---
title: "Artificial intelligence in insurance: claims and limits | Trustable"
url: "https://trustable.ai/insights/artificial-intelligence-in-insurance-evidence-and-limits"
description: "See how insurers use AI in underwriting, pricing, service and claims, with the inputs, evidence limits and accountable decisions that differ for each use."
---

Insights · November 6, 2026 · 12 min read

# Artificial intelligence in insurance: uses, evidence and limits

See how insurers use AI in underwriting, pricing, service and claims, with the inputs, evidence limits and accountable decisions that differ for each use.

In an illustrative motor claim, not a Trustable customer case, an adjuster receives a clean-looking repair invoice and damage photos. A summary tool repeats the invoice’s repair total and the photos’ apparent damage as established facts, leaving the reliability of the underlying evidence unresolved. Artificial intelligence in insurance supports decisions across a policy’s life, from underwriting and pricing to service and claims. Each use needs controls that match the decision it can affect and the evidence needed to support it.

## Claims contain several different evidence questions

Claims AI can organize a submission, identify apparent damage, suggest a reserve or help a handler navigate a file. Each output answers a different question.

At Trustable, our [claims authenticity checks](https://trustable.ai/insurance) focus on submitted documents, photos and recorded video, flagging signs of editing or AI generation for review. That is a narrower job than estimating a loss, interpreting coverage or authorizing a payout. We supply findings for the adjuster or special investigations unit (SIU); the insurer decides what those findings mean for the claim.

Allianz’s public statement about its [claims copilot deployment](https://www.allianz.com/en/mediacenter/news/articles/250205-smarter-claims-management-smoother-settlements.html) illustrates the division. This is an industry example, not a Trustable customer case. The system brings together claim and contract information, highlights discrepancies and drafts communications. Allianz describes coverage and payout decisions remaining with skilled claims professionals. That is a reported workflow, not evidence that the same design or performance transfers to every insurer.

The tasks in the illustrative motor claim remain separate:

- Document extraction identifies the amount written on the invoice. It establishes what the submission says.
- Damage assessment identifies visible damage and may estimate repair needs. It depends on the views and detail supplied.
- Authenticity analysis raises a concern about possible editing in a submitted file. It supplies a reason for further review.
- The adjuster assesses the evidence alongside the policy and loss circumstances to determine the claim’s treatment within assigned authority.

A concern in one task should remain visible to the others. If a photo’s authenticity is unresolved, a generated summary should attribute the damage to the submitted photo rather than present it as an established condition of the vehicle. An estimate derived from that photo inherits the unresolved dependency.

![Illustrative workflow: one submitted photo supplies a summary, estimate and file signals for adjuster review.](https://trustable.ai/assets/articles/insurer-ai/evidence-dependencies.0e8b0b7ebb25.webp)

Conversely, a file with no detected manipulation does not establish when the damage occurred, which event caused it or whether the policy covers it. A detector score or localized signal is not overall model accuracy, a calibrated probability of fraud, ground truth or proof of intent. Missing or unusable analysis is an evidence gap, never an authenticity clearance.

When the missing fact concerns current condition, a guided remote inspection or a site inspection can supply additional views through the insurer’s collection process. Fresh capture is a separate control from analyzing an existing submission. It cannot recreate a vehicle’s condition before repairs. Our [claim photo guide](https://trustable.ai/insights/ai-generated-claim-photos) covers the initial authenticity checks without turning one signal into a claim decision.

## The same AI label covers different jobs

Predictive models learn patterns from data to produce estimates or scores. Computer vision interprets images. Language systems extract information, answer questions or generate text. An AI agent can also call tools and carry out permitted actions. These functions can sit in the same workflow while retaining separate responsibilities. EIOPA, the European Insurance and Occupational Pensions Authority, distinguishes [predictive and generative uses](https://www.eiopa.europa.eu/publications/traditional-ai-generative-ai-implications-insurance-sector_en), including assistants whose outputs still need human validation.

This distinction matters when several systems work on the same file. Reading an invoice accurately does not resolve whether its amount was altered. A summary can repeat an unsupported statement even when the writing is clear.

## How insurers use AI across the policy lifecycle

The National Association of Insurance Commissioners (NAIC) documents [insurer AI uses](https://content.naic.org/insurance-topics/artificial-intelligence) across underwriting, pricing, service and claims. Separating the jobs makes their limits clearer.

