Chapter 36: The Future of Underwriting: AI, Climate Change, and the Underwriter of 2035

"The best way to predict the future is to invent it." — Alan Kay, computer scientist, in a 1971 talk at Xerox PARC. For an underwriter the line cuts two ways. The future of your profession is not a forecast you wait to receive; it is a thing the industry — and you — are building right now, one model, one override, one repriced book at a time. But Kay's optimism has a hard edge in insurance: the one future you cannot invent away is the physical one. The climate is doing what the physics says it will do, on its own schedule, and no amount of clever product design repeals it. The underwriter of 2035 lives where those two truths meet.

Overview

Imagine the desk you will sit at a decade from now. The submission does not arrive as a PDF you read once and price once; it arrives as a live feed. A sensor on the sprinkler riser reports water pressure every minute. A satellite re-images the roof every few weeks and flags the day a seam starts to fail. A telematics box in every truck streams hard-braking events to a dashboard that re-scores the fleet nightly. The model that scored the account at bind keeps scoring it, every day, against data that did not exist when you wrote it. And the same account sits on a coast where the hundred-year flood is arriving every fifteen, where the reinsurance behind your policy costs more each renewal, and where two of the largest carriers in the country have stopped writing new business entirely. That is not science fiction. Every piece of it exists in pilot form today. The question this chapter answers is what the underwriter does in that world — and the answer is the conviction this whole book has been building toward: the routine gets automated, the judgment does not, and the climate makes the judgment harder and more valuable than it has ever been.

Two forces are remaking the profession at once, and they pull in different directions. The first is artificial intelligence, which is moving from a tool that scores a static submission (Chapter 32) to a system that monitors a risk continuously and drafts the underwriter's work — a co-pilot that never sleeps. The second is climate change, which is not a tool at all but a physical fact that is repricing whole property lines faster than the rate-filing process can keep up, pushing some risks out of the realm of the insurable, and forcing a public-private reckoning about who bears the catastrophe the private market will no longer take. AI makes underwriting faster, cheaper, and more consistent. Climate makes it harder, more uncertain, and more consequential. The underwriter who can hold both — who can wield the AI and still exercise the judgment that the climate-stressed, novel, contested risk demands — is the professional this chapter argues will be the most valuable in insurance in 2035.

We will look forward without pretending to certainty, because honest forecasting names what it cannot know.

In this chapter, you will learn to:

• Define continuous underwriting and explain how real-time monitoring changes the underwriting decision from a one-time gate into an ongoing relationship.
• Define AI-augmented underwriting and locate the line between what the co-pilot does and what the underwriter must still own.
• Explain how climate change is repricing property insurance, and why a moving baseline is harder than a one-time shock.
• Define insurability and its limits, and judge when a risk crosses from priceable to uninsurable.
• Evaluate parametric, embedded, and on-demand products for what they solve and what they leave uncovered.
• Name the skills and the career opportunity that the transformation creates.

Learning Paths

This is a forward-looking chapter, so every track should read it as a map of where its own corner of the profession is heading — but each will weight it differently.

🏠 Personal Lines: The climate sections (§36.3, §36.4) hit personal property hardest — homeowners on the coast and in the wildland-urban interface are where insurability is failing first, and the public-private response (FAIR plans, the NFIP — Chapter 15) is where you will spend your career. Watch the on-demand and embedded products in §36.5; personal lines is where they are spreading fastest. 🏢 Commercial Lines: Continuous and IoT-based underwriting (§36.1) will reach commercial property and fleet first, because the sensors pay for themselves there. Harbor Steel is the worked example. The insurability-under-stress discussion (§36.4) is the one that decides whether you can keep writing whole industries and regions. 📊 Analytics: §36.1 and §36.2 are your frontier — the move from batch scoring (Chapter 32) to streaming, the rise of large language models as drafting co-pilots, and the governance problem that comes with both. Note carefully where the chapter says the model still cannot go. 📜 Certification: This chapter is lighter on testable mechanics and heavier on professional judgment, but the AINS/CPCU material on emerging risk, the limits of insurability, and the underwriter's evolving role lives here. The career section (§36.7) maps directly to Chapter 37.

36.1 Continuous underwriting: risk assessed in real time

Begin, as always, with the decision in front of you, because the future changes when you make it more than what it is. For the whole history of the craft, underwriting has been a snapshot. A submission arrives, you assess the risk as it is on that day, you price it, you bind it, and then — for a year, until renewal — you look away. The risk lives its life unobserved. The truck whose brakes you inspected at bind wears them out in month three; you do not know. The roof you rated as sound springs a leak in month seven; you find out only when the claim arrives. The whole structure of the annual policy is built around a single act of assessment frozen at a point in time, with a true-up twelve months later. That snapshot model was not a choice; it was a limit. You could not watch a thousand roofs every day, so you looked once and trusted.

