There is a number your operation almost certainly isn’t tracking: the number of submissions your underwriting team declines to quote just because there isn’t enough capacity to get to it.
A 2026 analysis by V7 Labs put a precise shape to the problem that most underwriting leaders already sense. A mid-sized carrier’s underwriting team might receive 30 to 50 commercial submissions per week. Each submission can run to 200 pages or more: loss runs, inspection reports, schedules of values, ACORD forms, prior policy wordings, broker narratives. Before an underwriter can form any view on the risk, someone has to read all of it, extract what matters, cross-reference what’s inconsistent, and flag what’s missing. That work, according to the same analysis, consumes more than 60 percent of a senior underwriter’s working week.
Which means the person you hired to assess and price risk is spending most of their time doing something that has nothing to do with assessing or pricing risk.
This article talks about how document AI agents remove the processing work from in front of underwriting judgment, which documents cost the most time, and how to build an underwriting workflow using Kudra.
What’s Actually Sitting Between Your Underwriters and the Risk
Walk through a single commercial submission and map where the time goes.
A broker sends a package. It contains a completed ACORD application, five years of loss runs from three prior carriers (each in a different format), an inspection report from 18 months ago, a schedule of values across 12 locations, and a broker narrative explaining why a large GL claim from 2023 should be viewed as an isolated incident.
Before an underwriter can evaluate that risk, someone has to:
- Identify every document in the package and confirm nothing is missing
- Normalize the loss runs, three different carriers, three different formats, some scanned, some Excel, one embedded in a PDF into a format that allows year-over-year comparison
- Extract incurred and paid figures, calculate frequency and severity trends, and flag any loss that exceeds an internal threshold
- Pull out the location data from the SOV, verify that all 12 locations are present and geocoded correctly
- Read the broker narrative and identify any statements that require follow-up
- Log everything into the underwriting system before work can begin
In most commercial lines shops, this takes 90 minutes to three hours per submission. For a team processing 40 submissions a week, that’s 60 to 120 hours of senior staff time spent on document reconstruction before a single underwriting decision is made.
That’s the capacity your operation is leaving on the table. Every week.
The Submission Backlog Is a Selection Problem in Disguise
when underwriting capacity is constrained by document processing time, the submissions that get worked are often not the submissions that should get priority.
Urgent broker requests jump the queue. Familiar relationships get faster attention. Simpler submissions get cleared first because they’re faster to process. The complex risks that require (and deserve) the most careful underwriting judgment end up waiting the longest, or getting declined not on merit but because the team simply ran out of week.
This matters for portfolio quality. The Guidewire London Market Tech Barometer, published in February 2026, surveyed more than 250 brokers placing risk in the London Market. 78 percent said an insurer’s technology capability either strongly influences or is decisive in placement decisions. More than half already use digital or algorithmic underwriting processes on the broker side. When your team’s turnaround time is three days and a competing carrier is at same-day for clean commercial risks, the selection dynamic shifts. You don’t just lose deal velocity, you lose the ability to shape your own book.
What a Document AI Agent Changes in Underwriting
The term gets overused, so it’s worth being precise about what a document AI agent does in a commercial underwriting context, and what it doesn’t do.
When a submission lands, a document AI agent reads the full package. It identifies each document type without being told (loss run, ACORD form, SOV, inspection report) and begins processing each one according to its structure. Loss runs get normalized regardless of carrier format, whether they arrive as Excel, scanned PDF, or an image embedded in an email. Incurred and paid figures are extracted, frequency and severity are calculated, and any claim that exceeds a configurable threshold is flagged automatically. The SOV gets mapped location by location. The ACORD data is extracted and cross-referenced against any prior policy on file.
What arrives at the underwriter’s desk is not the 200-page submission. It’s a structured briefing: coverage requested, key risk characteristics, loss history summary, any flags that need their attention, and a note on anything missing from the package.
The underwriter opens the briefing, reviews what the agent surfaced, applies their judgment, and moves to a decision. The 90 minutes of document reconstruction is already done. All that’s left is the part that actually requires them.
ScienceSoft’s 2026 analysis of AI in commercial underwriting found a 67 percent decrease in manual underwriting tasks and a reduction in average decision time from days to minutes for straightforward submissions. Hiscox, in a widely cited case, reduced underwriting turnaround from 72 hours to 180 seconds after deploying AI document processing on commercial submissions. Markel reported a 113 percent increase in underwriting productivity (measured as written premium per FTE) after restructuring their submission intake with AI.
These outcomes are from removing the document processing work that was sitting in front of them.
The Three Underwriting Documents That Cost the Most Time
Not all documents in a submission are equally expensive to process manually. Three account for the majority of underwriting prep time in commercial lines, and each has a distinct pattern that makes AI processing particularly effective.
Loss Runs
Loss runs are the most format-chaotic document type in commercial underwriting. They arrive from multiple prior carriers, each with its own structure, terminology, and level of detail. A manual review involves normalizing across formats, calculating totals, identifying trends, and cross-checking dates against policy periods. An experienced analyst doing this carefully takes 45 to 90 minutes per account.
A document AI agent handles format normalization automatically. It recognizes a claims table regardless of how it’s structured, extracts the relevant figures, calculates frequency and severity trends across the full loss history, flags anything that crosses a threshold, and outputs a structured summary. Processing time drops to under two minutes at consistent accuracy. At 40 submissions per week, that alone recovers three to four hours of senior staff time daily.
