AI Underwriting: How It Works and the Companies Building It

On this page (8)

1. TL;DR
2. What AI underwriting does, stage by stage
3. How it differs by line of business
4. Where humans stay in the loop
5. The main types of AI underwriting vendors
6. Regulation and fairness
7. How to evaluate an AI underwriting vendor
8. FAQs

TL;DR

• AI underwriting means using machine learning and language models for the work around a risk decision: reading submissions, filling gaps with third-party data, scoring risk against appetite, supporting the price and routing the file.
• In personal lines, small commercial and life accelerated underwriting, low-risk applications can be approved with no person involved. In mid-market and specialty commercial lines, AI prepares the file and the underwriter still decides.
• Vendors fall into five groups: submission intake, underwriting workbenches, data enrichment, risk scoring models and pricing platforms. Most carriers use more than one.
• Regulators treat AI decisions like any other underwriting decision. The NAIC Model Bulletin (adopted by 29 jurisdictions as of October 8, 2026), Colorado's SB21-169 rules and NY DFS Circular Letter No. 7 all expect governance, testing for unfair discrimination and oversight of vendors.

AI underwriting is the use of machine learning and language models to do the work around an underwriting decision: reading the application or broker submission, filling in missing facts from outside data, scoring the risk, checking it against appetite and guidelines, supporting the price, and routing the file. For simple risks such as personal auto, homeowners, small business packages and some life policies, carriers let the system approve applications without a person. For larger commercial and specialty risks, the AI prepares the file and recommends a next step, and an underwriter makes the call.

The rest of this guide covers what that looks like at each stage, by line of business, where people stay involved, which companies build it, and the rules that apply.

What AI underwriting does, stage by stage

Underwriting has the same basic steps whether a person or a model does them. AI shows up at each one.

Stage	What the AI does	Example
Submission intake	Reads broker emails and attachments and turns them into structured fields	A wholesale submission with an ACORD 125, a statement of values and five years of loss runs becomes one record with locations, values and claims
Clearance and triage	Checks for duplicates and conflicts, confirms completeness, ranks submissions by fit and likelihood to bind	A queue sorted so the underwriter opens the in-appetite, likely-to-bind accounts first
Data enrichment	Pulls facts the applicant did not provide from public and third-party sources	Business class, operations, hours and liquor exposure for a restaurant from its name and address
Risk scoring	Produces a score for a peril or for the account as a whole	A property-level wildfire score for a homeowners application
Appetite and guideline matching	Compares the risk to the carrier’s written guidelines and authority levels	A flag that a schedule is more than 30% unsprinklered, which needs a referral
Pricing support	Feeds risk factors into the rating model, or runs the model inside the workbench	An underwriter comparing pricing scenarios before quoting
Decision and routing	Approves, declines or refers under rules the carrier sets	A low-risk small business application bound with no underwriter touch; an out-of-appetite risk declined with a generated letter

Intake is where most commercial AI projects start, because the documents are the bottleneck. Loss runs are a good example: each carrier prints them in its own layout, and someone has to rekey claims before the underwriter can use them. Our guide to what a loss run report is explains the format, and the loss run parsing software comparison covers the tools that extract them.

How it differs by line of business

Personal lines and small commercial

Personal auto, homeowners and small business owners’ policies have high volumes, standard forms and rating plans built from known variables. That makes them suited to straight-through processing: the application comes in, data is prefilled and verified, rules and models produce a decision, and the policy is issued with no underwriter involved. Anything unusual is referred.

AI adds two things here. Prefill and enrichment shorten the application and catch misstatements, and peril models score each property or business individually rather than by ZIP code. Earnix says its Underwrite-It product combines rules with machine learning to fully automate underwriting decisions for straight-through processing. Planck, now part of Applied Systems, describes its commercial data as enabling automated underwriting and straight-through processing. Gradient AI lists straight-through processing of low-risk applications as one of its P&C underwriting uses.