The technology of Chapter 31 dissolves that limit. The sensors, the satellites, the telematics, the internet-connected equipment that data-driven underwriting put on the desk do not stop reporting after the bind. They keep streaming. And once the data is continuous, the underwriting can be too. This is continuous underwriting: the practice of monitoring an insured risk in real time, with live data, throughout the policy period — so that the risk picture is updated continuously rather than captured once at the point of sale and revisited only at renewal. The assessment stops being a photograph and becomes a video.

📋 At the Desk Continuous underwriting is easiest to understand as the collapse of the gap between underwriting and loss control. In the snapshot world those are two departments: you (underwriting) decide whether to write the risk, and someone else (loss control, Chapter 9) visits once to recommend improvements. In the continuous world they fuse. The sensor that tells you the sprinkler pressure dropped is doing underwriting (the risk just got worse) and loss control (someone should fix it today) in the same signal. The most valuable thing continuous underwriting does is not re-price the policy mid-term — most policies legally can't be re-priced mid-term anyway. It is to prevent the loss by catching the deterioration while it is still cheap to fix. The best claim, as Chapter 1 said, is the one that never happens; continuous monitoring is the machinery for making more of them never happen.

Notice what continuous underwriting does and does not change. It does not repeal the policy contract. A property policy bound for a year is a year-long promise (Chapter 5); you generally cannot raise the rate in month four because a sensor flagged a problem, any more than you could before (rate regulation, Chapter 4, governs that). What the live data changes is everything around the contract: whether you offer a loss-control intervention now instead of at the next inspection; whether you fund or require a repair as a condition of renewal; what you know when you set the next term's price; whether you non-renew a risk that is deteriorating instead of discovering the deterioration in the claim file. Continuous underwriting is less a new way to price a policy than a new way to manage a relationship with a risk across its whole life.

THE SNAPSHOT MODEL vs. CONTINUOUS UNDERWRITING            [constructed teaching example]

  SNAPSHOT (today's annual policy)
    bind ──────────── (12 months, unobserved) ──────────── renewal
     │                                                        │
   assess once                                          assess again
   price once                                           re-price
                  ↑ the loss happens HERE, unseen, in month 7

  CONTINUOUS (the 2035 desk)
    bind ─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─●─ renewal
     │     ↑ sensor flags pressure drop, month 7          │
   assess  → intervene NOW (loss control), log for renewal pricing,
            require fix as a renewal condition — the loss is PREVENTED

The diagram is schematic, but it captures the shift: the value moves from the two endpoints (bind, renewal) to the continuous middle, where the loss used to hide. This is the most concrete sense in which "the future of underwriting" is not a slogan. For sensor-rich commercial risks — property with monitored fire protection, fleets with telematics, equipment with breakdown sensors — the continuous model is already arriving. It will reach the rest of the book's lines unevenly, fastest where the data is cheap and the exposure is expensive, slowest where a sensor cannot see the risk at all (a products-liability exposure, a professional-liability exposure — there is no sensor for a negligent weld or a bad legal opinion).

⚠️ Underwriting Trap The seductive error of continuous underwriting is to mistake more data for more certainty, and to let the live feed lull you into writing risks you would otherwise decline — "we'll just monitor it." Two problems. First, monitoring is not mitigation: a sensor that tells you the building is burning does not put the fire out, and many of the worst losses (a catastrophe, a sudden structural failure, a fraud) give you no useful warning window even with perfect instrumentation. Second, the data stops the moment the insured unplugs the box — and the insured most likely to unplug it is exactly the one whose risk just got worse (adverse selection, Chapter 1, wearing a new costume). Continuous underwriting sharpens a good decision; it does not rescue a bad one. Write risks you would write anyway, and let the monitoring make them better — never write risks you wouldn't, on the promise that you'll watch them fail in real time.

There is a compliance dimension here that will only grow. Continuous data raises continuous questions about consent, privacy, and use. The insured agreed to share telematics or sensor data — but for what, exactly? Can you use a hard-braking pattern to non-renew? Can you share the feed with a third party? The privacy and data-ethics line that Chapter 35 draws around algorithmic decision-making applies with full force to a stream that never stops. The underwriter of 2035 will need to know not just what the data says but what the law and the contract permit them to do with it — and that boundary is still being drawn.

36.2 AI as co-pilot, not replacement

Now the force everyone asks about first: will artificial intelligence take the underwriter's job? You have the tools to answer it precisely, because the book has been building the answer for thirty-five chapters. The honest reply is a division of labor, not a yes or a no — and the cleanest name for that division is the one the industry has settled on: AI is a co-pilot, not an autopilot.

