Research on AI in auto insurance abroad has matured along three principal lines: risk assessment and pricing, loss assessment and claims, and fraud detection.

On risk assessment and pricing, Lee and Antonio (2015) [2] compared neural networks, generalised linear models, generalised additive models, and decision trees in predicting claim frequency and reported that neural networks performed best in predictive accuracy. Blier-Wong et al. (2020) [3] provided a systematic review of machine learning in property and casualty insurance pricing and reserving, highlighting the integration of unstructured data such as images and geospatial information as a key future direction. Hanafy and Ming (2021) [4] compared logistic regression, XGBoost, Random Forest, decision trees, naïve Bayes, and k-nearest neighbours, and demonstrated the strength of ensemble methods on auto insurance big data.

On loss assessment and claims, Singh et al. (2019) [5] developed an end-to-end loss-assessment system that automates damage estimation and claim amount calculation from images of damaged vehicles. Zhang et al. (2023) [6] introduced AI-assisted decision-making to facilitate dispute resolution in motor vehicle accident claims through rigorous cost estimation procedures.

On fraud detection and prevention, Aslam et al. (2022) [7] constructed an anti-fraud framework combining Boruta-based feature selection with logistic regression, support vector machines, and naïve Bayes; Maiano et al. (2023) [8] deployed deep-learning image recognition to verify the authenticity of claim images, contributing a multimodal pipeline for fraud detection.

Domestic research has tracked international frontiers but with comparatively few empirical studies. On pricing, Meng et al. (2017) [9] demonstrated the effectiveness of support vector machines, gradient boosting trees, and multilayer perceptrons in predicting claim probabilities and aggregate payouts in auto insurance. Yan et al. (2022) [10] showed empirically that a GA-XGBoost algorithm outperforms SVM and CART in evaluating UBI driver-behaviour risk.

On loss assessment and claims, Zhou et al. (2023) [11] examined the application of insurance technology in claim investigation and loss assessment, proposing a workflow that combines remote guidance with image collection.

On fraud detection, Ye (2011) [12] introduced a hybrid approach combining statistical regression with neural networks for auto-insurance fraud identification; Chen and Li (2022) [13] compared six machine-learning methods and found that gradient-boosting decision trees and Random Forest deliver the best combination of predictive performance and robustness.

Synthesising the existing literature suggests three important gaps. First, both international and domestic research has tended to compare algorithms within a single workflow segment, leaving the cross-segment, system-level mechanisms by which AI empowers the auto insurance value chain under-investigated. Second, domestic research is dominated by theoretical analysis and policy recommendations, with relatively limited empirical work; international algorithmic optimisation has, on aggregate, advanced ahead of domestic studies. Third, empirical examination of AI-enabled auto insurance services from the consumer-perception angle remains particularly scarce, leaving the industry’s assessment of AI value largely tied to supply-side indicators with insufficient attention to demand-side experience. By integrating literature analysis with 260 consumer questionnaires, this paper engages both supply and demand perspectives and aims to address these gaps.

Traditional auto insurance pricing rests on historical actuarial data and standard statistical models, which are limited in capturing individual-level behavioural variation and dynamic market conditions. AI—particularly machine learning and deep learning—processes high-volume, high-dimensional data to extract key risk factors such as driving behaviour, vehicle condition, geographic location, and weather conditions, supporting a refined transition from “vehicle-based” to “driver-based” pricing. Usage-based insurance (UBI) models, which adjust premiums dynamically using telematics data, offer a leading example of this paradigm shift. Privacy-preserving computing and federated learning further enable insurers to undertake joint modelling with automakers and third-party platforms without directly exchanging raw data, expanding the data base on which pricing models can draw.

Vehicle damage assessment is the core link in auto insurance claims, directly determining the indemnity amount. Computer vision and deep learning algorithms can process accident-scene photographs or video footage to automatically identify damage location, classify damage type, and assess severity, thereby producing structured repair plans and indemnity recommendations. With pre-trained image recognition models and repair knowledge bases, insurers can complete loss assessment for the great majority of low-amount cases within minutes—an order-of-magnitude improvement over the multi-day cycles typical of traditional workflows. The objectivity of AI recognition also helps to mitigate the information asymmetry and bargaining disadvantage that insurers traditionally face vis-à-vis repair shops and the insured.

