What is shadow AI in insurance?

Shadow AI in insurance is any AI tool employees use for work without approval from IT, security, or compliance. Common examples include pasting a loss run, medical record, or submission into a free public chatbot.

Why is shadow AI risky for insurers?

Insurers handle sensitive policyholder data, including medical history, Social Security numbers, and loss histories. When that data goes into a public AI tool, the insurer loses control over where it's stored, who can see it, and whether it trains another company's model.

How can insurers reduce shadow AI?

Insurers reduce shadow AI by finding the tasks employees hand to public AI tools and replacing those workarounds with approved AI built for insurance. Start with the workflows where data is most sensitive and volume is highest, then publish a clear policy that names approved tools.

How is Bevaya different from general-purpose AI?

Bevaya is an AI agent platform built for insurance, with models trained on insurance data rather than general web content. The platform is SOC 2 Type 2 attested, keeps each customer's data isolated, logs every action in an immutable audit trail, and routes low-confidence results to human reviewers.

What is least privilege for AI agents?

It means giving each AI agent only the data and permissions its task requires. An AI agent that indexes claim files gets those files and nothing else.

Do AI agents need access controls separate from employees?

Yes. AI agents should follow the same access rules as people, with boundaries set around each AI agent's job.

How often should insurers review AI agent access?

At least once a year, and whenever a workflow changes. New York's cybersecurity regulation requires annual access reviews for users, which makes it a practical baseline for AI agents too.

What's the difference between RBAC and AI agent access controls?

RBAC sets what each person can see and do based on their role. AI agent access controls set what each AI agent can reach. Strong programs use both.

Will insurance AI always need human review?

Some review always remains, but it shrinks as the AI learns. As corrections add up, fewer fields fall below your confidence threshold, so reviewers handle only the work that needs a person.

How long does it take for corrections to improve the AI?

Improvement arrives with each model release, not the moment a reviewer saves an edit. Corrections are used to retrain the model, and each new version is tested before it goes live.

How can we tell if our insurance AI is improving?

Track your no-touch rate over time, and compare accuracy against verified correct answers month by month.

Which certificate forms can AI create?

ACORD 25 and other ACORD certificates. AI identifies the form the request needs, fills it from live policy data, and attaches the required endorsement language.

Does AI-generated COI creation still need human review?

Only when a holder requirement doesn’t match the policy. AI flags the gap, explains the reason, and routes it to a reviewer. Matching requests issue without manual steps.

What happens when a policy renews or coverage changes?

AI refreshes active certificates so holders always have current proof of coverage on file, and logs an audit trail for each certificate it issues.

Why do insurance AI projects fail?

Most stall on process and people, not technology. AI repeats whatever process it's given, so undocumented steps and inconsistent handoffs become errors at scale. BCG finds two-thirds of the AI transformation challenge hinges on people.

Which insurance processes should be automated with AI first?

Start with high-volume work that brokers and policyholders feel directly, such as new business submissions or claims intake. These have clear inputs, measurable turnaround, and fast payback once the process is consistent.

How do you measure whether AI improved an insurance process?

Set a baseline before you automate. Track turnaround time, error rates, and missed deadlines, then compare the same metrics after AI is running. Broker complaints and rework volume show whether the change holds up in practice.

Does AI replace insurance staff?

No. AI agents take on repetitive tasks like data entry and document review, so people can focus on judgment, relationships, and service. Human reviewers still handle exceptions and own final decisions.

What is an insurance-trained AI model?

An insurance-trained AI model learns from real insurance documents such as loss runs, claim forms, and policies, rather than general web data. That training teaches it the field names, layouts, and conventions carriers actually use. Bevaya's InsurGPT is trained on more than 300 million insurance documents.

Can ChatGPT or other general AI read claims documents?

It can read the text, but it often misses the fields that matter most. In Bevaya's benchmark, the strongest general-purpose model read claim-tying fields correctly 72.8% of the time, against 86.6% for an insurance-trained model. General models also return answers without confidence checks or an audit trail.

Does claims AI make coverage or reserve decisions?

No. Claims AI reads documents, sets up files, and recommends next steps, but adjusters keep every claims decision. Values the AI isn't confident about go to a person before they reach the claim system.

What should claims leaders ask AI vendors about accuracy?