Let us define the term the careful way. AI-augmented underwriting is the practice in which artificial intelligence — predictive models (Chapter 32), and increasingly large language models that can read, summarize, and draft in natural language — handles the mechanical and high-volume work of underwriting (gathering and summarizing the submission, scoring the routine risk, drafting the quote, flagging the exception) while the human underwriter retains the judgment, the override, and the accountability for the decision. The "augmented" is load-bearing: the model amplifies the underwriter rather than replacing them, the way a power tool amplifies a carpenter. The carpenter still decides where the cut goes.

To see why the division falls where it does, separate the underwriting task into two kinds of work, exactly as the book has all along. There is the high-volume, well-understood, data-rich work — the small commercial BOP (Chapter 20), the standard personal auto (Chapter 14), the simplified-issue life policy (Chapter 17). For that work, the machine is already better than a person: faster, cheaper, more consistent, tireless, and — critically — auditable in a way a tired human at 4 p.m. is not. Straight-through processing (Chapter 20) and the predictive models (Chapter 32) already write enormous volumes of this business with no human in the loop, and AI will write more of it. Defending that is not nostalgia; it is arithmetic. No human desk can price ten million auto policies a year one at a time, and no insurer should want it to.

Then there is the complex, novel, contested, relationship-dependent work — the Harbor Steel package with its catastrophe exposure and its loss-run story (the whole book); the first policy for a risk that barely existed a decade ago (cyber, Chapter 24; a new climate peril; an AI-liability exposure that does not yet have a form); the account where the model says decline and the file says write (Chapter 32); the broker relationship that brings you the best risks first (Chapter 39). For that work the machine is a powerful assistant and a poor decider, for reasons this book has named repeatedly: it cannot read context it was never given, it cannot price a risk with no history, it cannot be held accountable to a regulator or an underwriting committee, and it optimizes the objective it was handed rather than the one you actually have.

🤖 Model vs. Judgment The large language model is a genuinely new kind of co-pilot, and it is worth being precise about what it changes and what it doesn't. What it changes: the drafting and reading work that used to eat an underwriter's day. An LLM can read a hundred-page submission and a five-year loss run and produce a two-page summary, a draft risk assessment, and a first-cut list of questions for the broker in seconds. It can draft the quote letter, the subjectivity list, the decline letter (Chapter 13). Used well, it gives the underwriter back the hours that judgment actually needs. What it does not change: the LLM does not know anything — it predicts plausible text, and plausible text can be confidently wrong (the industry word is hallucination). It will state a coverage position that sounds right and is not, invent a regulatory rule that does not exist, summarize a loss run and quietly drop the one claim that mattered. So the rule is the same rule as for every model in this book, sharpened: the co-pilot drafts, the underwriter verifies and owns. You sign the quote, not the model. "The AI wrote it" is not a defense you can give to a court, a regulator, or your manager — and in a few years it may be malpractice to have offered it.

The phrase to retire is "AI will replace underwriters." The phrase to keep is "underwriters who use AI will replace underwriters who don't." The mechanical underwriter — the one whose value was looking up the rate, re-keying the application, and stamping the file — is being automated away, and should be; that was never the craft. The judgment underwriter — the one whose value is reading the risk, structuring the deal, overriding the model with documented reasons, and defending the decision — is being amplified, handed an assistant that does the drudgery so the judgment has room to work. Which underwriter you become is, as the part introduction said, entirely your choice, and this book exists to make the second one possible.

⚖️ Compliance Corner AI-augmented underwriting collides head-on with the fairness law of Chapter 35, and the collision is getting regulatory attention right now. A model — and an LLM is a model — can encode bias it was never told to (proxy discrimination and algorithmic bias, both owned by Chapter 35), and a system that drafts and explains its own decisions can produce explanations that sound principled while the real driver is a protected-class proxy. Regulators have noticed: the NAIC has issued model guidance on the use of artificial intelligence by insurers, and states such as Colorado (SB21-169) have moved to require insurers to test their algorithms and data for unfairly discriminatory outcomes. The direction of travel is clear: as AI does more of the decision, the burden to prove the decision is fair — to test it, document it, and explain it — falls more heavily on the insurer, and therefore on the underwriter who stands behind it. The future is not less regulated. It is differently, and in places more heavily, regulated, precisely because the machine can hide unfairness inside math.

36.3 Climate change reshapes every property line

Set the AI down for a moment, because the second force needs no algorithm to be the most consequential thing happening to insurance. The physical climate is changing, and insurance — the industry whose entire business is pricing the probability of physical loss — is on the front line of feeling it. Hurricanes draw energy from warmer oceans. Wildfire seasons run longer and burn hotter in drier, hotter landscapes. Convective storms, hail, and flood are arriving in patterns that the historical record no longer reliably predicts. For the property underwriter, this is not a distant ethical concern; it is a repricing of the core product, already underway.

