AI use cases in insurance: 12 systems teams build in 2026
Twelve AI systems carriers, MGAs and brokers are building across underwriting, claims and service, with regulatory checks and illustrative team budgets.
By the Ryz Labs team · Updated October 2026
The AI use cases insurers build first are commercial submission intake, claims triage, medical and claim file summarization, subrogation detection and SIU referral support, followed by policy servicing agents. Under the NAIC model bulletin on insurers' use of AI, which many states have adopted, each of these needs a documented governance program, so the systems that ship are the ones designed for testing and oversight from day one.
This guide covers twelve systems across underwriting, claims and policyholder service at carriers, MGAs and brokers, with the data, integrations, failure modes and an illustrative team and budget for each. Ryz serves enterprises across industries, including insurance. Our AI pod teams build in your cloud and repos, working on your team.
About the numbers. Team shapes and build times are illustrative planning ranges, not quotes. Costs use only Ryz's public rates: mid-level $7,000 to $10,000, senior $10,000 to $15,000 and tech lead $15,000 and up per engineer per month, quoted per team, with ranges at the $15,000 lead floor.
The rules that shape insurance AI
Three regulatory threads matter before any build starts. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, expects a written AI systems program covering governance, risk management, internal controls and oversight of third-party models, and states that adopt it can ask for that documentation in market conduct exams. The New York Department of Financial Services' Circular Letter No. 7 (2024) sets expectations for using AI systems and external consumer data in underwriting and pricing, including testing for unfair discrimination. Colorado's SB21-169 and its implementing regulations require governance and testing for insurers that use external consumer data and algorithms. Separately, unfair claims settlement practices laws still apply to every claim decision, whatever software helped make it.
Underwriting
1. Commercial submission intake
Broker submissions arrive as emails with ACORD forms, statements of values, loss runs and supplemental applications attached. The system classifies each attachment, extracts insured, class, locations, values and requested coverages, checks appetite, and creates a clean submission record so underwriters work the best-fit risks first.
- Data: submission emails and attachments, appetite guides, historical bound and declined submissions.
- Integrations: underwriting workbench or policy admin system (Guidewire, Duck Creek, Sapiens), email, CRM, third-party data enrichment.
- What makes it hard: statements of values with thousands of rows in inconsistent layouts, and duplicate submissions from competing brokers. See AI underwriting document processing.
- Illustrative team and build: tech lead plus three senior engineers, four to six months: $45,000 to $60,000 a month, $180,000 to $360,000 in total.
2. Loss run extraction and normalization
Loss runs from prior carriers come in hundreds of formats. The system extracts claims, dates, paid, reserved and incurred amounts, normalizes them into one table, and checks that extracted totals equal the totals printed on the report.
- Data: loss run PDFs and spreadsheets, labeled examples per carrier format.
- Integrations: submission intake, rating and pricing tools.
- What makes it hard: totals must tie exactly; any mismatch goes to a person rather than being smoothed over.
- Illustrative team and build: tech lead plus two senior engineers, three to four months: $35,000 to $45,000 a month, $105,000 to $180,000 in total.
3. Underwriting guidelines assistant
Underwriters ask "can we write a restaurant with a liquor exposure over this threshold in this state?" The system answers from the current guidelines and referral authorities only, cites the section and tells the underwriter when a referral is required.
- Data: underwriting manuals, authority letters, referral rules, filed rules.
- Integrations: the underwriting workbench, document management, identity provider.
- What makes it hard: guidelines vary by state, program and effective date. Retrieval must filter by all three. See RAG development.
- Illustrative team and build: tech lead plus one senior and one mid-level engineer, two to three months: $32,000 to $40,000 a month, $64,000 to $120,000 in total.
4. Property risk from imagery
Computer vision models read aerial and street-level imagery to estimate roof condition, roof shape, vegetation near structures, pools and other features that matter for property risk, and compare them with what the application says.
- Data: licensed aerial imagery, inspection reports, labeled property features, claims outcomes.
- Integrations: imagery vendor APIs, rating engine, inspection ordering.