Ask how well the model reads the fields that tie a document to a claim, not just its overall accuracy. One wrong claim number barely moves a headline percentage but misfiles the whole record. Ask to see results field by field.

Can AI combine loss runs from several carriers into one loss history?

Yes. AI pulls every claim from each carrier's loss run into the same fields, whatever the format. The underwriter sees one consolidated loss history across carriers and policy years instead of reconciling them by hand.

Why can't general-purpose AI read loss runs accurately?

Loss run formats and conventions live in private carrier documents that general models never trained on. Values like total incurred often have to be calculated rather than read, so general models guess or leave the field blank. Newer general model releases improved by only a point or two.

How accurate is AI on loss runs in production?

With confidence scoring, a second Verifier model, and human review of flagged values, Bevaya's loss run AI reaches 98%+ accuracy in production. More than 90% of loss runs complete without anyone touching them.

Does AI replace underwriters in loss run review?

No. AI reads and reconciles the loss runs and sends uncertain values to a reviewer. Underwriters still judge the risk, set the price, and decide whether to quote.

Which general-purpose AI models did Bevaya test?

Twelve leading general-purpose models from four model families and three generations, tested in 18 configurations, some with and without reasoning. Each was accessed through its API with an optimized prompt and no insurance-specific training. The full list and every result are on Bevaya's benchmarks page.

How was field accuracy scored in the benchmark?

Field accuracy is the share of fields read correctly on each loss run, averaged across all 346 documents. Every model read the same documents and was scored the same way. Bevaya also measured the share of loss runs needing no correction, where the gap was wider.

Can verification and human review make up for a general-purpose model?

Only partly. Verification and review catch errors, but they start from the model's accuracy, and a model reading at 85% leaves roughly twice as many wrong fields as one at 93%. That means more files opened by people and less work that goes straight through.

Will AI replace jobs at insurance companies?

AI agents take over repetitive tasks like data extraction and document intake, while judgment calls stay with people. Staff typically shift to higher-value work like complex claims and risk assessment.

How should insurance staff be trained on AI tools?

Tie training to how each role changes, not a generic tool walkthrough. Give people permission to ask basic questions well past week one.

Who should lead change management for an AI rollout?

Leadership should own it visibly, since leadership support strongly predicts whether staff trust a new tool. Day-to-day managers should handle questions and fix early friction quickly.

How can you tell if staff are accepting a new AI tool?

Watch whether people keep using it and reporting issues, or quietly go back to the old process. Silence is often a warning sign, not a good one.

What loss run formats can AI read?

Carrier PDFs, Excel exports, scanned images, and faxes. Bevaya’s Loss Run Processing AI Agent reads hundreds of carrier and TPA formats out of the box, with no per-carrier configuration.

How does AI handle loss runs from multiple carriers?

It normalizes terminology, aligns valuation dates, and merges everything into one continuous loss history. Fields labeled “case reserves,” “indemnity reserves,” and “indemnity case” on different reports all map to the same underlying value.

Does AI know whether ALAE counts toward total incurred?

Yes. Whether ALAE sits inside or outside total incurred varies by carrier, and the AI accounts for that when it rolls up the numbers.

Does AI replace the underwriter’s judgment on loss runs?

No. The AI handles extraction, the math, and reconciliation, and the underwriter still decides how to price the risk. Anything low-confidence goes to a reviewer with the source document linked.

What tasks are insurance brokers automating with AI first?

Submission intake, policy checking, certificate of insurance creation, and renewal preparation. These are high-volume, document-heavy tasks where reading and acting on documents saves the most time.

Is AI replacing insurance brokers?

No. AI takes on reading, keying, and cross-checking documents so producers and account managers spend more time advising clients and winning business.

How is AI different from OCR or RPA for brokers?

OCR and RPA follow fixed rules and fail when a document looks unfamiliar. AI reads a document in context, decides what to do, and carries out the next step, which is what lets it handle varied carrier forms.

Does the EU AI Act apply to US insurers?

Yes, if they place an AI system on the EU market or their AI output affects EU customers. Where the company is headquartered doesn't matter.

Is claims AI considered high-risk under the EU AI Act?

Not by default. The Act's high-risk list covers life and health insurance pricing and risk assessment, but GDPR still applies to automated claim decisions.

What are the penalties for violating the EU AI Act?

Fines reach up to €35 million or 7% of global annual turnover for prohibited practices. Most other violations, including high-risk obligations, carry fines up to €15 million or 3%.

