On October 13, a San Francisco startup named Harbor came out of stealth with $10 million in seed funding and a product category that did not really exist a year ago: insurance for when an AI agent makes the mistake. The round was led by Base10 Partners, Abstract, and 8VC, according to Insurance Journal’s reporting. The four founders met at Stanford. They work with banks, insurers, and other regulated companies — the kinds of firms now handing real work to software agents.
Harbor matters to this desk for a simple reason. The question it is built on — who pays when the agent errs — is the same question underneath every errors and omissions conversation happening right now, asked about a worker who is not a person.
What is Harbor actually selling?
Two things, and the order matters.
First, a measurement platform. Harbor finds every AI agent a company is running and builds an inventory. It measures each agent’s operational, security, and regulatory risk against a framework mapped to the NIST AI Risk Management Framework and the EU AI Act. It checks agent decisions against the company’s own rules and blocks the ones that break them before they go through. And it re-measures every time an agent’s model, tools, or data change — because, as co-founder Jesús Santos argues, a risk that shifts every time the model shifts cannot be assessed once and filed away.
Second, insurance on what remains. The measurement is what makes the risk insurable. Whatever risk survives the controls, Harbor insures. That pairing is the whole idea: when the insurer pays for the loss, the insurer has every reason to make the measurement honest.
Why is an agent’s mistake different from an employee’s?
When an employee makes an error, the loss is usually a single event, and the existing tower responds the way it was designed to. Professional liability, cyber, and technology E&O policies pay for human mistakes every day. That is what they are for.
An agent’s error behaves differently. The same wrong decision can repeat hundreds or thousands of times before anyone notices — every transaction processed the same wrong way, every customer answered with the same wrong answer. By the time a human sees it, the single mistake has become a portfolio of losses.
Carriers see that shape and face a pricing problem: there is no claims history for it. Harbor says some insurers are responding by excluding AI from existing policies rather than underwriting a risk they cannot yet price. That exclusion trend is the gap Harbor was built to stand in.
Where does that leave the coverage a business already has?
Mostly, in need of a careful read. A professional indemnity or tech E&O program written for human work may say nothing clear about losses caused by software acting on its own instructions — and silence in a policy tends to get resolved the hard way, after the loss. The questions worth asking at renewal are suddenly specific. Does the wording contemplate automated decision-making? Has an AI exclusion or sublimit appeared in this year’s forms? If an agent repeats an error for three months before discovery, how does the claim get counted — one occurrence, or thousands?
That last question is not academic. It runs straight into the claims-made versus occurrence mechanics every producer already knows, with a fact pattern the forms were not drafted around.
Why should insurance people watch a four-day-old startup?
Because the exposure is arriving inside insurance’s own house first. Agents are being put to work on policy checking, intake, and document handling across the industry — the front line we covered in the policy-check agent piece. Every agency and carrier deploying those tools is creating exactly the exposure Harbor wants to insure. A market forming around your own operations is worth watching earlier than a market forming around someone else’s.
There is also a pattern here worth naming. When a new kind of work appears, the coverage follows a sequence: existing policies get stretched over it, then exclusions carve it back out, then purpose-built products price it properly. Agent risk is visibly in stage two moving into stage three. Harbor is an early stage-three entrant, and it will not be the last.
What don’t we know yet?
Plenty, and the honest answer matters more than a guess. Four days after launch, the public record does not say who carries Harbor’s paper, whether the coverage will be placed through brokers or sold direct to enterprises, what limits and pricing look like, or which customers have bound. Those are not small details — for producers, the distribution question is the whole ballgame, because it decides whether this becomes something you can sell or something that competes with you.
Watch for the follow-up announcements. The launch told us the category exists. The next release — a carrier partner, a distribution channel, a named program — will tell us how the category actually reaches a buyer.
The bottom line
A mistake that repeats itself a thousand times is not an employee error with a bigger number attached. It is a different kind of loss, and the market has started pricing it as one. Whether Harbor becomes the name attached to that market is unknowable in week one. That the market itself is forming — funded, staffed, and aimed at regulated industries — is now a matter of record.
Sources: Insurance Journal (Oct 13, 2026); FinSMEs (Oct 13, 2026); Digital Insurance, founder interview (Oct 2026); ProgramBusiness (Oct 2026).