Insurance doesn't need more AI features. It needs to be AI-native.

For much of the history of enterprise software, progress has been measured by a simple standard: the amount of work people no longer had to perform manually. Spreadsheets displaced hours of calculation, email transformed how information moved across organizations, and successive enterprise systems digitized forms, approvals, and records once kept on paper. Each advance made organizations faster and easier to scale, yet the intellectual structure of the work itself remained largely intact. Software changed how work was executed; it rarely questioned whether the work should be organized differently in the first place.

‍

Insurance has followed much the same trajectory. Over two decades, carriers have invested billions in document management, workflow engines, OCR, robotic process automation, and, more recently, generative AI. Yet the day-to-day experience of an underwriter remains curiously familiar. Submissions still arrive from every conceivable source, information still has to be gathered across dozens of documents, inconsistencies reconciled, and fragmented pieces assembled into a coherent understanding of the risk. The surrounding technology has grown far more sophisticated; the essential nature of the work has changed far less than we assume.

‍

I am occasionally reminded of the Ship of Theseus. If every plank of a ship is replaced over time until none of the original material remains, is it still the same ship? Insurance technology poses a similar puzzle. Manual keying has given way to extraction engines, filing cabinets to digital repositories, lengthy document reviews to AI-generated summaries. One component after another has been modernized, each an improvement in isolation, yet the broader process remains strikingly recognizable. We have become proficient at replacing individual planks while seldom asking whether the ship itself still represents the best possible design.

‍

That question is difficult to dismiss because underwriting has never truly been an exercise in reading documents. Documents are merely the medium through which information reaches the underwriter. The real discipline begins once the reading ends. Experienced underwriters reconcile statements across applications and loss runs, recognize contradictions between inspection reports and schedules of values, identify omissions that alter the interpretation of a submission, and determine whether the documents present a credible picture of the underlying risk. Underwriting has always been an exercise in judgment, with document processing serving merely as the mechanism through which that judgment is informed.

‍

This is not only a philosophical point; it has become an operational reality. A large share of the industry's most experienced underwriters is approaching retirement, and too few people are entering the discipline to replace those leaving it, a gap expected to persist into the coming decade. How experienced underwriters spend their hours therefore matters. When a professional who has bound hundreds of millions in premium is occupied confirming that a company name was entered correctly, the industry is underusing the expertise it can least afford to waste. The scarce resource in underwriting has never been documents or data. It has always been judgment, and judgment is exactly what repetitive preparation consumes.

‍

This distinction changes how one should think about artificial intelligence in underwriting. Much of the industry's attention has gravitated towards individual capabilities: extracting fields more accurately, summarizing submissions, flagging missing information. Each is genuine progress, but they share an assumption, that underwriting remains a sequence of human activities to which AI contributes incrementally. The technology becomes another instrument within the workflow, while the responsibility for assembling, validating, reasoning, and producing the underwriting-ready output continues to rest with the underwriter.

‍

Perhaps that is inevitable. We first ask how a new technology can improve existing work before asking whether it changes the definition of the work itself. The printing press was initially seen as a faster way to copy manuscripts; it took decades to appreciate that it had altered the economics of knowledge. Artificial intelligence may be nearing a similar inflection point in insurance. The more interesting question is no longer how many tasks it can accelerate, but whether those tasks should continue to exist as discrete human activities at all.

‍

Seen through that lens, underwriting looks different. The objective is no longer to help an underwriter navigate documents more efficiently; it is to present a submission that has already been interpreted, reconciled, and organized into a coherent representation of the risk. The value shifts from individual features towards the completeness of the outcome. The industry has long measured its teams by output while judging them by outcomes. If the outcome is the true measure, the underwriter's starting point ought to be refined information rather than raw material.

‍

But this reality is soon changing with the advancement of AI-Native Services (AINS), something that we are championing at Pibit. Internally, we are measuring every new update as a step towards building an AI-native insurance services ecosystem. It reflects a different way of designing insurance software: instead of asking how AI might improve individual stages of a workflow, asking whether the workflow should be partitioned into those stages at all. The wider technology industry is moving the same way. Y Combinator has made AI-native services a priority for its 2026 cohorts, backing companies that sell the finished result rather than a tool the customer must still operate. The phrase that captures the moment is that AI has stopped being a feature and become the foundation. It is hard to think of an industry to which that applies more naturally than insurance.

‍

The difference may seem semantic, but its consequences are profound. An organization built around AI-assisted workflows scales much as it always has, with technology making individuals more productive. One built around AI-native outcomes scales differently, because the unit of work changes. Human expertise stops being consumed by repetitive preparation and concentrates where it has always mattered most: interpreting uncertainty, exercising judgment, making decisions no dataset can resolve. In the figures carriers watch, gross written premium per underwriter, time to quote, and the share of quoted business that binds all begin to move, and in a portfolio of any size, modest gains translate into revenue measured in the millions. It matters all the more where hiring is not a simple remedy, since the people a carrier would need to add are already in short supply.

‍

The future of underwriting is unlikely to be shaped by whichever platform accumulates the most AI features. Those capabilities will keep improving, just as OCR and workflow engines did before them. The more consequential shift lies in reimagining software around complete underwriting outcomes rather than individual underwriting tasks. That is the philosophy behind AI-Native Services, and it is this shift, from feature-led software to outcome-led systems, that will define the next generation of underwriting technology. For the underwriter, it means opening a file that is already understood, and spending the day on the judgment that drew them to the work in the first place.