- What makes it hard: image dates (a roof can be replaced after the photo), and NAIC-style governance expectations for any model that influences an underwriting decision. See computer vision development.
- Illustrative team and build: tech lead plus three senior engineers, five to seven months: $45,000 to $60,000 a month, $225,000 to $420,000 in total.
5. Life underwriting medical record summaries
Life and disability underwriters read attending physician statements and medical records that can run to hundreds of pages. The system builds a timeline of diagnoses, medications, labs and procedures, cites the page for each, and highlights items that underwriting guidelines say need attention.
- Data: medical records obtained with the applicant's authorization, lab results, guideline manuals.
- Integrations: new business and underwriting systems, records retrieval vendors.
- What makes it hard: handwriting and scanned faxes, and strict handling of health information under state insurance privacy laws and the applicant's authorization. The underwriter decides; the summary only organizes.
- Illustrative team and build: tech lead plus three senior engineers, four to six months: $45,000 to $60,000 a month, $180,000 to $360,000 in total.
Claims
6. First notice of loss intake and triage
A voice or chat agent takes first notice of loss, collects the facts, photos and parties, checks coverage basics in the policy system, and routes the claim by severity and complexity, so a straightforward glass claim and a bodily injury claim never wait in the same queue.
- Data: FNOL transcripts, claim outcomes, routing rules, policy data.
- Integrations: claims system, telephony, policy admin, photo upload.
- What makes it hard: callers are often stressed; the agent must recognize injury, emergency or fraud signals and hand off to a person quickly.
- Illustrative team and build: tech lead plus four senior engineers, five to seven months: $55,000 to $75,000 a month, $275,000 to $525,000 in total.
7. Claim file summarization for adjusters
When a claim changes hands or reaches a review point, the system summarizes the file: coverage, liability notes, medical treatment, reserves history, demands and open tasks, with links to each source note or document.
- Data: adjuster notes, correspondence, medical bills and records, reserve history.
- Integrations: claims system, document management.
- What makes it hard: long files with contradictions. The summary must surface conflicts, not resolve them silently.
- Illustrative team and build: tech lead plus two senior engineers, three to five months: $35,000 to $45,000 a month, $105,000 to $225,000 in total.
8. Subrogation opportunity detection
Recoveries are missed when nobody reads the note that says "the other driver ran the light" or "the water heater was installed last month." The system reads notes and documents on closed and open claims, flags likely third-party liability or product defects, and drafts the referral.
- Data: claim notes, police reports, photos, past subrogation outcomes.
- Integrations: claims system, subrogation workflow.
- What makes it hard: statutes of limitation and deadlines vary by state, so referrals carry their deadline from the start.
- Illustrative team and build: tech lead plus two senior engineers, three to four months: $35,000 to $45,000 a month, $105,000 to $180,000 in total.
9. SIU referral support
A fraud model scores claims and documents for indicators (inconsistent dates, reused photos, repeat providers, altered invoices) and gives the special investigations unit a referral with the evidence attached. A Ryz AI pod built fraud detection for a global fleet company that scored 244K+ invoices in under 30 seconds each and surfaced $5.94M in fraud confirmed by the client's own fraud team (case study); the invoice-scoring pattern carries over to provider and repair invoices on claims.
- Data: claims, invoices, provider data, confirmed fraud outcomes.
- Integrations: claims system, SIU case management, industry databases.
- What makes it hard: a referral is not a denial. Unfair claims settlement practices rules still govern the claim, and the score must never delay a valid payment without a human decision.
- Illustrative team and build: tech lead plus three senior engineers, five to seven months: $45,000 to $60,000 a month, $225,000 to $420,000 in total.
Service and distribution
10. Policy servicing agent
A chat and voice agent handles ID cards, billing questions, address changes, adding a driver or vehicle, and certificate of insurance requests, by calling the policy admin system with the same permission checks a service rep has.
- Data: service transcripts, knowledge articles, policy data.
- Integrations: policy admin, billing, telephony, document generation.
- What makes it hard: some changes need a licensed agent or have state-specific rules. The agent must know where its authority ends.
- Illustrative team and build: tech lead plus four senior engineers, five to seven months: $55,000 to $75,000 a month, $275,000 to $525,000 in total.
