but firstsee the named perils
chapter i · the spark
every fault ends well.
on the sistine ceiling, two hands reach for each other and almost touch. that gap is where you live now. you built a mind, gave it your tools, and let go. the ideal hangs in museums. the real ships on friday. we underwrite the difference: the space between what your agent promises and what it does at 3am.
your spec, your evals, your promises to customers. the agent your team designed on the whiteboard. we read it like scripture, because it sets the bar every claim is measured against.
then we watch what the agent actually does in production. logs, traces, tool calls, edge cases. the distance between the ideal and the observed is the risk. and the premium.
before we underwrite, we interrogate. adversarial prompts, injection attempts, runaway-loop drills. we keep asking until your agent breaks. so the world doesn't get to ask first.
chapter ii · the crossing
shipping an agent is marching it over the alps: glorious on canvas, cold in practice. since january 2026, iso's cg 40 47 generative-ai exclusions have been attaching to gl renewals carrier by carrier, and ai carve-outs are appearing in cyber and e&o forms. goodfault writes the mountain's perils affirmatively, by name.
rogue refunds, mistaken discounts, wrongful payments, and unauthorized purchases the agent makes without a human.
the air canada peril: the agent invents a policy, price, or promise, and a court makes you honor it.
pii exfiltration, cross-tenant exposure, and ip disclosure caused by the agent's own actions.
deleted records, wiped configs, broken pipelines, and production systems the agent damages.
third parties manipulating your agent through prompt injection into any of the above.
adverse automated decisions in hiring, lending, or pricing, and the aggregation exposure they carry.
investigation and defense cost for eu ai act, revised product liability directive, and sector regulators.
your own downtime when a failing agent has to be pulled, paused, or rolled back.
no kill-switch, no logging, no injection testing, uncapped unattended financial authority, or anything cbrn / critical infrastructure. we don't insure agents without a recognized audit. that's a position, not a limitation.
one engine · two divisions
the risk was never about which model the ai runs. it is about what the ai is allowed to do, and how much damage it can cause before a human steps in. one underwriting engine prices all of it. a humanoid running an agentic policy model is both an agent and a machine, and we are built for where that line ends up.
division a · bits
customer support, commerce, coding, operations, screening, and workflow agents. eight named perils, written affirmatively, from rogue refunds to regulatory defense.
division b · atoms
when ai gets a body, errors become bodily injury and property damage. not just financial loss. humanoids, warehouse fleets, drones, ai devices, and the software layer of autonomous vehicles.
chapter iii · the watch
we are a carrier, not an auditor. hold a recognized ai-agent audit and you're in scope. upload it, get terms. no audit yet? we refer you out to recognized auditors and bind when you're certified. keeping those two roles separate is what makes our paper neutral.
a short intake: what it does, how much autonomy it holds, its financial authority, and its blast radius.
premium keyed to what the agent is allowed to do unattended. not to which model powers it. drop the authority, drop the price.
optional telemetry feed from your agent logs. renewal tracks real behavior, not a stale questionnaire, and earns a trace discount.
every policy is a story with chapters: the crossing, the watch, the wave, the chorus. walk them in order. or skip to the one you're living right now.
chapter iv · the wave
one day the wave arrives. these illustrative scenarios show what happens next, end to end. real named clients replace them as the book builds.
gf-0001
commerce co · certified support agent
a support agent was socially engineered into thousands of wrongful refunds over one weekend. the code-enforced per-transaction cap and daily circuit breaker, both policy conditions, contained it.
paid in 6 days via parametric trigger
gf-0002
saas · coding agent with prod access
a coding agent wiped a customer table. the rollback condition in the policy meant four hours of interruption instead of four days. and because the control worked, the premium fell at renewal.
4 hours of downtime, not 4 days · premium down at renewal
gf-0003
ai vendor · aiuc-1 certified
an ai vendor already held an aiuc-1 certificate. they uploaded it, the audit mapped straight to our intake, and coverage bound the same week with no back-and-forth.
zero re-audit · certificate did the underwriting
gf-0004
startup · no audit on arrival
a startup arrived with no audit and was out of scope. we referred them to a recognized auditor, they passed, and they returned to bind. we never touched the audit. that's the point.
then insured · neutrality intact
gf-0005
eu deployer · ai act obligations
an eu deployer used its certificate plus a goodfault policy as documented risk-transfer evidence. against high-risk fine exposure of up to 3 percent of global turnover.
compliance evidence in place before the clock runs
your agent's first fault is on us
start a policy in 20 minutes →
chapter v · the chorus
said in public, on the record. we collect the chorus because we underwrite exactly what it describes.
“every major technological leap in history, from electricity and automobiles to flight and the internet, was unlocked by private market insurance.”
epilogue · above the fog
what if the agent loops? what if the model drifts? what if a customer screenshot goes viral? this is where the story ends when you're covered: standing above the fog instead of inside it. the sea of what-ifs is still down there, it always will be, but every failure mode is named, capped, and carried. the view is the product.
tell us what your agent does, what tools it can reach, and what a bad day looks like. we return a premium in plain english. no actuarial latin, no 40-page pdf. most policies bind in under a week.
get a quotenot ready for a quote? describe your agent to the oracle and get an honest read of your riskiest failure mode. the one your roadmap is politely ignoring. free, anonymous, occasionally prophetic.