How insurers use AI across the policy lifecycle

| Use | Typical inputs | Useful AI output | Decision beyond that output |
| --- | --- | --- | --- |
| Underwriting | Application, loss history, property or applicant information | Risk classification or referral recommendation | Whether to accept the risk and on what terms |
| Pricing | Exposure, claims experience and rating variables | Expected loss estimate or risk score | The premium and rating approach the insurer applies |
| Customer service | Customer question, account information and policy documents | Answer, retrieved passage or draft response | Whether the response is appropriate and what action is authorized |
| Claims handling | Notice of loss, correspondence and claim documents | Extracted fields, summary or routing suggestion | Which facts are established and what handling follows |
| Damage assessment | Photos, inspection material and repair data | Damage classification or repair estimate | Cause, extent of covered loss and settlement value |
| Fraud analytics | Claim patterns, transaction history and linked records | Anomaly or investigation lead | Whether the concern warrants investigation and what it establishes |
| Evidence authenticity | Submitted documents, photos or recorded video | Signs of alteration or AI generation, with available locations and limitations | Whether the file can support a claim fact after review |

A system can contribute to several rows. Its approval for one task should not silently authorize another. Permission to summarize a claim does not include permission to settle it.

## Underwriting and pricing need a reason for the difference

Underwriting concerns risk selection and terms. Pricing concerns the amount charged. They share data, but their outputs need separate justification.

In its public statement on [AI-assisted roof assessment](https://www.zurich.com/about-us/ai-at-zurich), Zurich describes using aerial imagery in US underwriting. This is an industry example, not a Trustable customer case. The useful distinction is between a model’s assessment of a roof and the insurer’s treatment of the property. Image age, a recent repair or an incorrectly matched building can change what reliance is reasonable.

An underwriting record needs the information behind the recommendation, the applicable appetite rules and a way to resolve contradictory evidence. An underwriter’s ability to change the proposed action matters more than a nominal approval button.

For pricing, predictive performance alone leaves other questions open. The actuarial and pricing owners need a basis for the variables used, the relationship to insured risk, and the resulting treatment of policyholders. Data that improves a prediction can still introduce an unacceptable proxy or produce unfair outcomes.

New York’s [underwriting and pricing guidance](https://www.dfs.ny.gov/industry-guidance/circular-letters/cl2024-07) makes that distinction concrete. It addresses actuarial validity, unfair discrimination, governance and transparency for AI systems and external consumer data within its scope. It also leaves responsibility with the insurer when a third party supplies the tool. It is not a claims-handling rule for every jurisdiction.

## A service answer can become a consequential action

AI can help a service team locate policy wording or draft a reply. The risk changes when a customer relies on that reply to understand coverage, meet a deadline or change a policy.

A useful service design distinguishes general information from account-specific answers and actions. A general explanation of deductibles needs different access and review from a response about a particular claim. A policy change requires the customer’s authority and the insurer’s approved process, beyond a plausible conversation.

Generated answers can be confidently wrong, including their explanations and citations. The National Institute of Standards and Technology describes this as confabulation in its [generative AI risk profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf).

The service owner therefore needs approved source material, the relevant policy version and a human escalation path when the answer is unsupported or disputed. Access controls should prevent one customer’s information from appearing in another customer’s response. A missing policy passage should remain an unanswered question rather than an invitation for the model to complete the story.

The intended benefit is less time spent searching and drafting. The acceptance evidence is whether responses stay grounded in the right records and unresolved questions reach someone with authority to answer them.

## Automation changes what the system is allowed to do

A summarizer produces text. An agent that sends correspondence, changes a record or starts a payment can alter the insurer’s position. The distinction is authority.

For each use, the operating design needs a clear boundary between recommendations and permitted actions. A service assistant may draft an endorsement request while the policy system controls the change. A claims assistant may propose further evidence collection while the adjuster owns the request and its relevance.

Human oversight has substance when a named person can see the source material, understand the limitation, reject the recommendation and obtain further evidence. A reviewer who sees only a generated conclusion has little basis to challenge it.

Outsourcing the model does not outsource these decisions. Business owners define permitted uses; technical owners manage access and failures; risk and compliance teams provide challenge within their responsibilities. The decision record needs to connect what the system supplied with what the insurer actually did.