Here is the technical heart of why climate is so hard for insurance, and it reaches all the way back to Chapter 1. The law of large numbers and the whole apparatus of catastrophe modeling (Chapter 30) rest on an assumption that the past is a usable guide to the future — that the frequency and severity of perils, estimated from history, will hold well enough to price tomorrow. Catastrophe models (Chapter 30) already strain that assumption because catastrophes are rare; climate change breaks it, because it makes the baseline itself non-stationary — it moves. The "1-in-100-year" flood (return period, Chapter 30) defined on twentieth-century data may be a 1-in-30-year flood now, and a 1-in-15-year flood by 2050. When the baseline moves, the historical loss record is not just noisy; it is biased downward, systematically understating tomorrow's risk. An underwriter pricing coastal property off twenty years of loss history is, without a correction, pricing yesterday's climate into tomorrow's policy.

⚠️ Underwriting Trap The deadliest climate trap is the one that looks like good underwriting: trusting a clean loss run on a property whose risk is rising. A coastal account with no losses in the last five years looks like a good risk by every traditional measure (the loss run is the history that predicts the future, Chapter 8). But if the five quiet years were luck, and the underlying frequency of the destructive storm is climbing, the clean record is a trap — it tells you about a baseline that no longer exists. This is the climate version of the catastrophe error from Chapter 1: pricing for the average year when the tail is getting fatter. The disciplined move is to price off the forward-looking view — the catastrophe model's climate-conditioned output (Chapter 30), not the rear-view loss run — and to treat a clean coastal loss run as necessary but nowhere near sufficient.

Climate reshapes every property line, but unevenly, and naming the pattern is the underwriter's job:

• Hurricane / named windstorm (the Harbor Steel peril; homeowners, Chapter 15; commercial property, Chapter 19). Warmer water means, by the physics, more energy available for the strongest storms and more rainfall in them. Storm surge rides on a rising sea. The result is rising severity in the most exposed coastal zones — exactly where the protection gap (Chapter 30) is already widest.
• Wildfire (homeowners and commercial property in the wildland-urban interface). Hotter, drier, longer fire seasons in fire-prone landscapes have turned wildfire from a regional nuisance into a catastrophe-scale peril, and it has driven the most visible insurability crisis in the United States.
• Flood (the chronic under-insured peril; NFIP, Chapter 15). More intense rainfall and rising seas push flood beyond the historical floodplain maps that pricing and the public program were built on. Flood is where the gap between economic loss and insured loss — the protection gap — is largest of all.
• Severe convective storm (hail, tornado, straight-line wind — the quiet giant). Less dramatic than a hurricane, but a growing and geographically broad driver of property losses, and one the models have historically handled less well than the headline perils.

📋 At the Desk When climate enters your pricing, the single most useful discipline is to separate the trend from the event. Any one hurricane, fire, or flood is weather — an event, drawn from a distribution. The thing that should move your rate is the trend: the slow shift in the distribution itself. Underwriters get this backwards in both directions. After a big loss year, the temptation is to over-react to the event and spike the rate (and then cut it again after two quiet years — the underwriting cycle, Chapter 3, with a weather amplifier). In a quiet stretch, the temptation is to forget the trend entirely. The professional move is to hold the rate to the forward-looking expected loss — the climate-adjusted average annual loss (AAL, Chapter 30) — through both the loud years and the quiet ones, which is just rate adequacy (Chapter 11) applied to a moving target. It is the hardest version of the book's fourth theme, and the one the next decade will test most.

36.4 Insurability under stress — and the public-private response

Climate pushes us to the edge of the most important concept in this chapter, the one the whole book has circled since Chapter 1: insurability itself. We can now define it directly. Insurability is the condition of a risk being capable of being insured — possessing, well enough, the characteristics that make the insurance mechanism work (a large pool of similar exposures, definite and fortuitous loss, a calculable chance of loss, losses not catastrophic to the insurer, and an economically feasible premium — the criteria of Chapter 1) — at a price someone will both charge and pay. The last clause is the one that matters here. A risk does not become uninsurable because the math fails in the abstract. It becomes uninsurable when the price the math demands is higher than the price the market will bear — when the adequate rate (Chapter 11) is so high that customers won't (or by regulation can't) pay it, or insurers won't write it at the rate they're allowed to charge.

This is the crisis climate is creating, and it is essential to see that it is usually an affordability and availability crisis before it is a physics crisis. The risk can still be modeled; the AAL can still be computed; the catastrophe XOL (Chapter 27) can still, in principle, be priced. What fails is the meeting of the adequate price and the payable price. When a coastal or wildfire-exposed property's risk-adequate premium climbs past what homeowners can afford — or past what the state's rate regulation will approve — insurers face a choice with only bad options: write it at an inadequate rate and bleed (violating the third and fourth themes, the combined ratio and rate adequacy), or stop writing it. Increasingly, large carriers are choosing the second. Insurers have publicly pulled back from writing new homeowners business in the most exposed parts of states such as California and Florida, citing catastrophe exposure, reinsurance cost, and the difficulty of charging an adequate rate under the applicable regulation. That is not a hypothetical future; it is a present fact, and it is the sharpest test of insurability the industry has faced in a generation.

