The Pen, the Map, and the Model: 250 Years of P&C Underwriting
In 1771, seventy nine merchants, brokers, and underwriters each put one hundred pounds into a common fund and formalized what had been, until then, a crowd of freelancers doing business over coffee. That subscription created the Society of Lloyd's, and it is as good a birthdate as any for underwriting as an organized profession. Roughly 250 years later, the desk those men would recognize, a human reading documents about a risk and deciding whether the price clears the exposure, still exists. Almost everything around it has been torn down and rebuilt several times.
This article traces those rebuilds. Not as trivia, but because the history has a structure to it, and the structure is useful. Underwriting has been reinvented roughly every forty years, and nearly every reinvention followed the same sequence: a loss event breaks the prevailing method, a new data layer gets built in the wreckage, and pricing sophistication follows the data with a lag. The Great Fires gave us collective loss data and schedule rating. San Francisco gave us the claims paying reputation as a competitive asset. Hurricane Andrew gave us catastrophe models. Each time, people announced the death of the underwriter. Each time, the underwriter survived by moving up the stack. If you want to understand what AI is about to do to this profession, the honest answer is that the profession has been here before, five or six times, and the pattern is legible.
The Coffee House and the Signature
The word itself is the origin story. In the decades after Edward Lloyd opened his coffee house on Tower Street in London in the 1680s, merchants seeking to insure a ship and its cargo would circulate a slip describing the vessel, the voyage, the cargo, and the master. Men with capital would write their names under the description, each taking a share of the risk for a share of the premium. They wrote under. They were underwriters. The entire modern apparatus of risk selection descends from that physical act: a named individual, personally liable, reading a description of a risk and deciding whether to sign.
What made Lloyd's coffee house the venue rather than any other was information. Lloyd built his business on shipping intelligence, arrivals, departures, casualties, and by 1734 that intelligence had been formalized into Lloyd's List, one of the oldest continuously published journals on earth. This is the first lesson of the 250 years, and it shows up in every subsequent era: underwriting advantage has never really been about capital, which is a commodity, or courage, which is cheap. It has been about proprietary information reaching the person with the pen before it reaches the competition. The coffee house was a data platform. The underwriters were pricing the spread between what they knew and what the market knew.
Fire insurance grew up alongside marine, and it grew out of catastrophe. The Great Fire of London in 1666 destroyed some 13,000 houses and demonstrated that urban fire was a modelable peril: buildings were the exposure, construction was the rating variable, and proximity was the accumulation problem. Nicholas Barbon, a physician turned property developer, opened his Fire Office in 1681 and priced brick houses differently from timber ones. That distinction, charging more for frame than for masonry, is the first rating factor in property insurance and it survives, essentially intact, in every homeowners and commercial property manual in use today.
The American Experiment: Franklin, Surveys, and Saying No
The American industry begins, as many American things do, with Benjamin Franklin. The Philadelphia Contributionship for the Insurance of Houses from Loss by Fire, founded in 1752 and still operating, imported the mutual model and added something that deserves more credit than it gets: the survey. Before the Contributionship would insure a house, an inspector examined it and recorded its construction, its chimneys, its hazards. Risks that failed inspection were declined. The Contributionship famously refused to insure houses with trees in front of them, on the theory that trees interfered with firefighting, a decision that eventually spawned a competitor, the Mutual Assurance Company, which accepted tree lined properties and took the green tree as its fire mark.
It is worth pausing on how much modern practice is already present in 1752. Physical inspection before binding. Documented risk characteristics. Explicit declination criteria. Adverse selection playing out in real time, as the risks one carrier refused pooled at another carrier that priced them knowingly. The tools were a quill and a ledger, but the logic was loss control underwriting, and it worked well enough that the institution outlived the country's first two central banks. The Insurance Company of North America followed in 1792, the first American stock insurer, and brought marine underwriting discipline to a young economy that ran on shipping.
Reputation as an Underwriting Asset: New York, 1835
Through the early nineteenth century, American fire insurance was local, thinly capitalized, and untested at scale. The test arrived in December 1835, when a fire tore through lower Manhattan and destroyed several hundred buildings in the heart of the young nation's commercial district. The losses bankrupted the large majority of New York's fire insurers almost overnight. Policyholders discovered the difference between buying a promise and buying a funded promise.