Mid-market, specialty and E&S commercial

Here, submissions arrive by email from brokers, each risk is different, and pricing involves judgment. AI is used to cut the time before judgment begins: reading the submission, checking clearance, scoring appetite fit, summarizing loss history and surfacing red flags. Kalepa’s site shows the kind of signals involved, such as an incumbent carrier non-renewing, building permits for a recent remodel, or a clean driver history. The underwriter reviews and decides. Some carriers configure the system to decline clearly out-of-appetite submissions without a review.

Life insurance: accelerated underwriting

Traditional life underwriting relies on a paramedical exam and fluid tests and, according to the NAIC, can take up to a few months from application to issue. Accelerated underwriting skips the exam for eligible applicants and uses external data instead: prescription drug history, motor vehicle records, credit-based data and the Medical Information Bureau, scored by predictive models. The NAIC says this can cut the application process from several weeks to a few hours. Applicants whose data is insufficient still go through full underwriting with an exam.

AI is also used inside full underwriting. Sixfold says its life and health agent reviews prescription histories, lab results and driving records, surfaces key impairments and suggests a next step, with underwriters owning the final decision. Sixfold’s site quotes Guardian saying the product cut underwriter review time by 50%.

Where humans stay in the loop

Most products and most regulators agree on the same boundary: the further a decision is from a simple, rules-defined risk, the more a person should own it.

• Setting the rules. Underwriting leaders and actuaries write the appetite, guidelines, authority limits and rating plan. The AI applies them; it does not decide what the carrier wants to write.
• Referrals and exceptions. Risks outside authority, unusual classes, large limits and conflicting data go to a person.
• Complex and large risks. Specialty and large commercial accounts are quoted by underwriters, with AI preparing the file.
• Adverse decisions. Declines and surcharges carry regulatory risk. New York requires specific reasons for adverse decisions made with AI systems, and many carriers route them for review.
• Model oversight. Someone has to monitor drift, review errors and retest for unfair discrimination. The NAIC Model Bulletin text names the extent of human involvement in final decisions as one factor in how strong the controls should be.

The main types of AI underwriting vendors

AI underwriting is a stack of products, and most carriers use several. Tool profiles are linked where we have them.

Vendor type	What it does	Examples
Submission intake and digitization	Reads broker submissions and documents into structured data and routes them	Cytora
Underwriting workbench and scoring agents	Brings the submission, data, appetite fit, rating and quote into one underwriter view; scores and recommends	Kalepa, Sixfold, hyperexponential, Federato
Data enrichment	Builds a risk profile from public data so the applicant answers fewer questions	Planck
Risk scoring models	Scores a property, business or employer for a specific peril or loss outcome	ZestyAI, Gradient AI
Pricing and rating platforms	Builds and deploys pricing models and underwriting rules	Earnix, Akur8

Kalepa covers the full commercial path on one platform. Its site lists submission ingestion, clearance (including sanctions screening and producer appointment checks), triage, risk analysis, rating with models from Excel or external raters, quote and bind, and portfolio management.

Sixfold builds underwriting agents for commercial P&C and for life and health. On the P&C side it says the product reads loss runs and statements of values, scores risk with each fact cited to its source, evaluates appetite and portfolio fit, and recommends a next action.

hyperexponential (hx) calls itself a pricing and underwriting platform for commercial P&C and markets its product as an agentic underwriting workbench. Its site lists submission triage with ACORD extraction and appetite scoring, complex risk assessment with bespoke actuarial models, renewals, portfolio monitoring and reinsurance management. The company says carriers use it for decisions across more than $75bn of annual commercial P&C premium.

Cytora describes itself as a risk digitization platform for brokers, insurers and reinsurers. It turns client submissions, quotes, policies and invoices into schemas, then runs configurable workflows for clearance, appetite evaluation and routing. The company says the platform does not require model training.

Federato is known for portfolio-aware underwriting triage and now markets itself as an AI-native core system covering policy administration, billing and claims, with a Control Tower module for real-time portfolio steering.