📄 Read the Submission

text FIGURE 36.1 — "The risk that is modeled but not writable" [constructed teaching example] THE SUBMISSION A homeowners renewal on the coast: a well-built home, fully to code, no losses in the insured's ten years of ownership. The owner simply wants to keep the coverage they have. THE CONTEXT The catastrophe model's climate-conditioned named-storm AAL for the location has roughly doubled over a decade; reinsurance for the zone has hardened; the state's rate regulation has approved only part of the indicated increase. The risk-adequate premium is now far above both what the owner can comfortably pay and what the filed rate allows. WHAT IT SHOWS The risk is perfectly MODELABLE — we can price it. It is the meeting of the adequate price and the payable/approved price that has failed. This is an availability and affordability failure, not a measurement failure. WHAT IT DOESN'T It does not mean the homeowner is a "bad risk" or did anything wrong. A good, careful, loss-free insured can become uninsurable purely because the priced risk of their LOCATION outran the price the system will bear. THE DECISION At the company level: non-renew or restrict new business in the zone unless the rate can reach adequacy or the cat exposure can be ceded/shared. At the SYSTEM level: this is what FAIR plans, residual markets, and public backstops exist to catch (below). THE LESSON Insurability is not a fixed property of a risk; it is a moving relationship between the adequate price, the payable price, and the regulated price. Climate is moving all three.

So what catches the risks the private market will no longer take? The answer is the public-private response, and the underwriter needs to understand it because it is increasingly part of the landscape they operate in:

• Residual markets and FAIR plans. Most states maintain an insurer of last resort — a FAIR (Fair Access to Insurance Requirements) plan or a state-created entity — to provide basic property coverage to those the standard market won't write. These plans are filling up in stressed states, which concentrates catastrophe risk in exactly the entities least equipped to diversify it — a structural worry, not a solution.
• Public catastrophe programs. The National Flood Insurance Program (NFIP, Chapter 15) is the largest example: a federal program created precisely because private flood insurance was largely unavailable. It demonstrates both the promise (coverage exists) and the peril (a program that underprices the risk for affordability accumulates losses the taxpayer ultimately bears) of the public backstop.
• Mitigation and resilience. The most durable answer is to change the risk, not just who pays for it: building codes, wildfire defensible space, flood barriers, hardened roofs. Insurance can price and reward mitigation (the credit/debit logic of Chapter 11), and increasingly it must, because mitigation is the only lever that lowers the adequate price rather than just hiding it.
• Managed retreat and the hard limit. At the far edge is the possibility that some places become genuinely uninsurable at any payable price — that the honest underwriting answer is "this should not be rebuilt here." That is a societal decision, not an underwriting one, but underwriting is the messenger: the price is the signal, and a risk-based price that no one will pay is the market telling the truth about the physical risk.

⚖️ Compliance Corner Here the book's sixth theme — insurance serves a social function — reaches its hardest and most political edge, and the underwriter's duty is to be honest about the tension rather than to resolve it glibly (a discipline Chapter 35 owns in full). On one side is actuarial fairness (Chapter 35): the risk-based price is the true price, and suppressing it through regulation or subsidy distorts the signal, encourages people to build in harm's way, and accumulates losses someone eventually pays. On the other side is social fairness: a risk-based price that makes whole communities uninsurable, often communities with the least ability to move or to absorb the loss, is a real harm that "the model said so" does not answer. The climate-insurability crisis is where these two fairnesses collide most violently, and there is no clean underwriting answer — only the obligation to price honestly, to support mitigation that actually lowers risk, and to be clear-eyed that the protection gap (Chapter 30) is a societal problem the insurance mechanism alone cannot close.

36.5 The new product frontier: parametric, embedded, and on-demand

If continuous data and AI change how underwriting is done, and climate changes what must be priced, the third frontier changes the product itself. The traditional indemnity policy — pay the actual loss, proven and adjusted, up to a limit — is not the only shape insurance can take, and the next decade will see the others spread. Three are worth knowing, each a different answer to a weakness in the indemnity model. Each was introduced earlier in the book; here we ask what they become at scale, and — in the book's honest habit — what each leaves uncovered.