11. Outbound renewal and payment outreach
An outbound voice agent calls policyholders about upcoming renewals, missing documents or a payment before cancellation, answers basic questions and books a callback with an agent when needed. A Ryz AI pod built an AI voice platform that has placed 1M+ outbound calls.
- Data: policy and billing status, contact preferences, call outcomes.
- Integrations: telephony, CRM, policy admin, consent records.
- What makes it hard: calling consent and time-of-day rules under the TCPA and state law, and required notice language that cannot be paraphrased.
- Illustrative team and build: tech lead plus three senior engineers, four to six months: $45,000 to $60,000 a month, $180,000 to $360,000 in total.
12. Complaint classification and DOI response drafting
The system classifies complaints from every channel, links each to the policy and claim, and drafts the response to a department of insurance inquiry from the file, for compliance to finalize before the deadline.
- Data: complaints, DOI inquiries, claim and policy files, past responses.
- Integrations: complaint tracking, claims and policy systems.
- What makes it hard: deadlines and accuracy. Every factual statement must trace back to the file.
- Illustrative team and build: tech lead plus two senior engineers, three to four months: $35,000 to $45,000 a month, $105,000 to $180,000 in total.
How to pick your first use case
- Start where a professional already decides. Submission intake, file summaries and subrogation referrals help underwriters and adjusters without changing who decides, which keeps governance simple.
- Write the governance file as you build. Inventory, intended use, testing for unfair discrimination where relevant, and monitoring plans are easier to produce during the build than after an exam request.
- Pick a line of business with volume. One high-volume line gives you enough labeled examples to evaluate properly.
- Separate operational AI from rating and underwriting decisions. Systems that read documents or summarize files for a person carry less regulatory weight than models that change price or eligibility. Starting with the former lets you build the governance muscle, the inventory and the testing habits before you touch a filed rating plan.
- Reuse the document layer. Intake, loss runs and claim summaries share parsing, retrieval and review screens.
What it costs
Cost is team size times months times monthly rate. Ryz rates: mid-level engineers $7,000 to $10,000 per month, senior engineers $10,000 to $15,000, tech leads $15,000 and up, quoted per team.
- Underwriting pilot (tech lead plus two senior engineers) for four months: $35,000 to $45,000 a month, $140,000 to $180,000 in total.
- Claims pod (tech lead plus four senior engineers) for six months: $55,000 to $75,000 a month, $330,000 to $450,000.
- Full pod of about seven senior engineers including a tech lead and an ML engineer: $75,000 to $105,000 a month.
Cloud and model usage run in your accounts and are not included. You meet the named engineers first, and a pod can usually start in under two weeks once you say go. Ryz engineers have experience working within HIPAA and SOC 2 controls; certifications for your environment stay with you. If you want a packaged claims platform with its own models, license one; Ryz pods build in your stack.
Related
FAQ
What are the main AI use cases in insurance?
The most common production systems are commercial submission intake, loss run extraction, claim triage and file summaries, subrogation detection, SIU referral support, servicing agents and complaint response drafting. Pricing and underwriting models carry the most regulatory scrutiny.
What does the NAIC model bulletin require?
It expects insurers to maintain a written program for the responsible use of AI systems, covering governance, risk management, internal controls, testing and oversight of third-party systems, and to be able to show regulators how AI-supported decisions comply with insurance laws. Check which states have adopted it for your lines.
Can AI deny an insurance claim?
Claim decisions remain subject to unfair claims settlement practices laws and the insurer's own accountability. Most carriers use AI to organize files, flag issues and draft, and keep coverage and denial decisions with licensed adjusters.
Is AI fraud scoring allowed in claims?
Yes, insurers widely use fraud indicators to refer claims to SIU. The score supports an investigation; it should not by itself delay or deny a valid claim, and it should be tested and monitored under your governance program.
Can a Ryz pod work with Guidewire or Duck Creek?
Yes. Pods build integrations against your policy, billing and claims platforms through their APIs and integration layers, in your cloud and repos, and you own all code and IP.
Questions we didn't answer? Email info@ryzlabs.com.