## Oversight follows the use and the jurisdiction

There is no single approval called “AI approved” that settles every insurance use case.

In the US, the NAIC’s [AI model bulletin](https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf) sets out expectations for an insurer’s written AI program, governance, risk management and third-party oversight. It is model guidance for jurisdictions, rather than a nationwide AI statute. State requirements and the insurance activity determine the applicable obligations.

In the EU, EIOPA’s [AI governance opinion](https://www.eiopa.europa.eu/eiopa-publishes-opinion-ai-governance-and-risk-management-2025-08-06_en) clarifies existing insurance-sector expectations using a risk-based, proportionate approach. It addresses areas including data governance, fairness, record-keeping and human oversight. The opinion excludes AI systems classified as high-risk or prohibited under the AI Act; EIOPA identifies risk assessment and pricing for individuals in life and health insurance as high-risk uses under that Act. Other insurance uses remain subject to applicable sectoral legislation.

These distinctions belong in the use-case record. The insurer’s legal and compliance owners map the relevant line of business, jurisdiction and decision to its requirements. A fraud-screening service, a pricing model and a customer chatbot should not inherit each other’s regulatory treatment simply because all use AI.

## Evidence of value should match the job

The opportunity is practical: less document search, earlier attention to missing information, more consistent handling and better-supported decisions. Whether a deployment delivers that value depends on the task and its downstream effects.

Useful acceptance evidence differs by use:

Evidence of value should match the job

| Use | Evidence of fit | Failure the headline result can hide |
| --- | --- | --- |
| Underwriting or pricing | Validated predictions, relevant explanations and analysis of affected groups | Better aggregate predictions alongside poor treatment of a particular segment |
| Service assistance | Source-grounded answers, appropriate escalation and correct handling of account access | Fast replies that misstate a policy or expose unrelated information |
| Claims summaries | Accurate attribution, material contradictions retained and usable source references | A shorter summary that removes the uncertainty the adjuster needs |
| Damage assessment | Suitability for the asset, damage type and images supplied | A plausible estimate for damage the available views cannot establish |
| Authenticity review | Relevant findings, visible analysis limits and useful reviewer follow-up | More alerts without better evidence, or unavailable analysis counted as a pass |

Those are different evaluation jobs. A single accuracy figure cannot describe them all. Time saved at intake also says little about time spent resolving exceptions or correcting an unsupported decision.

A bounded deployment gives the insurer a clearer basis for adoption: one defined task, a receiving owner and a record of what changed. Broader authority follows evidence that the intended controls work for that use, including when the inputs are incomplete or the system cannot produce a usable result.

## Questions about insurance AI

### Is ordinary workflow automation the same as AI?

An automated process can follow predefined rules without an AI model. A rule that sends an incomplete claim to an intake queue is different from a model that interprets a free-text loss description. The insurer owns the rule or recommendation and the resulting action in both cases.

### Can AI authenticity checks deny a claim?

No. Trustable’s checks do not deny claims. They supply findings that can be wrong or incomplete, so a result must never be the sole basis for a denial. The adjuster or SIU reviews the concern alongside the policy, loss circumstances and other evidence. An authorized person at the insurer remains accountable for the claim decision.

### Can one claims model serve every insurance line?

Suitability does not transfer automatically. Vehicle photos, property estimates and clinical records contain different evidence and support different decisions. A motor damage-assessment deployment does not establish fitness for health-claim adjudication. The system’s approved scope needs to name the insurance line, the input types and the decision it supports.

## How Trustable helps

Our document and media authenticity findings support accountable claims review. We flag signs of editing or AI generation in submitted files so the adjuster or SIU can assess the concern alongside the rest of the evidence. We do not price policies or decide claims.

[Book a demo](https://cal.com/scamai/20min?utm_source=trustable.ai&utm_medium=website&utm_campaign=book_demo&utm_content=insight-artificial-intelligence-in-insurance-evidence-and-limits-body) to see our findings on a sample claim file and discuss how your SIU could use them to focus further evidence collection.

Trustable

## See it on a sample file.

Book a 20-minute demo. We agree on data handling before you share anything.

[Book a demo](https://cal.com/scamai/20min?utm_source=trustable.ai&utm_medium=website&utm_campaign=book_demo&utm_content=insight-artificial-intelligence-in-insurance-evidence-and-limits)