AI in insurance underwriting: how it works in commercial P&C
.png)
- AI in underwriting reads submissions, researches the account, and ranks risk so underwriters spend their time on judgment, pricing, and broker relationships instead of data entry.
- Adoption is uneven: Verisk reports 70% of surveyed respondents run AI in production for underwriting, while WTW found only 16% of North American P&C insurers use it to augment human underwriting today.
- Judge AI underwriting on outcomes (turnaround, hit ratio, premium per underwriter, loss ratio) and on whether every data point traces back to its source for regulators.
AI in underwriting is the use of machine learning and language models to read submissions, pull and check the data, research the account, and score the risk before an underwriter makes the call. In commercial P&C, it does the preparation work. The underwriter still decides whether to quote, at what terms, and at what price.
That definition sounds simple. The practice is not. According to a March 2026 Verisk and Reuters Insights survey, 70% of respondents have moved AI into full production for underwriting use cases. Yet a WTW survey of 59 North American P&C insurers found only 16% currently use AI to augment human underwriting, with 60% planning to make it a priority by 2028. The gap between those two numbers is where most carriers and MGAs sit right now: AI is running somewhere in the building, but it has not changed how risks get selected and priced.
This guide covers what AI in insurance underwriting actually does, where it fits in the workflow, which outcomes it should move, and how to evaluate an AI underwriting vendor without being sold a demo.
What is AI in underwriting?
AI in underwriting is a set of capabilities, not one product. In commercial lines it usually covers four jobs:
- Reading. Classifying and extracting data from ACORD forms, loss runs, schedules of values, supplementals, driver schedules, and the broker email itself.
- Checking. Validating extracted values, catching conflicts across documents, and flagging what is missing.
- Researching. Pulling external data for the line of business, such as license boards, sanctions lists, DOT and FMCSA safety records, or property hazard data.
- Ranking. Scoring risk factors, matching the account against appetite, and putting the best-fit submissions at the top of the queue.
What AI underwriting is not: a replacement for the rating engine, the policy admin system, or the underwriter's authority. The strongest deployments sit on top of the existing stack and hand the underwriter a prepared, sourced account. The weakest ones produce a summary nobody can verify.
Why AI in insurance underwriting matters now
Three pressures are pushing AI from pilot to production in commercial P&C.
Profitability is diverging. WTW found that P&C insurers with more sophisticated analytics and AI posted combined ratios six percentage points lower and premium growth three points higher than slower adopters between 2022 and 2024. That is a correlation, not proof of cause. It is still a gap no CUO wants to be on the wrong side of.
Capacity is fixed. Submission volume keeps rising while experienced underwriters retire. A team that spends an hour or more assembling each file quotes fewer accounts, answers brokers slower, and loses the risks it most wanted to write.
Regulators now expect a paper trail. About 25 jurisdictions have adopted the NAIC Model Bulletin on the Use of AI Systems by Insurers, according to a 2026 state-by-state tracker. The bulletin expects insurers to govern AI they use, including models from third-party vendors, and to explain decisions that AI supports.
How is AI used in insurance underwriting?
AI shows up at every stage of the commercial underwriting workflow. The underwriter's role changes at each stage. It does not disappear.
| Workflow stage | What AI does | What the underwriter still decides |
|---|---|---|
| Submission intake and clearance | Captures the email and attachments, detects duplicates, checks completeness, requests missing items | Whether a borderline account deserves a closer look |
| Document extraction | Classifies each document and extracts named insured, exposures, payroll, revenue, vehicles, values, limits, and loss history | Nothing, if the data is validated. Everything, if it is not |
| Data validation | Flags conflicts across forms, stale loss runs, and figures that do not reconcile | How to resolve a real discrepancy with the broker |
| External research | Pulls line-specific sources such as license status, sanctions checks, safety ratings, and hazard data | Whether a finding changes appetite or terms |
| Appetite and triage | Compares the account with coded appetite rules and ranks the queue by fit and value | Exceptions, referrals, and strategic accounts |
| Risk scoring | Scores individual risk factors as positive, negative, or neutral and shows each one | The final risk view, terms, and conditions |
| Pricing and quote | Delivers rater-ready data to the rating engine or workbench | Price, deductibles, limits, and whether to bind |
Submission intake and clearance
Intake is where most AI underwriting programs start, because it is where the most time disappears. A broker email arrives with five attachments and a note about documents still to come. AI reads the broker email and its context, classifies each attachment, checks completeness, and runs a first appetite pass before anyone opens the file. Out-of-appetite accounts get declined in minutes instead of after an hour of prep. For the step-by-step mechanics, see our guide to submission intake automation in commercial P&C.
