How AI is used in commercial underwriting

AI in commercial underwriting mostly means two things: software that reads and organizes the submission so an underwriter can spend time on judgment, and statistical models that score risks and suggest a price or a yes/no. Few commercial carriers let a model bind a mid-size account with no human involved. Small, simple accounts are the exception, and that is where you will meet the most automation.

Why commercial underwriting is a good fit for automation

A commercial underwriting file is messy. An agent sends an ACORD application, a spreadsheet of locations, five years of loss runs in five formats, a prior policy, perhaps a safety manual. Before an underwriter can think about price, someone has to extract the data, check it, and enter it into a rating system. That clerical step eats a large part of an underwriter's day, which is why the first AI deployments went there.

The second reason is volume. Many carriers quote only a fraction of the submissions they receive. A tool that triages the rest in minutes and routes the good ones to the right desk changes how quickly a broker gets an answer.

Where AI shows up in the workflow

1. Submission intake and data extraction

Document-reading models pull exposure data (payroll by class code, revenue, square footage, vehicle schedules, statement of values lines) from PDFs, emails and spreadsheets, then populate the carrier's system. Some tools also compare the application with the loss runs and flag mismatches, such as a claim on the loss run that the application denies.

2. Appetite triage and prioritization

Models score each submission against the carrier's appetite: class of business, size, location, loss history, hazard. High-fit accounts go to the front of the queue; out-of-appetite ones get a fast decline or are referred elsewhere. This is the most common use, and the one that changes what a broker experiences.

3. Third-party data enrichment

The carrier checks what you told it against outside data: business registries, public records, property data, geocoding, hazard scores for flood, wildfire and hail, satellite or aerial imagery (see aerial imagery and drones in property underwriting), and web and review data that describe what a business actually does. Mismatches between the application and the data drive follow-up questions.

4. Risk scoring and pricing

Machine-learning models, often tree-based ensembles rather than anything exotic, predict claim frequency or severity from dozens or hundreds of variables. The output usually feeds a human or rules-based decision rather than replacing it. In rate-regulated lines the model has to fit inside the filed rating plan: the carrier can only move price within the filed rules and the schedule rating credits and debits it is allowed to use. See commercial rate filing: prior approval, file-and-use and deregulation.

5. Cyber and external-scan underwriting

Cyber is the line where outside-in data is most developed. Insurers scan a policyholder's internet-facing systems for open ports, unpatched software and exposed credentials, and some carriers built their business on that data. Coalition and At-Bay are cyber specialists that describe scanning as part of their underwriting and monitoring. A scan does not replace the application's answers on multi-factor authentication or backups; it checks them. See the cyber insurance application: what underwriters ask and how to answer.

6. Portfolio and accumulation monitoring

Models track concentration by geography, class and peril, so that a carrier does not discover after a hurricane that it wrote too much in one county. This sits close to catastrophe models: how insurers price hurricanes, quakes and fires and is less visible to the buyer.

Where it works well and where it does not

AI performs best on high-volume, homogeneous, small-premium business: the small contractor, the salon, the consultant. NEXT Insurance and other digital carriers quote this segment in minutes from a short questionnaire plus data lookups. See insurtech and small business insurance: what changed and what didn't.

It performs worst on complex, low-data accounts: a manufacturer with unusual processes, a large contractor with a shifting mix of jobs, a first-year company with no loss history. There, the underwriter's questions about the actual operation matter more than any score. What helps in that segment is still a clean submission, a credible account of your risk controls, and a broker who knows the market.

What it means for the submission you send

A hypothetical example: a 25-employee fabrication shop sends a submission with payroll split across two class codes, a loss run showing two property claims, and a one-line description of operations. An intake model extracts the payroll and compares the description with the class code. If the description reads like a different trade, the file gets flagged, and the underwriter's first email is a question, not a quote.

Practical steps:

1. Describe operations in specific, consistent language across the application, website and loss narrative. Models and underwriters both compare them.
2. Make exposure numbers match everywhere: payroll, revenue, vehicles, locations. A mismatch is the most common reason an automated check flags a file.
3. Send clean, current loss runs and explain any large claim in a sentence or two, with the corrective action. Our guide on how to prepare an insurance submission underwriters will read covers the details.
4. Expect more outside-in verification. If a carrier scans your cyber exposure or pulls aerial imagery, fix obvious issues before you apply: close exposed remote access, repair a visibly damaged roof.
5. Ask your broker what data sources the carrier used when a quote is surprising. Correctable errors in third-party data are real.

The limits and the rules

Models inherit the flaws of their data. A property score built on old imagery can penalize a roof you replaced; a business-type model can misread a company with a generic name. Insurers also face rules on how they use models. The NAIC's December 2023 Model Bulletin on the use of AI by insurers expects written governance, testing for unfair discrimination, and oversight of third-party vendors, and a majority of jurisdictions have now adopted it. Colorado's statute, SB21-169, targets life, private passenger auto and health benefit plans rather than commercial lines. The details are in AI regulation in insurance: the NAIC bulletin and state rules. Commercial underwriting is generally less tightly constrained than personal lines, but the bulletin as written reaches any insurer decision that relies on an AI system.

For you as a buyer, the relevant questions are practical. Can the carrier explain a decline or a surprising surcharge? Who fixes bad data? Does your broker see the same information the underwriter used? A carrier or MGA that cannot answer is telling you something about how much human review the file gets. Related background is in Generative AI in the insurance industry and Managing general agents (MGAs): what they are and what they mean for buyers.

Key takeaways

• AI in commercial underwriting mainly speeds intake, triage and data enrichment; humans still decide on most mid-size and complex accounts.
• It is most automated for small, simple classes and in cyber, where outside-in scanning is common.
• Consistent descriptions and matching exposure numbers across the application, loss runs and public data prevent most automated flags.
• Third-party data can be wrong; ask your broker how a surprising quote was built and correct bad inputs.
• The NAIC's 2023 AI bulletin and state rules set governance expectations for insurers using models; see the regulation guide for specifics.