Parametric insurance (owned by Chapter 26) pays a fixed amount when a measured trigger is met — a hurricane of category X passing within Y miles, an earthquake above a magnitude, a rainfall total, a wind speed — rather than paying the proven loss. Its great virtues are speed and certainty: there is no claim to adjust, so the money arrives in days, and both sides know the terms in advance. For climate perils this is powerful — a coastal business can get cash to keep the lights on the week after the storm, not the quarter after — and it pairs naturally with the IoT sensors of §36.1, which can verify a trigger automatically. What parametric cannot do is match the payout to the actual loss. That mismatch has a name: basis risk — the gap between the parametric payout and the policyholder's true loss. The storm passes Y-plus-one miles away and does enormous damage: the trigger doesn't fire, and the insured collects nothing. The trigger fires but the insured happened to suffer little: they collect a windfall. Parametric trades indemnity's slow precision for speed and simplicity, and the price of that trade is basis risk. It is a supplement to indemnity coverage, not usually a replacement for it.

Embedded insurance (owned by Chapter 34) is coverage sold inside another transaction at the moment of need — the rental-car damage waiver at checkout, the shipping protection in the online cart, the appliance warranty at purchase, the travel coverage with the airline ticket. Its power is distribution: it reaches the customer at the exact instant the risk becomes salient, at near-zero acquisition cost, often via an API (Chapter 34) that quotes and binds in milliseconds. For the underwriter, embedded insurance pushes the selection decision almost entirely into the model and the product design — there is no human in the loop at the point of sale, which makes the up-front class definition, the rating algorithm, and the guardrails everything (the straight-through-processing logic of Chapter 20, industrialized). Its risk is the mirror of its power: coverage sold frictionlessly is coverage bought without much thought, which raises real questions of suitability and value that regulators are beginning to ask.

On-demand insurance is coverage you switch on and off for the period you need it — insure the drone for the afternoon you fly it, the car for the hours you drive it, the contents for the trip. It is the natural child of continuous data (§36.1): if you can monitor usage in real time, you can price it in real time. Its appeal is obvious and its adverse-selection problem (Chapter 1) is equally obvious: people will tend to switch the coverage on precisely when they perceive the risk to be highest, which is the oldest enemy in the book in a new app. On-demand products survive only with pricing and design disciplined enough to defeat that selection — which is, once again, underwriting judgment, embedded now in the product's rules.

🤖 Model vs. Judgment These products do not eliminate underwriting; they relocate it. In a traditional policy the judgment happens at the desk, account by account. In a parametric, embedded, or on-demand product the judgment happens up front, once, in the design: choosing the trigger and accepting the basis risk; defining the class that the embedded algorithm will write blind; building the anti-selection guardrails into the on-demand rules. The decision is made by a human — but earlier, at the level of the product rather than the policy, and then executed ten million times by the machine. This is the deepest sense in which "the underwriter of 2035" is not replaced but moved upstream: the most valuable judgment increasingly shapes the algorithm and the product before a single risk is written, rather than adjudicating risks one at a time. The skill is the same — read the risk, price it, structure it, defend it — exercised at a new altitude.

36.6 The skills that will matter in 2035

Pull the threads together and ask the question every reader of this book is really asking: given all this, what should I be good at? The transformation does not make underwriting skill obsolete; it raises the floor on the routine and raises the ceiling on the judgment, which means it changes the mix of what an underwriter is paid for. Here is the honest map of the skills that will matter, and it is striking how many of them are the oldest skills in the book, now more valuable rather than less.

• Judgment under uncertainty — the irreplaceable core. The whole book's first theme. When the model can price the routine, the human's value concentrates in the cases the model can't: the novel risk, the contested override, the climate-stressed account with no usable history. The underwriter of 2035 is paid precisely for the judgment that does not automate. This skill goes up in value, not down.
• Model literacy — reading, questioning, and overriding the machine. You will not build the models (Chapter 32), but you must read them well enough to know what they can and cannot see, to spot when a score is driven by a proxy (Chapter 35), and to override with documented reasons (the discipline of Chapter 32) and defend the override to a committee. "Fluent in the model without being captured by it" is the defining technical skill of the next decade.
• Climate and catastrophe fluency. Every property underwriter will need to think in forward-looking, climate-conditioned terms (§36.3, Chapter 30) — to price the trend, not just the event, and to reason about insurability under stress (§36.4) rather than treating it as fixed.
• Data judgment. Not data science — data judgment: knowing which of the flood of available data (Chapter 31) is signal and which is noise, which is permitted to use (Chapter 35) and which is not, and when a clean number is hiding a dirty assumption.
• Communication, negotiation, and trust. The relationship skills (Chapter 39) the machine cannot perform: explaining a price to an angry broker, saying no without burning the relationship, building the trust that brings the best submissions to your desk first. As the routine automates, the human-to-human work becomes a larger share of the job, not a smaller one.
• Ethical reasoning. The ability to see where risk-based pricing shades into unfair discrimination (Chapter 35), to weigh actuarial against social fairness honestly, and to refuse the decision that is profitable and wrong. In an age when the machine can hide bias inside math, the human who can find it is worth more, not less.