Into that wreckage stepped Eliphalet Terry, president of the Hartford Fire Insurance Company, who according to the company's own lore traveled by sleigh from Hartford to New York, pledged his personal fortune behind the company's obligations, and announced that Hartford would pay its claims in full and would gladly write new business. Whatever embellishment the story has picked up over nearly two centuries, the commercial effect was real and durable: Hartford emerged from the fire with a national reputation, and the industry learned that a catastrophe is also a marketing event. Claims paying behavior in the worst week of the market's life is the most expensive and most effective advertising an insurer will ever buy. Underwriters have been managing to that truth ever since, because the ability to pay in the tail event is a direct function of the discipline applied on every ordinary day before it.
The Great Fires and the First Data Layer
The second half of the nineteenth century is when underwriting stopped being a purely individual craft and became an industry with shared infrastructure. The National Board of Fire Underwriters formed in 1866 to bring order to rates and to push, decade after decade, for building codes, water systems, and professional fire departments. A year later, in 1867, a surveyor named D. A. Sanborn began publishing the fire insurance maps that carry his name: block by block diagrams of American cities recording every building's footprint, construction, height, use, and proximity to its neighbors.
The Sanborn maps deserve recognition as the first structured dataset in American underwriting. For nearly a century, an underwriter pricing a warehouse in a city he had never visited could open a folio and see the risk and, critically, the accumulation around it. Congested frame districts, the exposure of one building to its neighbor, the location of water mains. This was exposure management on paper, and it directly prefigures the exposure databases that catastrophe modelers would build 120 years later. The pattern to notice: the data layer was built first, as shared infrastructure, and pricing sophistication followed it.
It followed because it had to. The Great Chicago Fire of 1871 burned more than three square miles of the city and generated insured losses that overwhelmed the market; dozens of insurers failed outright and many more paid cents on the dollar. Boston burned the following year. The industry's response over the next generation was to make fire rating systematic. The Universal Mercantile Schedule in the 1890s, and A. F. Dean's analytic system published in 1902, decomposed a fire rate into a base rate for a standard building in a standard city, with published debits and credits for every deficiency and improvement: construction, occupancy, protection, exposure. An underwriter applying Dean's schedule was executing an algorithm, decades before anyone would have used the word. Schedule rating was the first attempt to make underwriting reproducible, auditable, and independent of the individual judgment of whoever held the pen. It was also the first time underwriters complained that the new system reduced them to clerks, a complaint that will recur in this story roughly once per generation.
San Francisco, 1906: The Cable That Built a Century of Trust
The earthquake and fire that destroyed San Francisco in April 1906 produced claims of roughly 235 million dollars, an almost incomprehensible sum against the era's capital bases, and it stress tested every clause in every policy. Many policies excluded earthquake; the fire that followed the earthquake burned for days; the proximate cause litigation could have consumed a decade. Some companies paid in full. Others discounted claims or hid behind exclusions, and their names were remembered on the West Coast for fifty years.
The most famous response came from Cuthbert Heath, the Lloyd's underwriter who had pioneered much of Lloyd's expansion beyond marine, and who cabled his San Francisco agent an instruction that became the market's founding legend in America: pay all of our policyholders in full, irrespective of the terms of their policies. As underwriting decisions go, it was technically indefensible and commercially brilliant. Lloyd's paid out enormous sums and purchased, with them, the American market's trust for the next century. The 1906 lesson compounds the 1835 lesson: the underwriting function does not end at binding. How an institution behaves at claim time is priced into everything it writes afterward, and the underwriters who understood that were managing a balance sheet asset that never appeared on the balance sheet.
New Perils, New Science: The Casualty Revolution
Property insurance had two and a half centuries of loss experience by 1900. The new century immediately produced two perils with none at all. When Gilbert Loomis bought what is generally credited as the first American automobile liability policy from Travelers in 1897, one thousand dollars of coverage for the damage his horseless carriage might do, the company priced it using horse drawn vehicle rates, because that was the only loss experience on earth that seemed adjacent. It is the purest example in the industry's history of underwriting a risk with no data: you reason by analogy, you price with a margin for ignorance, and you write small until the losses teach you.
