By the numbers
Insurance fraud costs the U.S. an estimated $308.6 billion a year, the pain that AI fraud detection targets (Coalition Against Insurance Fraud, via Triple-I, 2022).
Deloitte estimates P&C insurers could save up to US$160 billion by 2032 by deploying AI-driven, real-time fraud analytics (Deloitte 2026 insurance outlook).
Source: Deloitte Insights
McKinsey estimates generative AI could unlock $50-70 billion in additional insurance-industry revenue (McKinsey, via Reinsurance News, 2025).
Source: McKinsey (via Reinsurance News)
With AI-assisted underwriting, McKinsey reports quoting times cut from several weeks to days, and some commercial lines from multiple days to a few hours (McKinsey, via Reinsurance News, 2025).
Source: McKinsey (via Reinsurance News)
Gartner predicts at least 30% of generative AI projects are abandoned after proof of concept, blaming poor data quality, weak risk controls, cost, and unclear value (Gartner, 2024).
Source: Gartner (via THE Journal)
The EU AI Act (Annex III) classifies AI used for risk assessment and pricing in life and health insurance as high-risk, triggering documentation and oversight duties (EU Regulation 2024/1689).
Source: EU AI Act, Annex III
Key terms
- Straight-through processing (STP)
- Handling a transaction end to end, a quote, a policy issue, or a claim, with little or no manual intervention. In insurance, higher STP rates mean routine cases are auto-adjudicated while people focus on complex or high-risk ones.
- Document intelligence / Intelligent Document Processing (IDP)
- AI that reads unstructured insurance documents such as ACORD forms, submissions, medical records, and adjuster notes and turns them into structured, validated data, ideally with a source citation and confidence score for each field.
- ACORD forms
- Standardized insurance forms and data standards maintained by ACORD, such as certificates of insurance and loss runs. They are everywhere but usually arrive as PDFs, scans, and email attachments that still need extraction.
- Agentic AI
- AI that plans and carries out multi-step tasks, pulling data, calling tools, and drafting decisions, rather than answering a single prompt. In insurance it appears as underwriting 'work cells' or claims-triage agents that still route to a human for approval.
- Retrieval-augmented generation (RAG)
- A technique that grounds a language model's answers in specific retrieved documents, a policy wording, a guideline, or a claim file, so outputs cite real sources instead of relying on the model's memory, reducing hallucination in regulated workflows.
- Proxy discrimination / disparate impact
- When a model produces unfair outcomes for protected groups even without using protected attributes, because other data acts as a stand-in. Regulators including the NAIC and the EU AI Act expect insurers to test for and mitigate it.
At a glance
| Dimension | Build in-house | Insurtech SaaS / point solution | Boutique AI engineering partner |
|---|---|---|---|
| Time to production | 6-18+ months; needs ML hiring | Fast to switch on, but generic to your book | Weeks to a few months on your data |
| Insurance domain + core-system fit | You control it, if you have the talent | Pre-built for common lines; limited tailoring | Custom-built to your lines and workflows |
| Data + model ownership | Fully owned | Often vendor-owned; data may leave your walls | Negotiable, can be client-owned |
| Accuracy evaluation + governance | Depends on internal maturity | Vendor benchmarks, not your documents | Should ship an evaluation harness on your data |
| Regulatory documentation (NAIC / EU AI Act) | Your team must produce it | Varies by vendor; verify | Built into delivery if the partner is mature |
| Best fit when | AI is a core, permanent capability | A common, well-bounded use case | A high-value custom build you need in production fast |
What to check before you buy
- Insist on an evaluation harness scored against a labeled sample of your own documents, with a stated accuracy target, and treat any guaranteed accuracy number as a red flag.
- Require every extracted field and AI decision to be traceable to its source, with a document citation and confidence score, so an auditor or regulator can follow the reasoning.
- Confirm a human-in-the-loop design with no auto-decision on adverse actions: declines, pricing increases, or claim denials must route to a person.
- Ask for documentation aligned to the NAIC Model Bulletin (a written AI systems program and governance) and, for EU exposure, EU AI Act Annex III conformity.
- Test for proxy discrimination and disparate impact across protected classes before go-live, not after a complaint.
- Verify integration with your policy admin and claims core systems (for example Guidewire, Duck Creek, or a mainframe) and support for ACORD data standards.
- Start with a time-boxed feasibility sprint on real documents to prove value before committing to a multi-year build.
- Nail down data ownership, model IP, hosting location, and exit terms to avoid vendor lock-in.
Frequently asked questions
Where does AI pay off fastest in insurance?
Document-heavy, high-volume workflows usually return value first: claims intake and triage, fraud detection, and turning submissions and ACORD forms into structured data for underwriting. Deloitte estimates AI fraud analytics alone could save P&C insurers up to $160 billion by 2032, and McKinsey reports AI-assisted quoting falling from weeks to days. Start where you have volume, clear rules, and labeled data.
Is AI-driven underwriting and pricing allowed by regulators?
Yes, but under growing scrutiny. In the U.S., the NAIC Model Bulletin (adopted December 2023 and taken up by roughly half the states) expects insurers to run a governed AI systems program and stay accountable for third-party models. In the EU, the AI Act classifies life and health insurance risk-assessment and pricing AI as high-risk, with documentation and oversight duties. Existing anti-discrimination and unfair-trade-practice law still applies regardless of the model.
Why do so many insurance AI projects stall?
Gartner predicts at least 30% of generative AI projects are abandoned after proof of concept, usually because of poor data quality, weak risk controls, cost, or unclear business value rather than the model itself. Deloitte likewise finds about 90% of insurance leaders see the urgency of reinventing how work gets done, but only around 25% have taken tangible action. The fixes are unglamorous: clean data, tight scope, real evaluation, and integration into core systems.
Can AI fully automate claims (straight-through processing)?
Simple, well-structured claims can be largely auto-adjudicated, and STP rates are rising. Full automation of complex or contested claims is neither realistic nor advisable, and both good governance and regulators expect a human in the loop for denials and other adverse actions. The practical target is auto-handling routine cases while routing exceptions to adjusters with AI-prepared context.
Should we build in-house, buy a platform, or hire a partner?
It depends on how core and how custom the capability is. Buy a platform for a common, well-bounded use case; build in-house if AI is a permanent core competency and you have the talent to run it; hire a boutique AI engineering partner for a high-value custom build you need in production quickly, where some ship on roughly 90-day timelines. Weigh domain fit, data ownership, evaluation rigor, and regulatory documentation rather than brand names or headcount.
How accurate is AI at extracting data from insurance documents, and how do we measure it?
Accuracy depends heavily on document type and quality, so measure it on your own documents with a labeled test set, not a vendor demo. Mature teams ship an evaluation harness and report error rates per field. For reference, CONE RED has run bank-statement automation across 21 banks at a measured 0.24% transaction-direction error rate and built a Fortune 100 document classifier processing 5,000+ documents a month to a target of at least 95% accuracy, live in eight weeks. Treat any 'guaranteed' accuracy number with suspicion.
How do we keep AI from hallucinating or discriminating in regulated decisions?
Ground outputs in real documents using retrieval, and require citations and confidence scores so answers are traceable rather than invented. For fairness, test for proxy discrimination and disparate impact across protected classes and keep humans accountable for adverse actions. Both the NAIC bulletin and the EU AI Act push toward explainability, documentation, and testing, so build those in from the start rather than bolting them on later.
This guide is educational and even-handed: CONE RED is one option among several, and no honest firm guarantees a fixed accuracy, compliance, or business outcome. Every statistic above links to a source you can check; figures are current as of publication and may change.