What should insurers do now to prepare for the 2027 deadline?

Start recording every AI decision, routing uncertain results to human review, and keeping a version history of model changes. That's the same documentation E&O and D&O underwriters already ask for at renewal.

What is an ACORD form?

ACORD forms are standardized insurance forms that brokers use to submit applications and coverage details to carriers. Their consistent format makes them a natural starting point for automation.

How accurate is AI ACORD form extraction?

Insurance-trained AI can extract ACORD fields with 99%+ accuracy. Bevaya's ACORD Form Extraction AI agent reaches that level across every form and variant, and routes anything uncertain to a reviewer.

How long does it take to deploy AI for ACORD forms?

Carriers, MGAs, and TPAs typically go live with Bevaya in 8 to 12 weeks. Brokers keep sending forms the same way, so nothing changes on their end.

Can extracted ACORD data flow into existing underwriting systems?

Yes. The extracted, validated data lands directly in the underwriting workbench, so underwriters work in the tools they already use.

What is AI governance in insurance?

AI governance is the set of policies, controls, and monitoring an insurer uses to keep its AI systems accurate and accountable. It covers who owns each model, how it's tested before deployment, and how it's monitored once live.

What is the NAIC AI Model Bulletin?

It's guidance from the National Association of Insurance Commissioners on how insurers should govern AI. States that adopt it expect a written AI program covering accountability, testing, and ongoing monitoring.

How often should insurers review AI governance metrics?

Continuously, with a formal review at least monthly. Metrics pulled by hand each quarter arrive too late to catch a problem before it compounds.

Insurance AI Blog

Bevaya ExpertsOctober 27, 20267 min read

Shadow AI in Insurance: The Unapproved AI Tools Haunting Your Data

Somewhere in your organization, someone is pasting a loss run into a free chatbot. ...

Start Reading

Bevaya ExpertsOctober 22, 202615 min read

October 2026: Insurance AI Trends & Highlights

What happened in AI and insurance during October 2026? See the month's most important ...

Start Reading

Bevaya ExpertsOctober 20, 20267 min read

How Should Insurers Control What AI Agents Can Access?

When a new underwriter, adjuster, or account manager starts, someone decides which files ...

Start Reading

Bevaya ExpertsOctober 13, 20266 min read

How Insurance AI Learns From Your Team's Corrections

An underwriter fixes a field the AI got wrong on a submission. The next time a similar ...

Start Reading

Bevaya ExpertsOctober 7, 20265 min read

How Does AI Automate Certificate of Insurance Creation?

A general contractor emails a certificate request and needs it by end of day. The request ...

Start Reading

Bevaya ExpertsSeptember 29, 20265 min read

Insurance AI Results Depend on Process Quality

Insurance organizations deploy AI to move submissions, claims, and service requests ...

Start Reading

Bevaya ExpertsSeptember 24, 202616 min read

September 2026: Insurance AI Trends & Highlights

Read up on September's leading AI and insurance stories driving industry momentum, ...

Start Reading

Bevaya ExpertsSeptember 23, 20267 min read

Why Claims AI Needs Insurance-Trained Models, Not General AI

Claims teams are absorbing more complexity with fewer senior people. High deductibles ...

Start Reading

Bevaya ExpertsSeptember 23, 20266 min read

AI for Loss Runs: What Bevaya's Benchmark Means for Underwriting Teams

A commercial submission rarely arrives with one loss run. It arrives with three or four, ...

Start Reading

Bevaya ExpertsSeptember 23, 20266 min read

Insurance-Trained vs. General-Purpose AI on Loss Runs: What Bevaya's Benchmark Found

With the attention general-purpose AI has drawn, a fair question follows: is it good ...

Start Reading

Bevaya ExpertsSeptember 22, 20267 min read

5 Reasons Insurance Staff Resist AI and How to Plan for Each One

Before you roll out an AI agent, plan for this. A veteran adjuster will double-check ...

Start Reading

Bevaya ExpertsSeptember 15, 20265 min read

How AI Automates Loss Runs

Underwriters at carriers and MGAs lose up to 40% of their time to manual submission work, ...

Start Reading

Bevaya ExpertsSeptember 9, 20265 min read

How Are Insurance Brokers Using AI?

It's 4pm on a Friday. A client needs a certificate of insurance before a vendor contract ...

