Insurance as AI Risk Infrastructure: A Generative-Agent Simulation of AI Adoption

Yixuan Yuan; Dedai Wei; Chudong Qian; Jielin Feng; Ziyue Lin; Yuheng Zhao; He Cao; Erasmo Purificato

cs.MAAug 15, 2026

Insurance as AI Risk Infrastructure: A Generative-Agent Simulation of AI Adoption

Authors: Yixuan Yuan, Dedai Wei, Chudong Qian, Jielin Feng, Ziyue Lin, Yuheng Zhao, He Cao, Erasmo Purificato,

Organizations: University of Macau · Fudan University · Immunocan AI · European Commission ·

Abstract

The rapid evolution of artificial intelligence (AI) tools has demonstrated immense potential to enhance societal well-being and operational efficiency. However, the inherent unreliability and uncertain operational consequences of modern AI systems, typified by large language models (LLMs), have created a significant barrier to enterprise adoption. Many enterprises remain hesitant to integrate these tools deeply into their workflows due to concerns about unpredictable losses and liability exposure. While existing technical safeguards primarily seek to reduce the likelihood or severity of AI-enabled workflow failures, they do not by themselves provide ex post financial protection when residual pecuniary tail losses materialize. In this paper, we introduce a socio-economic framework that complements these safeguards by transferring and absorbing the residual financial consequences of AI adoption through insurance. To evaluate this framework, we develop an LLM-driven agent-based social simulation (LABSS) system. We assess the behavioral validity of the simulation using established economic and sociological theories. Our analysis demonstrates that the proposed insurance framework reduces firm-level financial exposure, thereby accelerating the aggregate adoption of AI tools and improving firm solvency and aggregate capital.

Figures & tables

Figure 1: Motivation. (a) The aggregate social benefit of AI adoption decomposes into heterogeneous firm-level value trajectories, with some firms bearing concentrated tail losses. (b) Existing safeguards reduce failure likelihood but cannot compensate realized losses. (c) Insurance pools premiums to compensate affected firms.

Figure 2: Overview. The system integrates macro-environmental uncertainty with micro-cognitive decision-making, orchestrating a lifecycle from signal processing to bilateral negotiation.

Figure 3: The Macro-Environmental Interaction Cycle. This diagram illustrates the cyclical flow of information and capital between the environment, AI vendors, insurers, and buyer firms, governing the transition from signal emission to contract realization.

Figure 4: The AAD Framework. The micro-level cognitive process where firms filter signals (Attention), evaluate risks (Assessment), and execute strategies (Decision).

Item	Test	Statistic	Null hypothesis	Result	Decision
M1. Bounded attention	Menu-conformance audit	$\bar{V}_{s}$ , $\pi^{<4}_{s}$	$H_{0}:\bar{V}_{s}=4$ , $\pi^{<4}_{s}=0$	$\bar{V}_{s}=3$ ; $\pi^{<4}_{s}=1.00$	Reject $H_{0}$
M2. Peer diffusion	Peer-response regression	$\hat{\beta}$	$H_{0}:\beta\leq 0$	$0.0134,0.0173,0.0267$ ; all $p<0.001$	Reject $H_{0}$
M3. Contract inertia	Renewal-gate and retention audit	$\pi^{expiry}_{s}$ , $\rho^{inc}_{s}$	$H_{0}$ : no expiry gate or incumbent retention	$\pi^{expiry}_{s}=1.00$ ; $\rho^{inc}_{s}=0.9368$ – $0.9588$	Reject $H_{0}$
M4. Heterogeneity	Heterogeneity F-test	$F$ , partial $R^{2}$	$H_{0}$ : traits and industry add no fit	$F=656.96$ ; $p<10^{-300}$ ; partial $R^{2}=0.204$	Reject $H_{0}$

Table 1: Compact mechanism diagnostics before path-level validation.

Figure 5: Sectoral network diagnostic. Bubble correlation matrix of industry-level mean panic paths over the post-diffusion window. The upper triangle reports insurance, and the lower triangle reports no insurance. Circle area encodes $|r|$ , and color encodes the sign.

Figure 6: Market path composition in the formal insurance arm. Panels (a)–(b) show active AI use by vendor and GICS industry, and panels (c)–(d) show insurance coverage by insurer and GICS industry. Dark lines denote the corresponding aggregate levels.