Workers compensation arrived by statute rather than by market demand. Wisconsin's 1911 law was the first to survive constitutional challenge, and within a decade nearly every industrial state had followed. Compensation was unlike anything the fire underwriters knew: high frequency, statutorily defined benefits, classification by occupation, and an explicit public policy mandate. It demanded a new science, and it got one. The Casualty Actuarial Society was founded in 1914 largely to put workers compensation ratemaking on a statistical footing, and out of that project came experience rating, the idea that an individual insured's own loss history should modify its manual rate. Experience rating is such a fixture now that it is easy to miss what it represented: the first formal, credibility weighted blend of individual data and class data in pricing. Every predictive model in use today is, philosophically, a descendant of that blend.
The Regulatory Architecture: Paul, SEUA, and McCarran
No account of American underwriting is honest without the legal architecture, because for most of the twentieth century regulation shaped the desk as much as data did. In Paul v. Virginia in 1869, the Supreme Court held that insurance was not interstate commerce, which left regulation to the states and, conveniently, left the industry's cooperative ratemaking outside the reach of federal antitrust law. For seventy five years, fire rates in most states were made collectively through rating bureaus, and the underwriter's job was in large part the correct application of the bureau tariff.
The Supreme Court reversed itself in United States v. South-Eastern Underwriters Association in 1944, holding that insurance was interstate commerce after all and that the Sherman Act applied. The industry's response was swift and effective: the McCarran-Ferguson Act of 1945 handed regulation back to the states and preserved a limited antitrust exemption for cooperative activities regulated by state law. The consequence was the prior approval era, in which rates were filed, justified, and approved, and rating bureaus remained central. Only gradually, through open competition laws in the second half of the century, did pricing become a genuinely competitive activity. The Insurance Services Office, formed in 1971 from the consolidation of the old bureaus, evolved into an advisory organization publishing loss costs rather than final rates, and carriers took over the last mile of pricing themselves. That last mile is where the modern pricing arms race would be run.
The Cash Flow Hangover and the Long Tail's Revenge
The 1970s and 1980s taught the industry two expensive lessons about discipline. The first was cash flow underwriting: with double digit interest rates on offer, carriers in the late 1970s and early 1980s wrote liability business at prices everyone knew were inadequate, on the theory that investment income on the float would cover the gap. When interest rates fell and losses developed, the result was the mid 1980s liability crisis, a hard market so severe that commercial insureds could not buy coverage at any price, municipalities went bare, and the corporate risk management and captive movement gained permanent momentum. The lesson, relearned in every soft market since, is that underwriting profit and investment profit are not fungible, because only one of them is under the underwriter's control.
The second lesson was the long tail itself. Asbestos and environmental liability claims, brought under occurrence policies written decades earlier, produced losses that no one pricing those policies in the 1950s and 1960s had imagined, on legal theories that did not exist when the contracts were signed. At Lloyd's, the compounding of long tail American liability through the LMX reinsurance spiral nearly destroyed the institution; the Reconstruction and Renewal settlement of 1996 ring fenced the old years into Equitas and ended the era of unlimited personal liability for Names. For underwriters, the asbestos experience permanently changed the meaning of the job. A casualty underwriting decision is not a one year bet. It is a decades long option written on the future of tort law, science, and society, and the premium charged today has to carry uncertainty that cannot be modeled, only respected.
The Quant Turn: Segmentation Becomes the Game
While commercial lines were absorbing those lessons, personal lines were quietly having a statistical revolution. Beginning in the United Kingdom in the 1980s and spreading through the 1990s, actuaries replaced one way rating analyses with generalized linear models, which could price dozens of variables simultaneously and honestly. In the United States, credit based insurance scores, developed with Fair Isaac in the early 1990s, turned out to be among the most predictive rating variables ever introduced in personal auto and home, and among the most socially contested, a controversy that has never fully closed and prefigures every current argument about algorithmic fairness in AI. California's Proposition 103 in 1988 stands as the era's reminder that rating is not a purely technical activity; it is conducted under a social license, and variables the public will not accept do not survive, whatever their lift charts say.