Start Reading

Bevaya ExpertsSeptember 1, 20267 min read

What the EU's AI Regulations Mean for Insurance Underwriting and Claims

The EU's AI regulations changed for insurance carriers, MGAs, and TPAs last month, and ...

Start Reading

Bevaya ExpertsAugust 27, 202615 min read

August 2026: Insurance AI Trends & Highlights

The key AI and insurance stories driving market activity, regulatory developments, and ...

Start Reading

Bevaya ExpertsAugust 25, 20265 min read

How Does AI Extract Data From ACORD Forms?

Nearly every commercial submission starts with an ACORD form, and every quote can stall ...

Start Reading

Bevaya ExpertsAugust 19, 20268 min read

How to Measure AI Governance in Insurance

By now, most insurance leaders have heard they need AI governance in place. What almost ...

Start Reading

Bevaya ExpertsAugust 11, 20267 min read

How AI Extracts and Routes Legal Demand Details In Insurance Claims

A demand letter lands in a shared claims inbox on a Friday afternoon, sitting between a ...

Start Reading

Bevaya ExpertsAugust 4, 20267 min read

How Resistance to Change Affects Your Insurance AI Budget Case

An AI budget case for insurance needs more than proof of what the technology can do. One ...

Start Reading

Bevaya ExpertsJuly 30, 202620 min read

July 2026: Insurance AI Trends & Highlights

Check out the latest news and insights to stay in the know about the trends and ...

Start Reading

Bevaya ExpertsJuly 28, 20265 min read

How Does AI Extract Data From a Statement of Values?

A property submission arrives with a Statement of Values spread across six spreadsheet ...

Start Reading

Bevaya ExpertsJuly 22, 20268 min read

How Does AI Classify Insurance Claims Documents?

A single claims packet can run 30 or 40 pages, carrying a demand letter, a police report, ...

Start Reading

Bevaya ExpertsJuly 14, 20269 min read

What a Good QA Process Looks Like for Insurance AI

Most insurance teams spend significant time evaluating AI before they buy it. They ...

Start Reading

Bevaya ExpertsJuly 7, 20265 min read

3 Reasons Insurance AI Pilots Stall and What It Takes to Reach Production

Eighty-six percent of insurers are confident AI can deliver on their business goals, but ...

Start Reading

Bevaya ExpertsJune 30, 20268 min read

How AI Automates Claim to Policy Comparison in Insurance

Every new loss raises the same question: is it covered? It sounds like a simple question, ...

Start Reading

The Platform →

Build

Run

Review

Govern

Underwriting Automation →

Claims Automation →

Policy Servicing →

AI Agent Library →

Technology

Accuracy

Trust

Stories

Who It's For

Why Bevaya

Featured Case Study →

Learn

Research

Updates

Featured Report →

About

News & Trust

Connect

Blog

Shadow AI in Insurance: The Unapproved AI Tools Haunting Your Data

October 2026: Insurance AI Trends & Highlights

How Should Insurers Control What AI Agents Can Access?

How Insurance AI Learns From Your Team's Corrections

How Does AI Automate Certificate of Insurance Creation?

Insurance AI Results Depend on Process Quality

September 2026: Insurance AI Trends & Highlights

Why Claims AI Needs Insurance-Trained Models, Not General AI

AI for Loss Runs: What Bevaya's Benchmark Means for Underwriting Teams

Insurance-Trained vs. General-Purpose AI on Loss Runs: What Bevaya's Benchmark Found

5 Reasons Insurance Staff Resist AI and How to Plan for Each One

How AI Automates Loss Runs

How Are Insurance Brokers Using AI?

What the EU's AI Regulations Mean for Insurance Underwriting and Claims

August 2026: Insurance AI Trends & Highlights

How Does AI Extract Data From ACORD Forms?

How to Measure AI Governance in Insurance

How AI Extracts and Routes Legal Demand Details In Insurance Claims

How Resistance to Change Affects Your Insurance AI Budget Case

July 2026: Insurance AI Trends & Highlights

How Does AI Extract Data From a Statement of Values?

How Does AI Classify Insurance Claims Documents?

What a Good QA Process Looks Like for Insurance AI

3 Reasons Insurance AI Pilots Stall and What It Takes to Reach Production

How AI Automates Claim to Policy Comparison in Insurance