Conventional auto-insurance claims involve multiple steps—reporting, inspection, loss assessment, review, payment, and case closure—which are time-consuming, costly, and prone to subjective judgement. AI-driven smart claims systems use NLP to extract structured information from incident reports, computer vision to identify photographic evidence automatically, and workflow automation to connect processing nodes, compressing claims-handling cycle times substantially; for low-amount cases, indemnity payment can be completed within minutes. Blockchain technology contributes tamper-resistant evidence storage, further enhancing process transparency and customer trust.

Auto insurance fraud is increasingly professionalised, organised, and concealed, and is poorly addressed by traditional manual review and rule-based heuristics. Machine learning models trained on historical claim data identify abnormal claim patterns automatically; combined with NLP for semantic analysis of incident narratives, computer vision for verifying scene photographs, and graph neural networks for detecting collusion among related parties, AI supports a multimodal anti-fraud architecture. Voice-based emotion recognition and biometric identification further enable insurers to detect anomalous emotional states or identity risks at the point of incident reporting, shifting the anti-fraud frontier upstream. AI-driven real-time monitoring and dynamic risk assessment likewise enable insurers to intervene at the early stages of risk accumulation.

Table 1 summarises the four nodes alongside their AI technology families and underlying mechanisms.

Table 1. Key nodes of the AI-enabled auto insurance value chain.

To examine the practical effectiveness of AI-enabled auto insurance services from the consumer-perception perspective, this study used an online questionnaire administered mainly through the Credamo questionnaire platform, supplemented by dissemination through WeChat groups, WeChat Moments, and online car-owner/insurance-related communities. The target respondents were adult drivers or potential auto-insurance consumers who were able to understand and evaluate AI-enabled auto-insurance services. The questionnaire did not impose a regional restriction and did not collect a separate province/city variable; therefore, the results should be interpreted as online-sample evidence rather than geographically representative national estimates. The instrument contains 18 items covering demographic information, driving experience, prior AI-enabled auto-insurance experience, AI claims, AI loss assessment, AI risk assessment, process-optimisation expectations, and open-ended personalisation needs. The main questionnaire items and response options are presented in the Appendix. Respondents who reported no prior AI-enabled auto-insurance experience were routed directly to expectation-oriented items, thereby separating service-experience evaluation from general demand-side expectations.

Table 2.Sample characteristics of valid respondents.

Note: Percentages are calculated using the 260 retained individual responses as the denominator.

A total of 260 valid individual responses were used for the empirical analysis. The sample characteristics are summarised in Table 2. Among the 260 retained respondents, 171 were female (65.77%) and 89 were male (34.23%). Respondents were concentrated in the 18 - 25 and 26 - 35 age brackets, and all retained individual responses reported some driving experience. A total of 165 respondents (63.46%) reported having previously used AI-enabled auto-insurance services, while 95 respondents (36.54%) reported no prior use. Since the questionnaire did not collect IP addresses or province/city information, the data processing did not involve geographic stratification. Data cleaning included: deleting non-individual summary rows, checking completeness of mandatory demographic and screening questions, and applying questionnaire skip logic when defining the analytical sample for each module.