The competitive dynamics of segmentation deserve emphasis because they are the cleanest demonstration in the industry's history of what happens to slow adopters. In a segmented market, the carrier with the finer rating plan does not merely win business; it wins the good business and sends the bad business, priced as average, to the carrier with the coarser plan. Adverse selection is not a static cost. It is a ratchet that transfers the profitable tail of the book from the unsophisticated to the sophisticated, quarter after quarter, and it is invisible in the aggregate numbers until the combined ratio has already gone wrong. Progressive built one of the great growth stories in the industry's history substantially on this mechanism. The strategic point transfers directly to the present moment: in any pricing technology transition, the cost of being late is not foregone efficiency. It is a book that quietly curdles.
Andrew, and the Forty Eight Hours That Changed Property Underwriting
Catastrophe modeling existed before Hurricane Andrew. Karen Clark founded Applied Insurance Research in 1987 and introduced a probabilistic hurricane model years before the storm; Risk Management Solutions emerged from Stanford in 1989. The industry mostly declined to believe them. The prevailing method for estimating hurricane exposure was a rule of thumb scaled off premium volume, informed by a couple of decades of quiet Atlantic activity, and as Swiss Re's head of catastrophe perils later put it, the industry was caught napping.
Andrew made landfall south of Miami in August 1992 and generated roughly 15 billion dollars in insured losses, multiples of what the premium based formulas had projected. According to the Insurance Information Institute, at least eight Florida insurers were rendered insolvent, with some accounts counting the failures higher, and others survived only because parent companies transferred capital in. The models, which had projected losses of Andrew's magnitude and been dismissed for it, were vindicated in the most expensive way possible.
What followed was the fastest methodological revolution in the industry's history. Within three years, catastrophe models moved from curiosity to requirement. The Bermuda reinsurance class of 1993 was built explicitly on modeled underwriting, eight new reinsurers deploying fresh capital through exceedance probability curves rather than intuition. Exposure management, knowing not just what you insured but where, to what limits, in what accumulations, became a core underwriting discipline rather than a back office afterthought, and terms like probable maximum loss and average annual loss entered the working vocabulary of every property underwriter. Florida turned itself into a regulatory laboratory, down to a state commission that certifies which models may be used in rate filings. September 11th, 2001 extended the same arc to terrorism, a peril the industry priced at approximately zero on September 10th, and the Terrorism Risk Insurance Act of 2002 formalized the recognition that some accumulations require a public backstop. Katrina in 2005 taught the models humility about storm surge, levee failure, and demand surge, and the models absorbed the lesson and improved, which is the point: the industry had finally built a data and modeling layer that could learn from events rather than merely be destroyed by them.
Telematics, Triage, and the Data Rich Desk
The first two decades of this century extended the quant turn from pricing into the underwriting workflow itself. Progressive's telematics program, piloted as TripSense in 2004 and scaled as Snapshot, moved personal auto rating from proxies about the driver to observations of the driving, the most direct data an insurer had ever collected on its own insured. Aerial and satellite imagery did something similar for property, replacing the drive by inspection with continuously refreshed roof condition and exposure data. In commercial lines, predictive models moved into triage and appetite: scoring submissions, flagging the risks that deserved senior attention, and straight through processing the small commercial business whose premium could never carry the cost of a human touch.
The InsurTech wave that crested after 2015 promised to finish the job and mostly did not, for a reason worth stating precisely because it sets up the present moment. Much of that generation attacked distribution and customer experience, the parts of the value chain visible from outside the industry, while the actual constraint on underwriting productivity sat where it had always sat: in the documents. Submissions arriving as unstructured email. Loss runs in forty carrier formats. Statements of values missing half their schedules. ACORD forms filled out approximately. The estimate most commercial underwriters would recognize is that the majority of desk time went to assembling and rekeying information before a single judgment was applied to it. The industry had built extraordinary models and was feeding them by hand.
The Present Turn: Language Models Meet the Submission Inbox
Which brings the story to the current chapter, and to why this one is genuinely different in kind rather than merely in degree. Every prior data revolution in this history structured the world in advance: Sanborn drew the maps, the bureaus standardized the stat plans, the modelers built the exposure databases. The revolution now underway is the first whose technology can work with the world unstructured. Large language models read the loss run, the SOV, the broker email, and the inspection report in their native disorder, and that capability lands precisely on the constraint that has governed the underwriting desk for a century.
