AI use cases in commercial underwriting split into two groups. Risk-level use cases score, price, and route one submission at a time: which risk to quote, what to charge for it, which queue it should sit in. Portfolio-level use cases steer the book as a whole: where exposure is concentrating, which segments are drifting out of appetite, where terms are quietly loosening across a whole class of business. Most underwriting AI investment today sits on the risk-level side.

That split isn't just a framing choice. In a 2025 survey of roughly 80 to 95 North American carriers, Celent found that AI activity clusters into four stages: experimental pilots, operational AI embedded in defined workflows, scaled intelligence that redesigns how humans and AI interact, and an adaptive enterprise stage with agentic AI managing micro-processes in real time. Very few carriers have reached either of the last two. The same research plots AI use cases on a grid by impact and differentiation, and most carriers are investing heavily in the high-impact, low-differentiation quadrant: document ingestion, summarization, straight-through processing for homogeneous risk. The harder, rarer work sits in the high-differentiation quadrant: embedding appetite upstream, continuous underwriting, portfolio optimization. That's the same split this article uses below.

Within underwriting specifically, Celent found 42% of carriers already using generative AI there, with the clearest current activity in document analysis and summarization (29%), submission ingestion and validation (28%), and case analysis and summarization (22%). The six use cases below show where that activity concentrates today, and where it's headed next.

Risk-level: scoring, routing, and preparing one submission at a time

1. Submission ingestion turns documents into usable data

Ingestion extracts and structures the data buried in broker submissions, loss runs, and ACORD forms, and it's become table stakes faster than almost anything else on this list. One hx customer alone automatically ingested 250,000 data points into the platform in a single 30-day period, data that would otherwise have been manually keyed. Roughly a quarter of hx's own customers are now using its AI tooling across pricing and underwriting workflows.

It's valuable precisely because it's no longer differentiating. It saves underwriters time they can spend on judgment instead of data entry, and it gets a quote back to a broker faster, but carriers who have it are table stakes carriers, not ahead of the market. The parts of commercial and specialty underwriting that haven't adopted it yet are the exception now, not the rule.

2. Underwriting agents surface the signal in external and third-party data

https://share.descript.com/view/a3BLEhoER8w

A D&O or E&O underwriter needs to know whether adverse media exists on a risk before pricing it. Classically that's a manual search, one underwriter working through results one at a time. An underwriting agent runs that search directly and surfaces what's relevant to the pricing and risk decision.

The same problem shows up with third-party data providers more broadly. A NatCat or financials data provider might return 100, 200, or 300 data elements for a single risk. An underwriter can't act on all of them, so historically that list gets manually pruned down to the four or five fields someone decided matter. The underwriting agent approach inverts that: it reads the full set of data and surfaces what's actually pertinent to this risk, rather than relying on a fixed, pre-pruned subset that may miss something relevant.

3. Appetite scoring and triage route submissions by value

Carriers increasingly score inbound submissions against appetite before an underwriter ever opens them, so the business reaching an underwriter's desk is more likely to be business the carrier actually wants and can competitively quote. That's a direct lever on quote-to-bind ratio and on how underwriting capacity gets spent, since capacity is finite and not every submission deserves equal attention.

4. Straight-through processing shifts from assisting to initiating

In the current generation of AI tooling, a human still initiates basic submission, triage, pricing, and issuing, and AI assists by summarizing, drafting, or recommending. The direction of travel is a shift in decision authority: AI initiates the triage, assignment, and documentation request, and escalates exceptions to a human rather than the other way round. That shift is furthest along in homogeneous, high-volume lines like small commercial, where the underwriting judgment required per risk is lowest and the case for full automation is strongest.

Portfolio-level: steering the book as a whole

5. Policy-wording comparison flags term drift at the account and portfolio level

https://share.descript.com/view/uWp8pmYlLHM

Manuscripted policy wording changes from one renewal to the next, and in a softening market those changes tend to move in one direction: terms get more generous even as rates hold. Rate deterioration alone doesn't show the full picture, because wording can loosen quietly underneath it.

Comparing policy wording year over year surfaces exactly what changed in the terms and conditions for an underwriter reviewing a single account. The same comparison rolled up across a book gives a portfolio-level view of whether wording is drifting across a whole class of business, catching deterioration that would otherwise stay invisible until it shows up in loss ratio.

6. Continuous monitoring gives real-time portfolio visibility and steering

Rather than reviewing exposure on a fixed schedule, continuous underwriting models watch accounts on an ongoing basis and alert on change: a shift in operations, a location move, a hazard profile drifting out of what was originally priced. The same approach supports loss-control targeting, deciding which accounts warrant an inspection versus a quarterly check-in versus an annual one, and whether an account is likely to actually follow through on loss-control recommendations if given them.

This is the harder half of the split above. Risk-level tools tell you whether one submission is good. Only a connected portfolio view tells you whether it's good for the book you're actually writing, and whether the book as a whole is moving in the direction you intended.

Where carriers are already running this in production

This isn't limited to one vendor's roadmap. Celent's research cites public examples across the industry: State Farm using digital knowledge assistants for employees and customer service chat, QBE running a cyber-underwriting AI assistant, AIG using large language models across its underlying processes, Travelers in new-business bond submission, and Chubb across underwriting, claims, and marketing. The common thread isn't the specific use case, it's that the technology has moved from pilot to production across multiple carriers at once, not just one early mover.

Where to start

Carriers don't need to adopt every use case above at once, and most of the gap between carriers isn't technical. Start with a clear, high-value use case, not a shiny-object pilot. Get executive sponsorship with real accountability attached to it, and actually measure the return, since a lot of carriers implement something and never check whether it delivered one.

Most carriers have already built or bought something for ingestion, appetite scoring, and triage. Fewer have a live, connected view of how those decisions add up across the book, policy wording drift, continuous exposure monitoring, portfolio optimization. That's usually where the bigger gap still sits.

How hx connects individual decisions to the book

Submission triage, pricing and rating, and portfolio intelligence run on hx's shared decision layer, so a risk-level decision and a portfolio-level view draw on the same data rather than living in separate tools an underwriter has to open side by side.

Actuarial teams build pricing and scoring logic in native Python and move approved changes through governed deployment workflows with fewer routine IT handoffs. API ingestion puts relevant data at the point of pricing, cutting re-keying and shortening quote turnaround. Every action taken in the platform is recorded and feeds back into portfolio analysis and what-if testing, helping teams validate models, adjust appetite faster, and keep loss ratios in check.

Book a demo to explore how hyperexponential connects risk-level scoring and pricing to governed portfolio steering.