Data analysis combines frequency analysis, mean Likert-score reporting, cross-tabulation, and subgroup comparison. In response to the questionnaire’s skip logic, the 95 respondents without prior AI-enabled auto-insurance experience are excluded from usage-specific satisfaction analyses but retained for expectation-oriented items. Consequently, the valid base differs across modules: claim-cycle and claim-efficiency results are calculated among 160 valid claim-experience responses; loss-assessment clarity and transparency are calculated among 163 valid loss-assessment responses; risk-assessment indicators are calculated among 165 experienced users; and process-optimisation expectations are calculated on the full sample of 260 respondents. Likert items were coded so that larger values indicate more positive evaluations. Five-point satisfaction items were coded from 1 = very dissatisfied to 5 = very satisfied; four-point clarity, accuracy, and personalisation items were coded from 1 to 4 in ascending order of positive evaluation. Cronbach’s alpha was calculated only for reflective Likert-type blocks with at least two items: the loss-assessment perception block (Q8 and Q9, α = 0.596) and the risk-assessment perception block (Q11 - Q13, α = 0.637). The claim-efficiency block contains a single satisfaction item and the process-optimisation block consists of multiple-response expectation items; therefore, Cronbach’s alpha is not treated as applicable for these two formative or single-item blocks. Mean Likert scores and reliability information for the core survey indicators are reported in Table 3. Basic subgroup comparisons were conducted using Welch’s t-test for prior-use status and Kruskal-Wallis tests for age and driving-experience groups, with the results presented in Table 4. The correspondence between questionnaire items and analytical modules is shown in the Appendix.

Table 3.Mean likert scores and reliability information for core survey indicators.

Note: Cronbach’s alpha was calculated for reflective multi-item perception blocks. Loss-assessment perception (Q8 - Q9): α = 0.596; risk-assessment perception (Q11 - Q13): α = 0.637. Cronbach’s alpha is not applicable to the single-item claim-efficiency satisfaction indicator or the multiple-response process-optimisation expectation block. Given the exploratory nature of this consumer-perception survey and the limited number of items in some modules, these reliability results should be interpreted with caution. Future studies may improve scale reliability by expanding the number of items in each construct.

Table 4.Basic subgroup comparisons of composite scores.

Note: Prior-use status is tested only for the process-optimisation demand score because respondents without prior AI-enabled auto-insurance experience skipped usage-specific satisfaction modules. Age and driving-experience comparisons use Kruskal-Wallis tests; prior-use comparison uses Welch’s t-test.

The claim-related analysis is based on respondents who had relevant AI-claim experience. Among the 160 valid claim-cycle responses, 115 respondents reported that the process from claim submission to indemnity receipt took less than one week, accounting for 71.88%. Claim-efficiency satisfaction is also positive: 122 of the 160 valid respondents selected “satisfied” or “very satisfied”, yielding a positive share of 76.25% and a mean score of 3.89 on a five-point scale. However, pain points remain concentrated in front-end document preparation and process interaction. Specifically, 62.50% of valid claim respondents selected excessive documentation requirements, 48.75% selected communication frictions with insurers, 48.13% selected complex application procedures, and 42.50% selected long review time. These results indicate that AI may have improved back-end processing speed, but the consumer-facing interface of documentation submission, guidance, and communication remains an important bottleneck.

For AI smart loss assessment, 147 of the 163 valid respondents regarded the loss-assessment information as clear or basically clear, corresponding to a positive clarity share of 90.18% and a mean clarity score of 3.08 on a four-point scale. Transparency satisfaction is also favourable: 127 of the 163 valid respondents were satisfied or very satisfied, with a mean transparency score of 3.88 on a five-point scale. At the same time, the two most frequently identified shortcomings were the lack of visualisation and unclear dispute-resolution mechanisms, each selected by 60.12% of valid loss-assessment respondents, followed by insufficient detail in loss-assessment reports (42.94%). The results suggest that AI image recognition has reached a relatively high level of perceived usability at the identification stage, but further improvement is needed in explainable presentation and post-assessment dispute handling.

Risk-assessment evaluation is based on 165 respondents with prior AI-enabled auto-insurance experience. In total, 138 respondents regarded AI risk assessment as accurate or comparatively accurate, producing a positive share of 83.64% and a mean score of 2.92 on a four-point scale. In addition, 155 respondents reported receiving personalised suggestions or services at least occasionally, corresponding to 93.94% and a mean score of 3.16 on a four-point scale. Satisfaction with risk-assessment functions is likewise positive: 135 respondents were satisfied or very satisfied, accounting for 81.82%, with a mean score of 3.96 on a five-point scale. Nevertheless, 75.15% of experienced users identified insufficient deep personalisation as a remaining problem, 55.76% reported that risk factors considered by the system were not comprehensive enough, and 44.24% noted untimely updates. These findings imply that consumers recognise the value of AI risk assessment, while expecting richer individual-level modelling and more dynamic feedback.

