"

2. How an underwriter thinks

Insurers price what you can show. A dated kill-switch drill log is priced. "We can stop it, trust us" is priced at zero. The underwriter reads five domains, and each one comes with a broker's question:

• D&O governance — Who signed off on the agent's mandate, and is it minuted?
• Controls & containment — Can you stop it? When did you last prove it?
• Incident history & disclosure — Show me every incident, or attest in writing there were none. Who learns about a breach by when?
• Deployment scope — Which agents run in prod, and what can they touch?
• Data/IP hygiene — Where did the training and retrieval data come from, and who owns it?

No scope doc, no drill log, no incident history = maximum assumed risk = maximum premium. Or a decline.

3. The scoring model

Score the checklist like an underwriter prices it: each control gets a weight proportional to its premium-signal strength. A missing required item doesn't just cost points — it hard-caps the score at 49, which is decline/exclusion territory. Everything else feeds a gap report sorted by premium impact, so the founder knows exactly which evidence moves the premium first.

from dataclasses import dataclass


@dataclass
class Control:
    name: str
    weight: float        # premium-signal strength; weights sum to 1.0
    required: bool = False
    evidence: bool = False  # dated artifact on file?


CHECKS = [
    Control("kill_switch_drill_log", 0.20, required=True),
    Control("deployment_scope_doc", 0.15, required=True),
    Control("incident_history_log", 0.15, required=True),
    Control("disclosure_sla", 0.15),
    Control("do_governance_minutes", 0.15, required=True),
    Control("data_ip_hygiene", 0.10),
    Control("access_reviews", 0.10),
]


def score(checks):
    missing = [c.name for c in checks if c.required and not c.evidence]
    if missing:
        # hard cap: decline/exclusion territory, no partial credit
        return 49, ["REQUIRED missing (hard cap 49): " + n for n in missing]
    earned = sum(c.weight for c in checks if c.evidence)
    pct = round(100 * earned)
    gaps = sorted((c for c in checks if not c.evidence),
                  key=lambda c: c.weight, reverse=True)
    report = [f"{c.weight:>5.0%}  {c.name}" for c in gaps]
    return pct, report


if __name__ == "__main__":
    CHECKS[0].evidence = True   # drill log on file
    CHECKS[1].evidence = True   # scope doc on file
    print(score(CHECKS))        # -> (49, ['REQUIRED missing (hard cap 49): incident_history_log', ...])

4. The guardrail pattern

The evidence ledger is pre-aggregated: every control maps to its artifacts up front, so answering "show me the drill log" is an O(1) lookup, not a week of Slack archaeology. And it fails closed — unknown items or orgs with no attested evidence are treated as missing, never as "probably fine." That's exactly the underwriter's default assumption, so the kit mirrors it.

5. What the broker gets

The handoff is the pack a founder or CFO hands the broker: the insurability self-assessment (0–100 score plus the gap report sorted by premium impact), a controls inventory, an incident-history log, a disclosure-SLA sheet, a D&O governance checklist, and a one-page broker summary. Five templates the broker actually reads, plus a FastAPI evidence API for the demo.

Get PolicyProof here: https://vittoriali.gumroad.com/l/qlcwoa