Planck builds commercial risk profiles from a business name and address, including classification by NAICS code and exposures such as liquor, delivery and operating hours. Applied Systems acquired it, and its site lists submission prefill, underwriting recommendations, policy monitoring and renewal recommendations.

ZestyAI scores individual properties for perils including wildfire, hail, wind, severe storm, non-weather fire and non-weather water, and offers roof age data. The company says it has more than 500 regulatory approvals, which matters for carriers that need to file a model as part of a rate or rule filing.

Gradient AI sells risk scoring for workers’ compensation, P&C and group health underwriting and claims. On its P&C page it reports client results including an 80% decrease in quote turnaround time; those are the company’s own figures.

Earnix and Akur8 sit on the pricing side. Earnix’s Underwrite-It manages underwriting rules and machine learning models together and lets underwriters simulate rule changes before deploying them. Akur8 builds transparent GLM and GAM pricing models for actuaries and deploys them through its rating engine.

For side-by-side detail, see the underwriting workbench software comparison, the insurance rating and pricing software list and the property data providers list. None of the vendors above publishes pricing; all are quoted by sales (checked October 2026).

Regulation and fairness

US regulators apply existing law, including unfair discrimination and unfair trade practice rules, and have added expectations about how insurers govern the models and data behind decisions.

NAIC Model Bulletin on the Use of AI Systems by Insurers

The NAIC adopted the Model Bulletin on December 4, 2023. It is a template for state insurance departments, not a law, and applies only where a state adopts it. The NAIC’s implementation map, current as of October 8, 2026, lists 29 adopting jurisdictions, including Illinois, Pennsylvania, New Jersey, Virginia, Washington and, most recently, North Dakota (October 5, 2026), plus the District of Columbia and Puerto Rico. The map lists California, Colorado, New York and Texas separately as having their own insurance-specific AI rules or guidance.

The bulletin, as adopted by Illinois in Company Bulletin 2024-08, expects each insurer to keep a written AIS Program that:

• covers governance, risk management controls and internal audit, with senior management accountable to the board;
• applies across the insurance life cycle, including underwriting and rating and pricing;
• covers AI systems built by the insurer and those bought from third-party vendors;
• documents data practices such as data lineage, quality and bias analysis;
• sets standards for diligence on vendors, and expects contracts that support audits and regulatory inquiries.

Insurers should expect examiners to ask for the written program and for documentation on third-party models.

Colorado SB21-169 and Regulation 10-1-1

Colorado’s SB21-169, signed July 6, 2021, requires insurers to test their use of external consumer data and information sources (ECDIS), algorithms and predictive models for unfair discrimination based on race, color, national or ethnic origin, religion, sex, sexual orientation, disability, gender identity or gender expression, and to fix harms they find. The Division of Insurance carried this out first for life insurance underwriting. Regulation 10-1-1 (3 CCR 702-10), the governance and risk management framework, took effect November 13, 2023. The amended version, effective October 15, 2025, covers life insurers, private passenger auto insurers and health benefit plan insurers. Insurers in those lines that do not use ECDIS must file an officer-signed attestation by December 1 each year. A separate quantitative testing regulation for life insurance was still at the draft proposed stage on the Division’s page when we checked. Our guide to Colorado SB 24-205 and the NAIC bulletin covers how Colorado’s broader AI act relates to these insurance rules.

NY DFS Circular Letter No. 7 (2024)

New York’s Department of Financial Services issued Circular Letter No. 7 on July 11, 2024. It applies to insurers authorized in New York, Article 43 corporations, HMOs, fraternal benefit societies and the State Insurance Fund, and covers underwriting and pricing that use ECDIS or AI systems. Its main expectations:

• Fairness testing before deployment and regularly after: check for disproportionate adverse effects, look for a legitimate rationale, and search for less discriminatory alternatives. External data must also be assessed as a possible proxy for protected classes.
• Governance with board oversight, senior management responsibility, written policies and internal audit.
• Transparency: reasons for adverse decisions that include the data sources used, and notice to declined applicants with a way to review the data.
• Vendor oversight: the insurer stays responsible for third-party tools and data and should keep audit rights in vendor contracts.

