AI for Underwriting

Submission intake and triage — the first and highest-value workflow

Submission triage is the single best starting point for AI in underwriting. It is high-volume, document-heavy, and the current manual process is a proven bottleneck.

The manual process today:

1. Broker email arrives with a submission package (ACORD forms, loss runs, supplemental applications, SOVs, sometimes a broker cover letter)
2. Underwriter opens each document, reads through them, and mentally assembles the risk profile
3. Key data is manually entered into the underwriting workbench or policy administration system
4. Underwriter makes a quick decision: review further, request more information, or decline
5. Elapsed time: 30-90 minutes per submission, depending on complexity

The AI-assisted process:

1. Upload the entire submission package to AI
2. AI extracts all key risk data: insured name, address, SIC/NAICS, ISO class, building construction, protection class, Total Insured Value, requested coverages and limits, years in business, loss history summary
3. AI compares the extracted data against your documented underwriting appetite
4. AI produces a structured triage summary with a preliminary recommendation
5. Underwriter reviews the summary, makes a judgment call, and either proceeds to full underwriting or declines
6. Elapsed time: 10-15 minutes including review

Here is a prompt template for commercial property submission triage:

I am a commercial property underwriter. Review this submission package and produce a structured triage summary.

Extract the following from the ACORD forms and supplemental application:
- Named insured and DBA
- Mailing address and all insured locations
- SIC/NAICS code and business description
- Building construction type (ISO construction class)
- Year built and any major renovations
- Square footage and number of stories
- Fire protection class
- Sprinkler protection (yes/no, type, coverage percentage)
- Total Insured Value (building and contents separately)
- Requested coverages and limits
- Deductibles requested
- Current carrier and expiring premium
- Years with current carrier
- Named insured's years in business

From the loss runs, extract:
- Total number of claims in the past 5 years
- Total incurred for each year
- Largest single loss (date, cause, incurred amount)
- 5-year loss ratio (total incurred / total earned premium if available)
- Frequency by cause of loss (fire, water, wind, theft, liability)

Flag any concerns:
- Losses trending upward
- Any single loss exceeding 25% of the building's TIV
- Frequent claims of the same type
- Gaps in coverage history
- Missing information needed for underwriting

Produce a preliminary appetite assessment based on these criteria:
[Insert your appetite guidelines here — e.g., "We write commercial property for ISO construction classes 1-4, fire protection classes 1-6, minimum 5 years in business, maximum TIV $50M per location, loss ratio below 60% over 5 years"]

Appetite matching — does this submission fit our book?

Every carrier has underwriting appetite guidelines — documented criteria defining what risks they want to write, what they will consider with referral, and what they will not touch. The problem is that these guidelines often live in lengthy underwriting manuals, memos from management, and the institutional memory of senior underwriters.

AI can operationalise your appetite in a way that makes it consistently applied across every underwriter on the team.

How to set this up:

1. Document your appetite criteria in a structured format. This is the most important step. If your appetite is not written down, AI cannot apply it.

2. Create an appetite reference document that AI can read alongside each submission. Example structure:

COMMERCIAL PROPERTY UNDERWRITING APPETITE

Preferred risks (can bind without referral):
- ISO construction classes 1-4 (fire resistive to masonry non-combustible)
- Protection classes 1-6
- Minimum 5 years in business
- 5-year loss ratio below 50%
- TIV per location: $1M - $25M
- Occupancy classes: office, retail, warehouse, light manufacturing
- Sprinkler protected required for TIV above $10M

Acceptable with referral:
- ISO construction class 5 (wood frame) with approved fire protection
- Protection classes 7-8
- 5-year loss ratio 50-70%
- TIV per location: $25M - $50M
- New ventures (less than 3 years in business) with experienced principals

Declined classes:
- Habitational (apartments, condos)
- Restaurants and bars
- Woodworking operations
- Auto body shops
- Any insured with arson conviction
- 5-year loss ratio above 70%

3. Include this document in every submission review prompt. AI will read both the submission and your appetite document, then assess whether the risk falls into preferred, referral, or decline territory.

The value here is consistency. When 15 underwriters apply appetite guidelines manually, interpretation varies. When AI applies documented criteria, every submission gets the same initial screen — and borderline cases are flagged for senior underwriter review rather than quietly declined or quietly bound.

Loss history analysis — the AI advantage

Loss history analysis is where AI delivers one of its clearest advantages over manual review. Underwriters currently spend significant time reading through years of loss runs — often in different formats from different carriers — trying to piece together a coherent picture of an insured's claims experience.

The challenge with manual loss run review:

• Loss runs come in dozens of formats (carrier-specific, TPA-specific, ACORD loss run format)
• Data must be mentally aggregated across multiple carriers and policy periods
• Open claims require judgment about ultimate value versus current incurred
• Large losses need to be separated from attritional frequency
• Development on open claims must be considered
• Cause of loss codes are inconsistent across carriers

What AI does with loss runs:

Upload all available loss runs and use this prompt template:

Review the attached loss runs for [Insured Name] covering the past [5/7/10] years.