Schedules of Values
SOV processing is where errors in commercial property underwriting typically originate. A 12-location SOV submitted as a formatted Excel file looks straightforward. But locations have different coverage structures, some have older valuations, and the submitted replacement cost figures often don’t match what’s in the prior policy. Catching these discrepancies manually requires comparing two documents side by side for every location.
An AI agent reads both the current SOV and the prior policy simultaneously, flags any location where the submitted value differs from the prior by more than a configurable percentage, and surfaces any location missing a required field. The underwriter sees a clean exception report rather than two spreadsheets to reconcile.
ACORD Forms and Broker Applications
Commercial ACORD applications are dense. Extracting the relevant fields (entity structure, SIC code, revenue, payroll, prior carrier, coverage requested, loss history attestation) and verifying they’re consistent with the other documents in the package is tedious and error-prone when done manually. An AI agent extracts every relevant field, checks internal consistency across the submission package, and flags any discrepancy (a stated loss figure that doesn’t match the loss run, a coverage request that differs from the prior policy) before the underwriter opens the file.
What This Means for the 2027 Budget Conversation
For underwriting operations, the 2027 budget conversation is no longer about whether to invest in AI, but where that investment will produce the fastest and most measurable return. Industry data already shows the impact: Morgan Stanley projects a 200-basis-point improvement in expense ratios by 2030 for insurers that operationalize AI at scale, while AM Best reports a 2.4-point drop in underwriting expense ratios among carriers already using advanced automation. The key is targeting workflow-level opportunities with clear metrics. Document processing in commercial underwriting stands out because both the baseline (current processing time) and the improvement (speed and accuracy after deployment) are easy to measure, and the capacity freed translates directly into higher-value underwriting work.
Organizations that defined specific KPIs before deploying document AI (not after) achieved positive ROI six to nine months faster than those that measured retrospectively, according to a 2026 Insurance Thought Leadership analysis of enterprise AI deployments. That framing matters for budget approval: the outcome should be defined before the check is written.
Putting It Into Practice: Running Loss Runs Through Kudra
Here’s what this looks like in a real underwriting workflow, using a set of loss runs from a mid-market commercial property submission as the example.
The documents arrive: five years of loss runs from three prior carriers, each in a different format. One is a scanned PDF, one is an Excel export, one is a formatted table embedded in a broker email. Under the current process, an analyst spends 90 minutes to two hours normalizing across formats, extracting figures, and building a summary before the file reaches the underwriter’s desk.
With Kudra, the same loss runs run through an automated document workflow the moment they arrive.
Step 1: Build Your Workflow
Each component does one job:
- OCR layer: Handles the reality that loss runs rarely arrive in clean digital formats. Scanned PDFs, image files, and embedded tables all get converted into processable text before any extraction begins
- Table extractor: Recognizes that a claims history table is structured data, not narrative text, and pulls it accordingly, preserving row and column relationships across all three carrier formats regardless of how differently they’re laid out
- Entity detector: Identifies the specific fields that matter — claim dates, incurred amounts, paid amounts, open reserves, coverage line, and policy period, and extracts them cleanly and consistently across every loss run in the set
- Text completion: Generates a plain-language summary of the full loss history: total incurred, frequency by year, severity trends, and any claim that crosses a configurable threshold
You configure this once. Every set of loss runs that flows through it gets the same structured output, in the same format, in under two minutes.
Step 2: The Underwriter Opens a Summary, Not a Stack of Files
When the workflow runs, what arrives at the underwriter’s desk is not three carrier documents in three different formats. It’s a single normalized summary: five-year loss history consolidated across all prior carriers, frequency and severity calculated, year-over-year trend surfaced, and any flagged item (a large open reserve, a claim with significant paid-to-incurred movement, a policy period with unusually high frequency) called out for review.
The underwriter examines the flags, forms their view on the loss history, and moves to a decision. What previously took 90 minutes of manual normalization takes the agent under two minutes. The judgment is still theirs. The reading isn’t.
Step 3: Ask the Loss Runs Anything With Kudra Chat
Kudra Chat, currently in early access, extends this further. Rather than waiting for the workflow to surface specific flags, underwriters and analysts can ask direct questions of the full loss run set in natural language.
“What is the total incurred across all GL claims in the last five years?”
“Which policy period had the highest claim frequency?”
“Are there any open claims with reserve movement above 20 percent?”
The system reads across all three carrier documents, traces the figures, and returns a sourced answer in seconds. Every response cites the specific document and section it drew from, so the underwriter gets an answer and a reference they can verify, not a number with no trail.
We’ve tested Kudra Chat on loss run sets exceeding 200 pages across multiple carrier formats. The figures stay accurate, the references stay traceable, and the answers stay grounded in what the documents say.
For commercial underwriting specifically, this changes how fast a loss history can be understood, not by shortcutting the analysis, but by removing the 90 minutes of normalization work that currently sits in front of it.
The Honest Framing
AI agents in underwriting don’t price risk. They don’t build relationships with brokers. They don’t apply the kind of qualitative judgment that separates a senior underwriter’s assessment from a junior one.
What they do is clear the path for that judgment to happen faster, more consistently, and across more submissions than any team can currently handle manually.
The carriers and MGAs that are moving fastest in 2026 are the ones that recognized the document processing problem for what it is: not an inevitable cost of doing commercial underwriting, but a solvable bottleneck sitting between their people and the work those people are actually there to do.
The 2027 budget is the right moment to solve it.