Document extraction and validation
Extraction is the foundation. If a payroll figure or a loss total is wrong, every step after it is wrong too, including the price. Insurance documents also drift: brokers reformat loss runs, carriers revise forms, and scanned supplementals arrive sideways. Template-bound tools break on that drift. Insurance-trained models that read context hold up better, and a human validation layer catches what the model misses.
External research and enrichment
Underwriters used to spend part of every file on web research. AI does this by line of business: a CPA firm gets license board and sanctions checks, a trucking account gets FMCSA and DOT records, a property schedule gets hazard data. One rule matters here. The application stays the ground truth. External data enriches the account; it does not overwrite what the insured submitted.
Risk scoring and prioritization
Most carriers still work submissions first in, first out. AI scoring lets the team work the best-fit accounts first. The scoring model matters. A deterministic, factor-level score that shows why an account ranked where it did is something an underwriter will trust and a regulator will accept. A single opaque number is neither.
AI underwriting vs manual underwriting
The difference is not that AI does the underwriting. It is where the underwriter's hours go.
| Dimension | Manual underwriting | AI-assisted underwriting |
|---|---|---|
| Time to prepare a file | 60 minutes or more per submission is common | Data delivered structured and validated, with the underwriter reviewing exceptions |
| Appetite check | After the file has been prepared | At the front, against coded rules |
| Queue order | Inbox order | Ranked by appetite fit, value, and urgency |
| Data entry | Re-keyed from PDFs into raters and CRMs | Extracted into structured fields and pushed to existing systems |
| External research | Ad hoc, depends on the underwriter | Consistent sources by line of business |
| Consistency | Varies by person and by day | Same factors scored the same way on every account |
| Audit trail | Notes and email threads | Every value linked to its source document and every edit logged |
| Underwriter focus | Admin prep plus risk analysis | Risk selection, pricing, and broker relationships |
What outcomes should AI in underwriting move?
Fields extracted and documents read are outputs. They are useful for measuring a vendor's operations. They are not why a carrier buys AI. The outcomes that matter sit on the income statement.
- Speed to quote. Brokers send the best risks to whoever answers first. Faster turnaround means more quotes on accounts you want.
- Hit ratio. Working the in-appetite accounts first, with complete data, raises the share of quotes that bind.
- Premium per underwriter. When prep work leaves the desk, each underwriter writes more business without adding headcount.
- Loss ratio. Accurate exposure and loss data, plus consistent scoring, means fewer mispriced risks on the books.
These outcomes compound. Even a 2-point improvement in loss ratio on a $500M book is $10M in underwriting margin. Speed without accuracy moves the first number and damages the last one.
Track these metrics from day one, not only extraction accuracy:
- Submission-to-quote time and time to first broker response
- Quote rate and bind rate on target accounts
- Decline speed on out-of-appetite risks
- Missing-information rate and rounds of broker follow-up
- Written premium per underwriter
- Loss ratio movement by segment over time
Where AI underwriting programs go wrong
Most stalled programs share a small set of causes. Our look at why AI underwriting pilots stall goes deeper, but the pattern is consistent:
- Buying speed and ignoring accuracy. A fast summary built on a wrong loss total is worse than a slow correct one.
- Designing without underwriters. They know which fields drive decisions and which are noise.
- Treating the model as the product. The hard part is validation, workflow fit, and integration with the PAS and CRM.
- Skipping the broker email. Effective dates, target limits, and missing-document notes often live there.
- No audit trail. If a value cannot be traced to its source, the underwriter will re-check it by hand, and the time savings vanish.
- Measuring pilots on outputs. Documents processed per hour says nothing about hit ratio or loss ratio.
AI underwriting governance and the NAIC bulletin
Governance has become the gating factor. Grant Thornton’s 2026 AI Impact Survey found that only 24% of insurance leaders are very confident they could pass an independent review of AI governance and controls within 90 days. Another 44% say governance or compliance challenges have contributed to an AI project failing or underperforming. Verisk's respondents named data privacy and security as their top concern, followed by the accuracy and reliability of AI decisions and explainability.