Produce the following analysis:

1. LOSS SUMMARY TABLE
For each policy year, show:
- Policy period
- Carrier
- Number of claims
- Total incurred (paid + outstanding reserves)
- Largest single loss
- Loss ratio (if earned premium is available)

2. LARGE LOSS DETAIL
For any claim with incurred above $[threshold], provide:
- Date of loss
- Cause of loss
- Description
- Current status (open/closed)
- Paid to date and outstanding reserve
- If open, note the age of the claim and whether reserves appear adequate based on the description

3. FREQUENCY ANALYSIS
- Claims per year trend (increasing, stable, decreasing)
- Most common cause of loss
- Any clustering of claims by location, type, or time period

4. DEVELOPMENT CONCERNS
- Flag any open claims that may develop adversely based on the description and current reserves
- Identify any late-reported claims (significant gap between date of loss and date reported)
- Note any claims with reserve increases over time

5. UNDERWRITING RED FLAGS
- Loss ratio trending upward over the review period
- Any single loss exceeding [X]% of premium or TIV
- Repeated claims of the same type suggesting a systemic issue
- Gaps in coverage history

The result: In 5-10 minutes, you have a structured loss analysis that would take 30-60 minutes to assemble manually. More importantly, AI catches patterns that humans miss when reading through pages of loss run data — the slow creep of reserve increases on an aging claim, the clustering of water damage losses at one location, the gap year where the insured may have been uninsured or with a carrier that is not showing loss data.

Comparable risk analysis and supplemental application review

Comparable risk analysis is one of the most judgment-intensive parts of underwriting — and one where AI can significantly accelerate the research phase.

When pricing an unusual or complex risk, experienced underwriters draw on their mental database of similar risks they have seen before. AI can formalise this process:

I am pricing a [risk description — e.g., "75,000 sq ft cold storage warehouse in Houston, TX, ISO construction class 4, protection class 3, sprinklered, TIV $18M"].

Based on the risk characteristics, identify the key rating factors that would influence pricing:
- Construction and occupancy hazards
- Protection class and fire suppression adequacy
- Geographic exposures (wind, flood zone, earthquake)
- Industry-specific loss drivers for this occupancy class
- Typical coverage concerns and common exclusions

What questions should I be asking the broker or insured that I might not have thought of based on the standard application?

This does not replace your pricing judgment or your actuarial-driven rate models. It accelerates the risk assessment process by ensuring you consider all relevant factors, including ones that the standard application may not capture.

Supplemental application review is another high-value AI workflow. Many specialty lines require lengthy supplemental applications — cyber liability questionnaires, professional liability supplementals, environmental impairment applications. AI can:

1. Extract all answers from the supplemental into a structured format
2. Flag inconsistencies between the supplemental and the ACORD application
3. Identify questions left blank or answered ambiguously
4. Compare answers against typical responses for the insured's industry class
5. Generate a list of follow-up questions for the broker

Referral prioritisation and workflow optimisation

Most carriers have a referral process where submissions outside standard authority must be escalated to senior underwriters or management. The problem is that referral queues become bottlenecks — senior underwriters spend time reviewing referrals that should never have reached them, while genuinely complex risks wait.

AI can improve referral prioritisation in two ways:

First, by reducing unnecessary referrals. When AI applies appetite criteria consistently during initial triage, submissions that clearly fall within underwriter authority proceed without escalation, and submissions that clearly fall outside appetite are declined before reaching the referral queue.

Second, by enriching referral packages. When a submission does require referral, AI can prepare a structured referral summary that gives the senior underwriter everything they need to make a quick decision:

Prepare a referral summary for senior underwriter review.

Risk: [Insured name and description]
Reason for referral: [Exceeds authority limit / unusual occupancy / adverse loss history / new class of business]

Include:
1. One-paragraph risk summary
2. Key risk data (construction, occupancy, protection, TIV, territory)
3. Loss history summary (5-year loss ratio, largest loss, trend)
4. Specific concern requiring senior review
5. Recommended action (quote with conditions / decline / request more info)
6. Supporting rationale for the recommendation

The result: senior underwriters can review a referral in 2-3 minutes instead of 15-20, because the submission data has already been extracted, organised, and analysed. They focus their judgment on the specific question requiring their expertise rather than re-reading the entire submission.

Throughput improvement is measurable. Carriers that have deployed AI for submission triage and referral prioritisation report 30-50% increases in underwriting throughput — measured in submissions reviewed per underwriter per day — with no degradation in loss ratio. In fact, loss ratios often improve because AI ensures every submission is consistently screened against appetite criteria.

Key takeaways

1. Submission triage is the highest-value AI workflow for commercial lines underwriting — it is high-volume, document-heavy, and the current manual process is a proven bottleneck.
2. Appetite matching becomes consistent when AI applies documented criteria to every submission, flagging borderline risks for senior review rather than leaving interpretation to individual underwriters.
3. Loss history analysis accelerates dramatically — AI standardises data across carrier formats, calculates aggregates, and surfaces patterns that humans miss when reading raw loss runs.
4. Comparable risk analysis and supplemental review are accelerated by AI's ability to extract, structure, and compare data — but pricing judgment remains firmly human.
5. Referral prioritisation improves when AI both reduces unnecessary referrals and enriches referral packages for faster senior review.

Next up: AI for Claims Processing.