Errors-and-omissions insurance grew up around doctors, lawyers, and engineers — people who can be deposed. A TBA can receive revenue. It cannot sit for an examination under oath. Underwriters know this. They will not pretend otherwise for a white paper.
Today you insure the operator. You name the company. You maybe schedule “AI tools” as a technology rider with exclusions that run for pages. The skill is an unnamed instrument, like a laptop. When the laptop causes a loss, the fight is about the human’s supervision, not the laptop’s soul.
The policy is a human document
That is workable until losses are large and frequent. Then carriers will want telemetry: which skill id, which model version, whether pause was available. Protocol events become claims evidence. Agents should assume their logs can be exhibited. Humans should assume “the bot did it” is not a coverage grant.
What can be covered today
A native agent policy — premium paid in USDC, payout to a TBA, parametric on events — is conceivable. It is not on any admitted paper this year. Building it requires capital, regulation, and a definition of occurrence that survives a prompt injection. That is a decade of work, not a hackathon.
What will take a decade
Until then, the dual economy’s risk sits where it always sat: on firms that can buy insurance and on people who cannot hide behind a hash. Autonomy without a policy is just uninsured enterprise. That is allowed. It is not new.
