Daniel Elechiguerra will happily throw himself out of an airplane. Ask him about his career, though, and he describes someone who grew up wary of risk, in a country where, he says, taking that kind of career gamble is unusual. His father, he recalls, was surprised that he even went to business school.
That tension between the thrill-seeker and the worrier turns out to be a pretty good résumé for his current job. Elechiguerra is a co-founder of Harbor, a San Francisco startup that wants to insure the AI agents companies are now letting loose on loans, payments and customer service. Harbor emerged from stealth on Oct. 13 with a $10 million seed round led by Base10, Abstract and 8VC.
"I always say it's not so rational to start a company," he told Insurance Business in an exclusive interview. "You have to be crazy."
Elechiguerra trained as an engineer and taught himself much of his machine learning as an undergraduate, long before "AI" became a boardroom word. He went to MIT for a master's in business analytics to formalize what he had picked up on his own.
MIT also showed him an interesting side to the engineering world. Technical people, he noticed, tend to chase narrow metrics without asking whether those numbers actually move the business. He wanted to understand the entrepreneurial side of business, so when his Fulbright scholarship required him to return to Spain, he joined McKinsey.
There he worked largely on risk modeling for financial institutions, with a little insurance mixed in. A consultant who could write code turned out to be a novelty. "Everybody was impressed," he said. Harbor's company page also lists a stint at Citadel.
The itch to build something of his own kept growing, but long consulting hours left no room for a side project. In his mind Stanford's Graduate School of Business offered a way out: a couple of years, as he put it, to try and fail without much downside.
And to make sure he wasn’t tempted by a retreat to employment, he didn’t leave himself an escape route. While classmates competed for summer internships, he skipped one. He didn't line up housing for the following year either. "I burned all the boats that I could," he said. "When you're risk-averse, you really need that push."
He is quick to say the timing helped. He had no children, and a failure would not have been the end of the world. "It was kind of like the right moment, right place, right setup," he said.
The risk-taking, he insists, is selective. He calls himself very risk-seeking as an investor and an adrenaline fan in his free time, skydiving included. It was career decisions, where the downside felt real, that he had to push himself through. Inside Harbor, that cautious streak now has a job: he often plays devil's advocate when teammates get carried away, a habit he thinks suits an insurance company.
The idea came quickly. In his first month at Stanford, Elechiguerra was already brainstorming with Jesús Santos, a younger Spaniard he knew from back home who had once come to him for advice. Santos, now Harbor's chief executive, had won a full-ride scholarship to Stanford's master's in management science and engineering. Before that, he graduated as class valedictorian in electrical and mechanical engineering at Madrid's Universidad Pontificia Comillas, having switched to University after choosing education over his budding soccer career in Toledo.
Santos packed a lot in early. He interned at McKinsey in Madrid and at Morgan Stanley in London. He also spent a year in strategy and operations at Crossmint, an agentic payments company, where he says he was the youngest hire. Before any of that, Elechiguerra said, Santos played youth football in Toledo against the academy sides of Spain's top clubs, then chose university over the pitch.
The pair soon pulled in two more founders. Cristian Bartolomé brought the deepest AI résumé of the group. He co-founded Factored, which builds machine learning teams for companies, and was co-founder and CTO of Pulsar, an AI platform for manufacturers. With DeepLearning.AI, he helped build the first version of Andrew Ng's MLOps specialization. He also served three times as a course assistant on Stanford's CS230 deep learning class, and spent a summer as a visiting researcher at UC Berkeley on NASA's Saffire program.
Luis Botín, whom Harbor lists as a Stanford math and computer science graduate with a background in AI safety research, took a different route. At Stanford he co-directed the social impact arm of BASES, the student entrepreneurship society, which ran a $100,000 startup incubator. He did summer stints in data at Milan's P101 and in software at the climate startup CO2 Revolution. He also competed in horse racing until an injury ended it. Elechiguerra jokes that this may explain the team's fixation on safety.
All four are engineers, as is most of Harbor's staff, and that is the point. Measuring an agent's risk, Elechiguerra argues, is the hardest part of the job.
Insurance appealed precisely because so few people find it exciting, he said. If the industry doesn't draw much engineering talent, a team of engineers might find an edge there.
They toyed with consumer ideas first, including data-driven car insurance in the spirit of Lemonade. They dropped them, he said, because in personal lines the battle is won on distribution rather than data. He also admits he had little appetite for selling cover to people who often don't understand what they're buying. Commercial buyers, who shop with more sophistication, and brand-new risks that incumbents are slow to touch looked like more fertile ground for growing a unicorn.
For Elechiguerra and his compatriots, AI agents ticked both boxes. If software is now doing work people used to do, the founders reasoned, someone needs to insure its mistakes. And while the losses an agent can cause look a lot like a human employee's, the way it fails does not. "Some agents might just not fail at all," Elechiguerra said, "and some others that have very poor governance and very poor controls can fail thousands of times before somebody even notices."
