A submitted document can pass every readability test and still fail the underwriting review.
The pages are there. The sections are present. On the surface, the language looks right.
But somewhere inside — in a liability clause that doesn't hold; in an abuse prevention policy that omits what the product requires; in a waiver that exists but won't survive a coverage dispute — the exposure the carrier thought it was managing, remains open.
This is not a document reading problem. It is a clause validation problem.
The insurance industry has invested heavily in AI for document processing, and the returns are beginning to show. WTW's 2026 Advanced Analytics and AI Survey found that P&C carriers using sophisticated analytics achieved combined ratios six points lower and premium growth three points higher than slower adopters over the past two years. But behind that aggregate performance gap is a more specific challenge that does not get named clearly enough: Most of the document AI deployed in underwriting today stops at extraction. It tells you what is in the document. It does not tell you whether what is there is sufficient.
That distinction — presence versus sufficiency — is where underwriting exposure quietly accumulates.
The misdiagnosis
The first instinct when applying AI to underwriting document review is to treat it as a reading problem. Get the right information out of the document reliably. Flag missing pages. Route submissions faster. These are tractable problems, and AI solves them well.
But extraction is not validation. Knowing that a hold harmless clause exists is different from knowing whether that clause actually transfers the relevant exposure. Knowing that an abuse prevention policy is present is different from knowing whether it meets the specific criteria the product requires — background screening by scope and frequency, training applied to the right roles, mandatory reporter obligations named explicitly where jurisdiction demands them.
McKinsey's research on large commercial lines puts the operational context plainly: underwriters spend 30 to 40 percent of their time on administrative tasks — rekeying data, checking completeness, manually executing analyses. That is not time spent on judgment. It is time spent on evidence gathering. When evidence gathering consumes that share of capacity, the deeper question of sufficiency gets less attention than it deserves.
The result is a predictable gap. Not random. Not occasional. The same clause types, across the same product lines, surface the same insufficiency patterns at claim time — because they were not evaluated at submission.
The problem is structural, not just operational
Clause validation failures are not primarily a technology problem or a staffing problem. They are a criteria problem.
Most document review processes — manual or AI-assisted — are built around presence detection. Does the document contain an indemnity clause? Does it name the required additional insured? Is a waiver present? These are binary questions, and they are the wrong questions for the lines of business where clause quality matters most.
In liability lines covering youth organizations, habitational risks, sports and recreation, and event venues, the distance between a clause that exists and a clause that performs is exactly where adverse claims originate. A general liability submission for a youth program may include a sexual abuse and molestation endorsement — and still fall short if the underlying abuse prevention policy omits the structural elements the product is written to require. A waiver may be present, correctly worded, and jurisdiction-compliant — and still be legally unenforceable because it was not executed against the current scope of operations.
Execution gaps deserve particular attention because they are both consequential and invisible to standard review. An unsigned waiver is not a waiver. An undated consent form may not be enforceable. A signature page that predates a material change in operations provides limited protection regardless of its language. These gaps do not announce themselves. They surface at claim time.
Jurisdictional variance adds another dimension. A document drafted for national operations typically reflects the obligations of a single state. What satisfies requirements in one jurisdiction may omit statutory language required in another. In multistate operations, reviewing for that variance consistently is impractical at manual review volume — and straightforward to build into a systematic process.
What AI can actually address — and what it cannot
The shift that matters is not from manual to automated review. It is from presence detection to sufficiency evaluation.
AI-assisted clause validation, built with the right criteria, changes what the underwriter is reviewing. Instead of reading through a document to locate relevant language, the underwriter receives a structured finding: the passage identified, the requirement evaluated against, and a determination of where it falls short. The judgment remains human. The evidence gathering — which McKinsey's data suggests consumes up to 40 percent of underwriting time in large commercial lines — becomes systematic.
That reallocation of attention is where the operational leverage lies. Celent's analysis from ITC 2025 observed this shift directly: the industry conversation around AI in underwriting has moved from automation to interpretation — from straight-through processing as the goal toward decision quality and transparency as the standard [2]. AI is functioning less as a faster underwriter and more as the infrastructure that directs underwriting attention to where it is genuinely needed.
What AI does not address is contextual sufficiency — the cases where a clause satisfies its stated requirement while leaving a gap the requirement did not anticipate. Indirect exposures embedded within permitted activities. Coverage treatment that is technically adequate for one territory but insufficient given the carrier's actual multi-state exposure. These require an experienced underwriter, and no validation framework substitutes for that expertise. The purpose of systematic clause review is to ensure that expertise is applied to cases that genuinely demand it, not distributed across completeness checks.
The real investment
Carriers exploring this space consistently encounter the same surprise: the technology is not the hard part.
The harder investment is in the criteria library — the explicit, product-level definitions of what each requirement means, what satisfies it, and what does not. That knowledge exists in underwriting. It lives in the judgment that experienced reviewers apply case by case. Getting it into a form that a systematic process can apply consistently is where most implementations either succeed or stall.
The second requirement is transparency. A system that produces a determination without showing its reasoning creates more burden, not less. Underwriters need to see what was found, where it was found, and why the determination was reached — not a confidence score. That traceability serves the review workflow, and it is what makes the output defensible when a submission goes to a coverage dispute.
WTW's 2026 survey data makes the business case visible. The performance gap between AI-advanced carriers and slower adopters — six points on combined ratio, three points on premium growth — did not emerge from technology investment alone [3]. It emerged from building the underlying criteria, process, and governance that make analytics actionable at the decision level.
The adoption trajectory in claims analytics tells a parallel story. Fraud detection analytics are used by one-third of carriers today and projected to reach nearly 70 percent within two years. Straight-through processing in claims workflow automation sits at 14 percent today, with 36 percent of carriers planning deployment shortly. The direction is not ambiguous.
The differentiator will not be which carriers deploy AI. It will be which carriers deploy it against a well-defined standard of what sufficient actually means — at the clause level, for each product, across every jurisdiction where they write business.
The clause validation gap is not a feature request. It is infrastructure — the layer between document intake and underwriting decision where the risk posture the carrier intends to write either holds or quietly does not. Building that layer well is not the most visible application of AI in insurance. It may be the most consequential.
Naveen Karakavalasa is Principal Manager of Emerging Technologies,T okio Marine North America Services. He specializes in the design and deployment of AI systems for underwriting and claims operations. He writes at nkspace.dev.
Any opinions expressed here are the author's own.
(Featured image credit: Piscine26/Adobe Stock)
Read additional thought leadership from contributor: How AI is changing document analysis in insurance
© Arc, All Rights Reserved. Request academic re-use from www.copyright.com. All other uses, submit a request to TMSalesOperations@arc-network.com. For more information visit Asset & Logo Licensing.