For an underwriting team, governance comes down to three questions an examiner will ask. Where did this value come from? What did the system do with it? Who approved the decision?
AI that produces clean data with no lineage is just a faster black box. The defensible version ties every extracted value to its source document, scores every risk factor visibly, and logs every human edit.
That is why field-level provenance belongs in the requirements, not the nice-to-have list.
How to evaluate AI underwriting vendors
Demos run on clean files. Your inbox does not. Use these criteria to separate an AI underwriting platform from a document reader with an AI label.
| Criterion | What to ask | Red flag |
|---|---|---|
| Accuracy | What field-level accuracy is written into the contract, and what happens when you miss it? | Accuracy quoted from a demo set, with no contractual commitment |
| Human validation | Who reviews the AI's output, your team or ours? | Your underwriters become the QA team |
| Format tolerance | Run it on your ugliest loss runs and handwritten supplementals | Needs templates per broker or carrier |
| Explainability | Show me why this account scored the way it did, factor by factor | A single score with no visible drivers |
| Provenance | Click any value and show me the source page | Summaries that cannot be traced to documents |
| Integration | How do you connect to our PAS, CRM, and rater, and how long does it take? | Requires replacing the policy admin system |
| Lines of business | Which commercial lines are live with clients today? | One line in production, the rest on a roadmap |
| Outcome evidence | What happened to turnaround, premium per underwriter, and loss ratio for clients like us? | Only output metrics such as pages processed |
If your team is weighing an internal build, our build vs buy analysis for underwriting automation covers the tradeoffs.
How to roll out AI in underwriting
Teams that get the most from AI for underwriting start narrow and expand once the data is trusted.
- Pick the highest-friction workflow, usually a high-volume shared inbox or a program with long turnaround.
- Define the fields that drive clearance, appetite, rating, and referral for that line of business.
- Start with classification and extraction, with validation in place, and feed corrections back.
- Add appetite rules, external research, and risk scoring once the data is trusted.
- Connect to the systems underwriters already use, then measure turnaround, hit ratio, and loss ratio against a baseline.
How Pibit.AI approaches AI in underwriting
Pibit.AI built the CURE™ platform for commercial P&C carriers and MGAs. It sits on top of existing policy admin systems rather than replacing them, and integration typically takes weeks, not months. Five modules cover the workflow: ClearCURE for intake and clearance, DocumentCURE for extraction, ResearchCURE for external data, RiskCURE for factor-level risk scoring, and WorkflowCURE as the underwriting workbench. Teams start with one or two modules and expand.
The difference is the validation layer. AI does the first pass, and Pibit's own review team validates the output, so the client's underwriters never become the QA team. That is what backs a contractual 99.9% field-level accuracy commitment. Every value links back to its source document, and every risk factor stays visible.
Carriers and MGAs using Pibit have seen up to a 20% higher new business hit ratio, twice the throughput per underwriter and up to 70% faster time to quote.
If you are scoping AI in underwriting, start with the messiest inbox you have and the loss runs your team dreads. That is where the time is hiding. See CURE run on your own submissions.
Frequently Asked Questions
AI in underwriting is the use of machine learning and language models to read submissions, extract and validate data, research the account, and score risk before an underwriter decides. In commercial P&C, AI handles the preparation work, such as loss runs, ACORD forms, and appetite checks, while the underwriter keeps authority over whether to quote, the terms, and the price.
No. AI replaces the manual preparation work, not the underwriting judgment. It reads documents, checks appetite, and ranks the queue, so underwriters spend more time on risk selection, pricing, and broker relationships. Verisk's 2026 survey found carriers pursuing measured automation that keeps human-in-the-loop judgment, and regulators expect a person to stay accountable for AI-supported decisions.
Accuracy depends on whether a human validates the AI's output. AI-only extraction varies with document quality and format drift, which is why many pilots stall. Platforms that pair AI with a validation layer do better: the Pibit CURE platform carries a contractual 99.9% field-level accuracy commitment, with every value linked back to its source document for audit.
Head of Marketing and Partnerships
.png)
Ready to optimize