📋 At the Desk If you want a single organizing idea for your own development, here it is: let the machine do what it does better than you, and become excellent at what it cannot do at all. Do not compete with the model on speed, consistency, or volume — you will lose, and you should. Compete on judgment, context, relationship, and accountability — the things that are still, and will remain, human. The trainee who spends the next five years becoming a faster manual rater is preparing for a job that is disappearing. The trainee who spends them becoming a sharper judge of risk — fluent in the tools, honest about their limits, trusted by brokers, and willing to own a hard decision — is preparing for the job that the transformation makes more valuable than it has ever been. That second path is the one this whole book has been training you for.

36.7 The career opportunity in the transformation

End where careers begin, because this is, finally, a hopeful chapter. It is easy to read "AI is automating underwriting" as a warning and feel that you have arrived at a shrinking profession. The opposite is true, and the distinction matters: tasks are being automated, but the judgment the profession exists to supply is becoming scarcer relative to demand, not more abundant. The risks are getting more complex, not less — cyber, climate, AI itself as a new liability exposure, supply-chain and systemic risks that didn't exist a generation ago. Every one of those needs an underwriter who can do what no model can: price a risk with thin history, structure a deal that makes a hard risk writable, and stand behind a judgment when the algorithm and the broker and the committee all disagree.

The career opportunity, concretely, runs along the seams this chapter has traced. There is opportunity at the model seam — the underwriters who can sit between the data scientists and the business, translate between them, and govern the models will be disproportionately valuable (the analytic path of Chapter 37). There is opportunity at the climate seam — catastrophe and climate-risk underwriting is one of the fastest-growing and least-automatable specialties in insurance, precisely because the baseline is moving and the judgment is hard. There is opportunity at the new-product seam — the parametric, embedded, and on-demand frontier (§36.5) needs people who can design the judgment into the product. And there is opportunity, as always, at the relationship seam — the broker-underwriter partnership (Chapter 39) that no API replaces.

🤖 Model vs. Judgment The most common career mistake the transformation invites is to pick a side: to become either a "people person" who fears the math or a "quant" who disdains the relationship and the context. The future punishes both. It rewards the rare professional who is bilingual — fluent enough in the models to use, question, and override them, and fluent enough in the human craft to read a risk, build a relationship, and own a decision. That bilingual underwriter — comfortable with the algorithm and unintimidated by the judgment — is the single most valuable professional this book can help you become, and the whole arc from Chapter 1 to here has been the curriculum for becoming them. The machine handles the routine; you handle what matters; and the combination is worth more than either alone. That is not a threat. It is the best job in insurance, and it is the one waiting for you.

🗂️ The Underwriting File

Harbor Steel in 2035 — a forward look. Step out of the present file for a chapter and imagine the Harbor Steel account written a decade from now, with the tools and pressures of this chapter in force. Nothing here changes the decision the book is building toward — that is the capstone's to state (Chapter 40) — but it is worth seeing how the same risk would be underwritten in the world you are about to enter.

Continuous underwriting (§36.1). The plant's sprinkler riser, its electrical panels (the source of the 2021 fire), and its hot-work areas (the source of the 2023 fire) would all be sensored. Instead of an infrared scan ordered once as a subjectivity and a hot-work program audited at renewal, the panel temperature and the welding-permit compliance would stream to a dashboard. The two fires that drove this whole file might never have happened — or would have been caught at the smoldering stage. The telematics you'd require on the 12-truck fleet (Chapter 23) would re-score the drivers nightly rather than at the next MVR pull. The underwriting would stop being an annual snapshot and become a live relationship with the risk.

AI as co-pilot (§36.2). The submission would arrive pre-filled (Chapter 31) and pre-scored (Chapter 32); an LLM would draft the risk assessment and the subjectivity list from the loss runs in seconds. But the judgment that defines this account — reading the two fires as a story about management being fixed, weighing the prior carrier's non-renewal, and ultimately overriding the model's decline-leaning score the way Chapter 32 did — would still be yours to make and yours to defend. The co-pilot does the drafting; you still write the override.

Climate repricing (§36.3, §36.4). This is the part that genuinely changes. The Port Hadley named-storm AAL (Chapter 30) would, on a climate-conditioned 2035 view, very plausibly be higher than the figure the file carries today — and the reinsurance behind the property line (the cat XOL of Chapter 27) would cost more. The forward-looking adequate price for the catastrophe exposure rises, and the affordability and insurability questions of §36.4 press harder on the whole coastal book Harbor Steel sits in (Chapter 29). The risk does not become un-writable — Harbor Steel is a commercial account with mitigation levers and a broker delivering controls — but the margin between the adequate price and the payable price narrows, and the account's fate ties ever more tightly to whether the cat aggregate in the Port Hadley zone (Chapter 30) still has room.