For tools that help document and test models against these rules, see AI governance software for insurers.

How to evaluate an AI underwriting vendor

1. Start from the bottleneck. Decide whether the problem is reading submissions, missing data, triage, risk selection or pricing. Each points to a different vendor type.
2. Pilot on your own submissions. Use recent files, including messy ones: scanned loss runs, multi-tab schedules, forwarded email chains. Measure fields corrected and recommendations your underwriters overturned.
3. Check that outputs cite their sources. An underwriter should be able to see which document or data source each fact and flag came from.
4. Ask who changes the rules. Find out whether underwriting leaders can update appetite and guidelines themselves, and whether every change is logged.
5. Ask for the governance file. Request the vendor’s model documentation, unfair discrimination testing approach, data sources and drift monitoring, in a form you can hand to a regulator under the NAIC bulletin, Colorado or New York rules.
6. Confirm the data rights. Find out where external data comes from, whether it is licensed for underwriting use in your states, and whether your submissions are used to train shared models.
7. Map the integration. List what the tool needs from your policy administration, rating and document systems, and what it writes back.
8. Review security. Ask for current SOC 2 or ISO 27001 reports and data residency options.
9. Get pricing in writing. Vendors in this category do not publish prices, so ask how the fee scales: per submission, per underwriter, by premium or by line.

FAQs

What is AI underwriting?

AI underwriting is the use of machine learning, predictive models and large language models in the insurance underwriting process. The models read applications and submission documents, pull in outside data, score the risk, check it against appetite and guidelines, and either recommend a decision to an underwriter or, for simple risks, make the decision under rules the carrier sets.

How is AI used in insurance underwriting today?

The common uses are submission intake (turning broker emails, ACORD forms, loss runs and schedules into data), triage and appetite matching, prefill of business or property details from public data, risk scores for perils such as wildfire, pricing models, and accelerated underwriting in life insurance that replaces the medical exam with data checks for eligible applicants.

What is automated underwriting in insurance?

Automated underwriting is a process where a system approves, declines or refers an application based on rules and models, without an underwriter reviewing it. It is common for personal auto, homeowners, small business packages and accelerated life products. Applications that fall outside the rules are referred to a person.

Will AI replace insurance underwriters?

Not on current evidence. Vendors such as Sixfold describe underwriters as reviewing and owning the final call, and regulators expect human oversight scaled to the potential harm of a decision. AI takes over document reading, data gathering and simple approvals, which moves underwriters toward complex risks, referrals and broker work.

What are the best AI underwriting companies?

It depends on the job. Kalepa, hyperexponential and Sixfold build underwriter workbenches and scoring agents for commercial lines; Cytora focuses on submission intake; Planck and ZestyAI supply data and risk scores; Gradient AI offers risk scoring for workers' compensation, P&C and group health; Earnix and Akur8 build pricing and rating platforms.

What is the best AI underwriting software for commercial insurance?

For commercial carriers and MGAs, the shortlist usually starts with an underwriting workbench (Kalepa, hyperexponential, Federato) or a submission scoring and intake layer that sits in front of existing systems (Sixfold, Cytora). The right choice depends on which systems you keep. See the full comparison of underwriting workbench software.

Is AI underwriting legal?

Yes, but the decisions must comply with the same insurance laws as any other underwriting decision, including unfair discrimination and unfair trade practice laws. States that adopted the NAIC Model Bulletin expect a written AI program; Colorado requires governance and testing from life, private passenger auto and health benefit plan insurers that use external consumer data; New York requires testing, explanations of adverse decisions and vendor oversight.

How accurate is AI underwriting?

There is no independent, industry-wide accuracy figure. Vendors publish customer results, such as faster quote turnaround or reduced review time, but these are self-reported and measured on their customers' own books. The practical test is a pilot on your own recent submissions, compared against what your underwriters decided.