Generative AI in Insurance, Explained

Generative AI is the phrase behind much of the current noise about AI in insurance. It is also applied too broadly. The useful distinction concerns the task: does the system estimate or classify something, or does it generate new content? Input and output format alone do not settle it. A classifier can read language, and a generative model can return structured data.

What generative AI is, and how it differs from the AI insurers already use

Insurers have used predictive and classification systems for years. A pricing model estimates loss, a fraud model ranks a claim, and an underwriting model classifies or scores an application. The task is to estimate an outcome or assign a category.

Generative AI produces new content: a paragraph, summary, answer, image, audio clip, or structured response assembled from a prompt and context. Most text systems in use today are built on a large language model, itself a kind of foundation model.

The distinction changes validation. A predictive model may be tested for calibration, rank ordering, stability, and error against observed outcomes. A generative system needs task-specific evaluation of factuality, grounding, completeness, consistency, and safe refusal. Both require evidence. They require different evidence.

Where generative AI shows up in insurance

The early, practical uses are unglamorous, which is a good sign. They cluster in document-heavy work, the reading and writing that fills an insurance day.

In underwriting, generative models read submissions, emails, PDFs, and images and pull the relevant facts into a system, work that used to mean manual data entry. In claims, they summarize long files so an adjuster can grasp a case quickly. In service, they draft the routine correspondence and power the chat windows customers now expect. Across all of these, the pattern is the same: the model handles the first draft or the raw extraction, and a person reviews and decides.

The money behind this is real. McKinsey has estimated that generative AI could add $50 billion to $70 billion in insurance industry revenue, with the impact concentrated in marketing, customer operations, and software engineering.¹ Treat that as a consultancy’s projection rather than a booked result, but the reason carriers are moving is not mysterious: the work it targets is expensive and slow.

The new risk: the confident wrong answer

Generative systems add a distinctive failure: the fluent answer that is unsupported or false. They can also expose confidential data, follow malicious instructions embedded in retrieved documents, or produce inconsistent answers to equivalent questions. Hallucination is the easiest failure to recognize, not the only one.

A generative model is not a policy database. It generates a response from patterns in its training and the context supplied at runtime. Sometimes it invents a policy provision, misstates a coverage limit, or cites a rule that does not exist, using the same tone it uses for a supported answer. In insurance, that presentation can make an error look like authorized advice.

For insurance, where a misstated coverage term can become a claim dispute years later, this is not a small problem. It is why the serious deployments keep a person between the model and the customer, and why “the model said so” is never a defense.

Why regulators look at it differently

Traditional model governance often starts with defined inputs, outputs, performance measures, and observed outcomes. A large language model’s explanation is generated text, so it cannot serve as the audit trail by itself. Explainability instead depends on preserved prompts, retrieved sources, model and configuration versions, evaluation results, and the downstream action. Model validation still applies, but the test set and acceptance criteria must match the task.

None of that puts generative AI outside the rules. It puts it inside the same AI governance framework as everything else, while making that framework harder to apply. The NAIC Model Bulletin is written around AI systems generally, and asks for governance, risk management, and internal controls scaled to what a system can do to a consumer.² It is not silent on the technique, either: it defines generative AI as its own class of system, and it tells examiners that validation, testing, and auditing should reflect whether what they are looking at rests on a predictive model or a generative one.² It stops short of writing a separate program for it. An examiner still expects you to know where the model is used, what it touches, and how you check it. The NAIC evaluation tool starts by counting models by operational area, then branches to governance, high-risk detail, or the data categories behind them depending on what those counts show.³ You just cannot lean on the same tests you use for a scoring model. Where generative tools sit relative to other AI across the business is mapped in the AI use cases by business line guide.

How to think about governing it

The starting move is the same one that governs all insurance AI: sort by what the output does to a consumer, not by how advanced the model is. A generative model that drafts internal notes is low stakes. One that writes something a customer relies on, or that feeds a coverage or claims decision, is not.

From there the controls follow the risk. Build an evaluation set from real insurance tasks and define acceptance criteria before deployment. Ground answers in approved documents where accuracy matters, preserve the prompt, retrieved sources, model version, response, and downstream action, and route uncertain or consequential outputs to a person. For customer-facing use, the evidence should show how unsupported answers are detected and what happens when the system cannot answer safely.

Generative AI is governed today, under rules already on the books. Whether it gets rules of its own is still open. At this article’s July 2026 review, the NAIC framework and state instruments treated it as ordinary AI rather than a separate regulatory category, and the Big Data and Artificial Intelligence Working Group was the place where a change would surface first. Until that changes, the newest tool in insurance runs on rules written before it existed, and the seams are being found by whoever hits one.

Footnotes

1. McKinsey & Company estimate, as reported by Reinsurance News, “Gen AI could unlock $50-$70bn in insurance revenue, estimates McKinsey & Company,” February 27, 2026: https://www.reinsurancene.ws/gen-ai-could-unlock-50-70bn-in-insurance-revenue-estimates-mckinsey-company/ . A consultancy projection reported via trade press, not a booked result; the trade report gives the figure as additional industry revenue. ↩

2. NAIC, “Model Bulletin: Use of Artificial Intelligence Systems by Insurers,” adopted December 4, 2023: https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf ↩ ↩²

3. NAIC, “AI Systems Evaluation Tool 4.0” (Exhibit A asks only for counts by operational area; Exhibit B governance, Exhibit C high-risk system detail, Exhibit D data details each branch from those counts): https://content.naic.org/sites/default/files/inline-files/AI%20Systems%20Evaluation%20Tool%204.0%20%28Clean%29.pdf ↩