The economic stakes are being sized in earnest. McKinsey estimates generative AI could unlock 50 to 70 billion dollars in insurance industry revenue, and describes the industry ascending a staircase from predictive analytics through generative document work toward agentic systems that coordinate entire workflows. The same firm's researchers have been blunt that the value will not come from tinkering, from running a few pilots, or from buying a patchwork of software from suppliers with minimal strategic intent. The early adoption data supports the warning: Evident's AI Use Case Tracker recorded 87 percent year over year growth in insurance AI use cases, while only around 40 percent of insurers report tangible business outcomes. A great deal of motion; considerably less traction. The gap between those two numbers is where this era's version of an old story is playing out, the difference between technology bolted onto a workflow by builders who met the business at the point of sale, and technology built into the workflow by people who understood loss runs and treaty renewals before they wrote a line of code.
History suggests how this resolves, because it has resolved the same way five times. Schedule rating did not eliminate the fire underwriter; it eliminated the recalculation of debits and credits and moved the underwriter's judgment to the risks the schedule could not see. Catastrophe models did not eliminate the property underwriter; they eliminated the premium based guesswork and moved the judgment to model validation, accumulation strategy, and the perils outside the model. The pattern is consistent: automation consumes the reproducible center of the work and pushes human judgment to the exceptions, the tail, and the portfolio. The underwriters of the next decade will supervise machine assembled risk pictures, spend their attention on the submissions that do not behave, and manage books rather than files. That is not the death of the craft. It is the fifth or sixth time the craft has shed its clerical layer and kept its judgment layer, and the judgment layer has always been where the combined ratio was actually made.
What 250 Years Actually Teaches
Compress two and a half centuries and four patterns survive the compression.
First, catastrophe is the industry's R&D department. Nearly every methodological leap in this story was purchased with a loss that broke the old method: 1666 and the birth of fire insurance, 1835 and the funded promise, 1871 and schedule rating, 1906 and the claims reputation, 1992 and the catastrophe model. The institutions that compounded through those moments were not the ones that avoided the loss. They were the ones prepared to be solvent, credible, and open for business the week after it.
Second, data layers precede pricing revolutions, in that order, every time. Lloyd's List before marine underwriting matured. Sanborn maps before schedule rating. Standardized stat plans before actuarial ratemaking. Exposure databases before catastrophe pricing. Digitized submissions before whatever underwriting becomes next. Whoever builds and owns the data layer sets the terms of the era that follows, which is why the current race to structure the industry's unstructured documents matters more than any individual model or vendor.
Third, the social license question never closes. Fire marks and the uninsured, redlining and its long shadow, credit scores, price optimization, and now algorithmic underwriting: every generation's most predictive new variable collides with the public's sense of fairness, and the variables that lose that collision disappear regardless of their statistical merit. Underwriting is conducted inside a political settlement, and the underwriters who forget it get their rating plans written by referendum, as California demonstrated in 1988.
Fourth, and most practically: in every transition, the losers were not the institutions that adopted the new method imperfectly. They were the institutions that adopted it late, because adverse selection is a ratchet, and by the time it shows up in the results, the good business has already left.
The Takeaway
The underwriter of 1771 signed a slip because he had better information than the market and the nerve to act on it. Everything since, the schedules, the actuaries, the models, the machines now reading the submission inbox, has been the industrialization of that single act. The tools have never been the constant. The constant has been that someone, ultimately, decides whether the price clears the risk, and that the quality of that decision compounds silently for years before the market grades it. The next decade will hand that decision more assistance than the previous 250 years combined. The institutions that win it will be the ones that treat the new tools the way Lloyd's treated its shipping lists and the Bermuda class of 1993 treated its models: not as a threat to judgment, and not as a replacement for it, but as the data layer underneath it. That has been the winning position for 250 years. There is no evidence the pattern is about to break.
Fabio Faschi is an Enterprise Insurance and AI leader, National Producer, Board Member of the Young Risk Professionals New York City chapter and Committee Chair at RISE with over a decade of experience in the insurance industry. He has built and scaled over a dozen national brokerages and SaaS-driven insurance platforms, and is the founder of ScholarusAI.com and Hogglet.com for Enterprise AI transformation and risk management. Fabio's expertise has been featured in publications like Forbes, Consumer Affairs, Realtor.com, Apartment Therapy, SFGATE, Bankrate and Lifehacker.
