The first useful thing to know about Justin Levine’s career is that insurance arrived sideways. He trained as a civil engineer at Georgia Tech and went to work in construction, where risk is less an abstraction than a daily companion. It lives in the subcontractor roster, the schedule, the quality log, the change order, and the hundreds of decisions that determine whether a project becomes a building or a cautionary meeting. At Hunter Roberts Construction Group in New York, Levine moved into enterprise risk. His first contractor-controlled insurance program was tied to VIA 57 West, Bjarke Ingels’s sloped Manhattan tetrahedron. He still calls it his favorite building. Nostalgia, in his case, comes with an insurance program attached.
That early view was from inside the contractor, not across a carrier’s desk. Levine could see how builders actually selected partners and managed work. He could also see how much knowledge was trapped in paper, spreadsheets, and the heads of experienced people. A conventional insurance application offered a still photograph. A jobsite produced a film.
A risk manager learns to build software
In 2015, Levine co-founded TradeTapp, a platform designed to help general contractors collect and analyze information about subcontractors. The idea came straight from the work. Choosing a trade partner is not clerical housekeeping. It is a bet on whether another company has the finances, capacity, safety record, and operating discipline to finish its part of a complicated project.
TradeTapp turned that qualification process into a structured software product. BuildingConnected acquired the company in August 2018. Autodesk then bought BuildingConnected in early 2019, carrying TradeTapp into a much larger construction technology system. Levine stayed in risk strategy and got a wider window onto the industry. Contractors were adopting digital tools for project management, documentation, reality capture, fleet operations, and quality control. The technology had begun to record how work was performed, yet insurance pricing seldom recognized those signals.
There was a business hiding in the mismatch. Construction software usually sold efficiency: fewer hours, cleaner coordination, faster workflows. Levine wondered whether the same tools could demonstrate something more valuable. If disciplined use of technology correlated with fewer or more predictable losses, software adoption could become an underwriting input. A contractor’s operating behavior might earn recognition in the price of insurance.
“The availability of clean, usable loss data.”Levine, on the one thing he would change in insurance
Shepherd starts where the forms stop
Levine co-founded Shepherd in 2020 with Mohamed El Mahallawy, a former Airbnb engineer, and Stephen Buonpane, a construction insurance veteran from Chubb. The combination mattered. Shepherd would need software people who understood systems and underwriters who understood what happens when an elegant theory meets a difficult account. It joined Y Combinator’s Winter 2021 batch and announced a $6.15 million seed round that September.
The company began with excess casualty for commercial construction, a line where a single program can involve several insurers stacked above one another. Speed is not cosmetic in that market. A broker may be assembling a tower of coverage while fielding questions from multiple parties, each with its own process. Shepherd’s early promise was a digital experience and a faster answer. Its more consequential bet was to connect underwriting with the technology contractors already used.
Shepherd Savings is the clearest expression of the thesis. The program evaluates qualifying technology adoption and usage and can offer upfront premium savings. It does not begin by asking a contractor to wear a new device or change a crew’s habits for an insurer. It begins with tools the contractor already chose. That removes a familiar obstacle in prevention programs: the party paying for better behavior is no longer waiting indefinitely for somebody else to notice.
In a 2024 study, Shepherd analyzed project-management behavior and loss records across more than 1,000 contractors. It reported correlations between Procore use and claim patterns, then incorporated its findings into underwriting. The company has been careful to distinguish correlation from causation, including analysis around adoption dates. The important operating choice is the feedback loop: collect the evidence, test it, price what can be supported, and keep learning.
Give the underwriter the wheel, lose the keystrokes
By 2026, Shepherd’s language had shifted from digitization to autonomous underwriting. The distinction is important. Digitization puts a form on a screen. Automation can read the incoming emails, PDFs, and spreadsheets, structure the information, check appetite, and prepare an account for judgment. Levine compares the roadmap to levels of autonomy in self-driving cars. The human role changes as the system takes on more of the route.
His preferred future for underwriters is not disappearance. It is elevation. He argues that they should become strategists and portfolio orchestrators, spending less time on data entry and more on selection, pricing, relationships, and the shape of a book. Shepherd says its underwriters already handle two to three times more accounts than peers in comparable commercial risk. It does not employ underwriting assistants. AI use is part of hiring and part of the job.
“Underwriters are going to become strategists and orchestrators and people who think about portfolios, not keystroking and data entry.”Justin Levine, 2026
This is a practical philosophy of automation: remove the work that merely moves information, then make the consequential work more visible. It also demands humility. Commercial insurance is full of long-tail consequences. A fast wrong answer can take years to reveal itself. Shepherd’s technology therefore depends on experienced underwriters as much as it depends on models. The unusual culture Levine has tried to build is a conversation between the two.
The AI boom needs power, walls, and a policy
In March 2026, Shepherd announced a $42 million Series B led by Intact Private Capital, with participation from Spark Capital, Costanoa Ventures, and others. The company said the round brought total funding to $67 million. It also reported sevenfold revenue growth over two years, more than 1,500 policies, over 600 customers, and more than $400 billion in insured project value.
The timing placed Shepherd beside a new kind of construction rush. AI may be sold as weightless intelligence, but its physical layer is emphatically heavy. Data centers require land, concrete, cooling, specialized equipment, and dedicated power. They also require casualty and property coverage assembled at a speed conventional commercial underwriting was not designed to match. Levine’s neat phrase is that the AI race moved from the cloud to the construction site.
Yet data centers are only a minority of Shepherd’s reported insured portfolio. The larger book includes commercial buildings, schools, life-science facilities, roads, bridges, tunnels, and other infrastructure. Construction was chosen as an entry point because every physical asset must first be built. In January 2026, Shepherd extended that logic into renewable energy and power, following solar, wind, and battery assets from development through construction and operation.
Lo-fi beats and people who argue well
Levine talks about the team with more warmth than he talks about capital. Years ago, he called the people the best thing Shepherd was building. The company mixes specialties that are not naturally fluent in one another: software engineering, data science, claims, actuarial work, and commercial underwriting. His hiring idea is to find “micro-entrepreneurs,” people with enough command of a discipline to question how it is practiced.
He also has ordinary founder habits. A lo-fi beats playlist runs in the office. He tries to keep one no-meeting day each week for uninterrupted work. After long stretches of remote work, he found himself wanting more face-to-face collaboration and the useful information carried in a reaction or a pause. Shepherd’s San Francisco office is part of that operating choice, especially for product and engineering teams.
The through-line is not disruption for its own theatrical sake. It is proximity. Levine learned insurance beside the people buying it. He learned software by turning their recurring task into a product. He learned the limitations of construction data by helping put that product inside a much larger platform. Shepherd is the next pass over the drawing: a company built on the belief that better knowledge of daily work can lead to a more accurate price for risk.
His old measure of success remains the more interesting one. He said Shepherd would have succeeded if incumbent insurers eventually looked more like Shepherd than Shepherd looked like them. That leaves plenty of room for the old industry to answer back. It also reveals what Levine wants to build: not merely a faster counter at the same insurance office, but a different view of the jobsite outside its window.