The expectation-oriented items were answered by the full sample of 260 respondents, including both experienced users and non-users. In claims handling, the most frequently selected improvement expectations were faster reporting response (73.08%), online claim guidance (71.92%), accelerated review (67.69%), and automated liability identification (65.38%). In loss assessment, respondents placed the strongest emphasis on more accurate results (81.54%) and real-time assessment at the accident scene (79.62%), followed by support for multiple accident scenarios (66.54%). In risk assessment, improving the timeliness of risk updates was the most salient expectation (81.15%), followed by including more comprehensive risk factors (74.23%) and providing visual and understandable risk reports (71.15%). Subgroup comparisons further show that non-users selected more process-optimisation expectations than experienced users on average (8.14 versus 6.85 selected items out of 10; Welch’s t-test p < 0.001), suggesting that consumers without prior experience may hold stronger expectations for AI-enabled service improvement.

Table 5.Consumer evaluation of AI-enabled auto insurance and key pain points.

Table 5 summarises consumer evaluations of AI-enabled auto-insurance services, the key pain points identified, and the corresponding improvement expectations across the three service dimensions. In aggregate, the results provide demand-side evidence that consumers perceive AI-enabled auto-insurance services positively in terms of efficiency, clarity, accuracy, and personalised recommendation. However, these findings are consumer-perception evidence rather than direct causal evidence from insurer-side operational performance data. Complex operational procedures, cumbersome documentation, insufficient visualisation, weak deep personalisation, and limited timeliness of risk updates remain the principal demand-side concerns and should figure prominently in subsequent technological iterations.

AI applications in auto insurance depend heavily on multi-source data—telematics, user behaviour, and third-party data—but data formats across automakers, repair shops, and insurers remain non-uniform and cross-platform sharing arrangements are underdeveloped. As parties to the claim, the insured may also answer selectively or omit relevant information during inquiries, further constraining the availability of reliable raw data. The resulting “data silos” limit both the training-sample scale and the cross-scenario generalisation of AI models, representing a foundational bottleneck that the industry must collectively address.

Vehicle damage assessment is intrinsically heterogeneous: identical damage to the same vehicle model can yield substantially different repair cost estimates depending on location; the same repair shop may price parts and labour differently across vehicle models; and even within identical vehicle-shop pairings, differences in collision angle and force lead to divergent damage profiles. Final pricing typically emerges from negotiation between the repair shop and the insured, with reference to local market rates, leaving the insurer in a comparatively passive position. AI image recognition can improve the objectivity of damage classification, but moving from “assisted recognition” to “unified pricing” requires industry-level standards and underlying data infrastructure beyond the reach of any individual insurer.

New energy vehicles increasingly feature a “vehicle-battery separation” model, but most existing dedicated products bundle the battery within whole-vehicle insurance, leaving stand-alone coverage for batteries or non-battery components underdeveloped. For intelligent connected vehicles, the industry has yet to develop dedicated terms and products and continues to rely on traditional compulsory motor liability and vehicle-damage products. The risk types, liable parties, and rate-making considerations differ markedly across automation levels, and emerging risks such as cybersecurity, data privacy, and safety of intended functionality have no clear counterpart in legacy products. At the same time, high loss ratios for new energy vehicles and high per-incident repair costs for intelligent connected vehicles create a pricing dilemma in which falling accident frequency coexists with rising repair costs, complicating the application of traditional actuarial models.

AI-enabled auto insurance relies heavily on the deep utilisation of personal data, but the Personal Information Protection Law and related regulations have raised the bar for data collection, storage, and use, requiring insurers to balance compliance with modelling needs. Internal resistance is also non-trivial: traditional business units may view AI deployment as a substitution threat, with underwriters and field investigators displaying limited adaptability to new technologies. Organisational change therefore constitutes a non-technical but consequential constraint.

Table 6 summarises the four challenges alongside potential AI-enabled responses.

Table 6.Key challenges in “AI + Auto Insurance” and potential pathways.