Pitching insurance to venture capitalists, an industry where many have been burned before, was its own test. What helped, he said, was that Harbor's risk-assessment engine could stand alone as a product for other carriers, which lowered the bet for investors. The round closed around late April or early May, months before Harbor announced it. Once a few investors were interested, he said, the rest came more easily, though he describes the process as stressful.
Read next: It may take real losses to sell AI insurance: CRC Specialty CEO
Elechiguerra says that many think that agent risk can simply be folded into cyber. He is unconvinced. Too often, he said, cyber gets treated as the bucket for anything that happens to be digital.
His case rests on how differently people and machines go wrong. A professional's error might trace back to poor sleep, a late night or an argument at home, factors no underwriter can efficiently measure. An agent's risk profile, by contrast, shifts mainly when it is redeployed, and between releases its behavior should stay roughly constant. That, he said, makes it something you can actually test.
Insurers could still bury the exposure inside existing lines and charge enough to stay profitable, he conceded. "But there's no point in doing so," he said, "because you can actually become more sophisticated and get risk signals and then price properly."
The standard market, meanwhile, is drawing its lines. ISO's generative AI endorsements became available for commercial general liability policies in January, and a review of state filings found the exclusion already in force on thousands of CGL policies across 49 states and Washington, DC. Berkley has gone further with an absolute AI exclusion on its D&O, E&O and fiduciary products. Researchers have also warned that carriers may be sitting on AI liability they never priced.
Elechiguerra draws the line at autonomy. A chatbot used like a search engine is just another tool, he said, and will probably stay inside existing covers. "You never excluded Google from your liability," he noted. Once an agent acts with no human in the loop, though, the risk distribution "completely changes."
He also expects friction in professional liability. Firms whose staff paste AI output into client work without checking it may find themselves in gray areas, or even excluded, he said, and E&O carriers could see claims climb as a result.
So how do you underwrite a piece of software? Harbor's approach, as Elechiguerra described it, works in three layers, each more revealing than the last.
First, the client grants access to the platform where its agent activity is already logged. Harbor reads those traces to see what each agent has done in production, which tools it calls and what workloads it handles. Second, using an API or sandbox the client provides, Harbor's engineers try to break the agent themselves, pushing it into situations its history may never have tested. Third, for clients willing to open up further, the team can review code repositories and prompts. He called that last step intrusive but very high signal, and said it usually isn't needed.
The tests are scoped using a risk taxonomy that Elechiguerra said was built with the Stanford AI Safety Hub, then mapped onto each specific use case to estimate severity and probability of failure. On its website, Harbor says its agent risk framework spans operational, security and regulatory risk and is mapped to the EU AI Act, ISO/IEC 42001, OWASP's Top 10 and the NIST AI Risk Management Framework, among others.
The output is a set of scores and reports that clients can use on their own, and that Elechiguerra called the strongest signal feeding Harbor's pricing. The goal, he said, is to maximize the signal collected relative to the friction imposed on the buyer. He added that clients willing to share more data also tend to be the better risks, a useful selection effect for a young book.
Read next: AI governance isn't keeping pace with AI adoption, AXA XL warns
Harbor intends to act as its own MGA, pricing with its own risk models and distributing the product itself. Elechiguerra said the company holds the necessary licenses and was in London recently to finalize capacity, though he declined to name the partners.
That detail matters for brokers. Harbor's launch post says it will insure the residual risk "with our own money behind the answer," while the MGA structure Elechiguerra described suggests a carrier partner will issue the paper. Limits per client and the identity of that carrier remain the open questions hanging over Harbor's bid to insure the gap left by spreading AI exclusions.
Harbor will not be alone. Lloyd's coverholder Armilla AI raised its standalone AI liability limit to $25 million per organization in January, and the Artificial Intelligence Underwriting Company launched last year with a $15 million seed round to pair insurance with an agent safety standard.
Read next: Shadow AI is already in your organization. The question is whether you know what it's doing
For now, Elechiguerra's biggest need is a seat at brokers' tables. New MGAs, he said, risk becoming just one more product in a crowded pitch deck, even in a niche with few rivals. "It's always hard to distribute to the distributors," he said.
He was candid that the debate over who pays when an agent misfires is colored by incentives. Brokers have their reasons to talk up the gap, he said, while capacity providers lean conservative because they carry the loss. How much each side actually understands about AI shapes the view, too.
He also sees a consumer angle emerging, even as the company targets enterprises. Asked about young people handing AI agents access to their bank accounts, he admitted his own team has had unwelcome surprises with personal agents. But he expects the practice to become normal, and said Harbor has seen strong demand for cover on agent-driven transactions. Insurance, he said, may be the missing trust layer: people accept something can go wrong, but don't want to absorb the loss themselves.
Not everyone thinks buyers are ready. CRC Specialty's Neil Kessler recently said existing AI products aren't flying off the shelves, and that it may take real losses before boards set aside budget, much as happened with early cyber and EPLI.
Elechiguerra believes the leap was the right call, though he adds that he hopes it goes well. "The more I look at it, the more obvious it was, even though it felt risky," he said. What's left, in his view, is up to the team. Asked whether Harbor could become a unicorn, he said the team dreams of it. "I really think it's just execution risk," he said. "So it's on us to actually deliver."