A new product on the table (§36.5). By 2035 the parametric wind supplement that Chapter 34 merely floated would be a live option — a trigger-based cover paying cash on a defined named-storm event to bridge the gap above the 5% named-windstorm deductible, accepted with its basis risk understood. What this chapter adds to the file: a forward look, not a decision. The disposition is unchanged — the account remains the quote-with-conditions the building chapters set, and the capstone (Chapter 40) states and defends the binding decision. What 2035 changes is the toolset and the pressure: more data, a sharper co-pilot, a heavier climate load, a new product option — and, through all of it, the same human judgment at the center, now more valuable than ever.

Conclusion

The future of underwriting is not the disappearance of the underwriter; it is the concentration of the underwriter's value into exactly the work that cannot be automated. Two forces are remaking the profession at once. Artificial intelligence is moving from a tool that scores a static submission to a system that monitors a risk continuously and drafts the underwriter's work — continuous underwriting that turns the annual snapshot into a live relationship, and AI-augmented underwriting in which the machine is a co-pilot, never an autopilot, that does the drudgery so the judgment has room to work. Climate change, meanwhile, is repricing every property line by making the catastrophe baseline non-stationary — a moving target that the rear-view loss run systematically understates — and pushing whole regions and perils to the edge of insurability, where the adequate price outruns the payable one and the private market retreats, leaving a public-private response to catch what it cannot. The product frontier — parametric, embedded, on-demand — relocates underwriting judgment upstream into the design, rather than eliminating it. And the skills that will matter in 2035 turn out to be the oldest in the book — judgment, model literacy, climate fluency, communication, ethics — now worth more, not less.

What remains uncertain is genuine and worth naming: how fast the regulation of AI-driven decisions will tighten; how far climate will push insurability before mitigation, capital, and public backstops catch up; which of the new products will scale and which will stumble (the InsurTech lesson of Chapter 34 counsels humility); and, honestly, how much of this forecast the next decade will simply rewrite. Forecasting names what it cannot know. But the central claim is as solid as anything in the book, because it is the book's whole argument arriving at its destination: the underwriter who can bridge judgment and analytics — fluent in the model, honest about its limits, clear-eyed about the climate, trusted by brokers, and willing to own a hard decision — will be the most valuable professional in insurance. That is the future. You are, by reading this far, already building it.

The last four chapters turn from the future of the work to the shape of the career that does it — from your first day as a trainee, to leading the underwriting function, to the broker relationship that feeds you, to the capstone where you assemble the complete Harbor Steel file and bind the coverage you have spent forty chapters learning to defend. The file is nearly closed. Let's finish the job.

Key Terms

• Continuous underwriting — the practice of monitoring an insured risk in real time, with live data (sensors, satellites, telematics), throughout the policy period, so the risk picture is updated continuously rather than captured once at the point of sale and revisited only at renewal.
• AI-augmented underwriting — the practice in which artificial intelligence handles the high-volume, mechanical underwriting work (gathering, summarizing, scoring, drafting) while the human underwriter retains the judgment, the override, and the accountability for the decision; AI as co-pilot, not autopilot.
• Insurability — the condition of a risk being capable of being insured: possessing, well enough, the characteristics that make the insurance mechanism work (Chapter 1) at a price someone will both charge and pay; its limit is reached not when the math fails but when the adequate price outruns the payable or permitted one.

Spaced Review

1. Distinguish continuous underwriting from the traditional annual snapshot, and explain why its greatest value is usually loss prevention rather than mid-term re-pricing. (§36.1)
2. A model scores a complex account and an LLM drafts the quote letter. Where, exactly, does AI-augmented underwriting put the model's work and where does it put the underwriter's — and why is "the AI wrote it" not a defense? (§36.2)
3. (Back to Chapter 1 / Chapter 30.) Climate change makes the catastrophe baseline non-stationary. Using the law of large numbers and the idea of a return period, explain why a clean five-year coastal loss run can be a trap rather than a reassurance. (§36.3; Ch. 1, Ch. 30)
4. (Back to Chapter 32 / Chapter 35.) The Harbor Steel model recommended decline (a 7) and the underwriter wrote it at a 6. In a 2035 world of continuous data and a heavier climate load, what about that decision stays the same, and what gets harder? Name the fairness tension a heavier coastal repricing raises. (§36.2, §36.4; Ch. 32, Ch. 35)
5. (The recurring pricing-discipline question.) An underwriter prices a coastal book off the last five quiet years rather than the climate-conditioned average annual loss. Would that decision help or hurt the combined ratio, and over what time horizon would the truth show up? (§36.3; Ch. 3, Ch. 11)