Item	Test	Statistic	Null hypothesis	Result	Decision
H1a. Active stock	Trajectory model comparison	$R^{2}$ , $\Delta AIC$	$H_{0}:\Delta AIC\leq 0$	$3$ paths $\times 100$ updates; mean $R^{2}=0.996$ , $\Delta AIC=476.7$	Reject $H_{0}$
H1b. Entry S-curve	Trajectory model comparison	$R^{2}$ , $\Delta AIC$ , peak	$H_{0}:\Delta AIC\leq 0$	$3$ paths $\times 100$ updates; mean $R^{2}=0.998$ , $\Delta AIC=532.1$ ; peak days $57$ – $69$	Reject $H_{0}$
H1c. Risk-transfer lead	Paired seed-endpoint contrast	$\bar{d}_{T}$ , $AUC_{s}$ , $t_{d}$	$H_{0}:\mathbb{E}[d_{s,T}]\leq 0$	$3$ pairs; $\bar{d}_{T}=+9.41$ pp; $AUC_{s}>0$ $(3/3)$ ; $p=0.025$	Reject $H_{0}$
H1d. Digital timing	Within-seed trajectory-label permutation	$\Delta^{\tau}$ , $p_{\mathrm{perm}}$	$H_{0}:\Delta^{\tau}\leq 0$	$300$ firms/seed; $\Delta^{\tau}=9.79$ updates; $p_{\mathrm{perm}}=0.001$	Reject $H_{0}$
H1e. AI size order	Within-seed trajectory-label permutation	$\pi^{AI}$ , $p_{\mathrm{perm}}$	$H_{0}:\pi^{AI}\leq 1/6$	$300$ firms/seed; $\pi^{AI}=0.614$ ; $p_{\mathrm{perm}}=0.001$	Reject $H_{0}$
H2a. Product diffusion	Coverage trajectory fit	Coverage, $R^{2}$	$H_{0}$ : no product-diffusion structure	$3$ paths $\times 100$ updates; active $R^{2}=0.994$ , entry $R^{2}=0.998$	Reject $H_{0}$

Table 2: Compact path-level validation of intermediate market behavior.

Figure 7: Bankruptcy dynamics. Cumulative bankruptcy counts by operational cycle in the paired counterfactual. Faint paths show seed-level integer trajectories; bold paths show the three-seed mean.

Figure 8: Capital decomposition. The left panel tracks the social total capital index relative to the no-insurance arm and the cumulative reduction in unabsorbed claimable loss. The right panels compare the final capital index and cumulative unabsorbed claimable loss across arms.

Figure 9: Vendor contract terms. Distribution of term lengths for successfully bound AI vendor contracts, measured in operational updates.

Figure 10: Diffusion and industry convergence. Aggregate AI adoption and the cross-industry adoption IQR, defined at each three-day observation as the difference between the 75th and 25th percentiles of adoption rates across the 11 GICS industries. Lower IQR indicates stronger convergence.

Figure 11: Vendor allocation. Final vendor market share, realized tail risk, and average task error in the insurance arm.

Figure 12: Sector-specific vendor composition. Final AI vendor composition by industry in the insurance-on arm. The dark grey ring denotes the uninsured share.

Figure 13: Insurance risk-pool ledger. Cumulative premiums and cumulative paid claims are reported in millions of dollars. The shaded band denotes the premium-funded buffer between the two cumulative flows, and the vertical bars report paid claims per operational update.

Figure 14: Panic tail and aggregate burden. The upper panel shows the 95th-percentile panic trajectory and the lower panel the cumulative mean panic burden, both in

10^{-3}

units; shading denotes the no-insurance excess.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter block	Exact formal setting
Population and horizon	300 buyer firms, comprising 197 small, 69 medium, and 34 large firms; 11 sectors; a 300-calendar-day horizon with one operational update every three days; paired seeds 42, 77, and 202; four AI vendors; four private insurers and one residual backstop.
Sector allocation	Communication services 28; consumer discretionary 36; consumer staples 26; energy 19; financials 32; health care 24; industrials 31; information technology 37; materials 19; real estate 24; utilities 24.
Firm cash and returns	Initial cash $=0.30A_{i}$ ; traditional return mean $=-0.00008$ , standard deviation $=0.0035$ , lower cap $=-0.025$ , and upper cap $=0.010$ ; AI gain scale $=0.078$ ; AI gain risk drag $=0.70$ .
Peer-information network	12 same-sector and 4 cross-sector neighbors, minimum degree 10, and recent-claim window 7 operational updates.
Vendor attention	The visible set contains at most three vendors. One slot is sampled uniformly from the full vendor pool, and the remaining slots are sampled without replacement using the Phase 3 visibility probability. Sector affinity equals 1.00 for a targeted sector, 0.75 for an all-sector vendor, and 0.35 otherwise; baseline visibility is $b=0.50$ . A numerical weight floor of 0.01 is applied in the implementation and is nonbinding under the formal vendor profiles.
Event and loss mapping	Risk-score floor $=0.02$ ; loss realization $r=0.70$ ; action-loss scale $s=3.35$ ; catastrophic threshold $=0.25$ ; tail multiplier coefficient $=12.0$ ; material-loss ratio $=0.0015$ ; claimable-loss ratio $=0.0030$ ; minimum claimable-loss ratio $=0.0012$ ; material-event score threshold $=0.46$ ; claimable-event score threshold $=0.54$ .

Appendix

Table 3: Formal runtime and state-transition parameters. The simulation advances once every three calendar days. Values are shared by the insurance-on and insurance-off arms unless the insurance market itself is disabled.

Vendor	Fee	Productivity lift	Risk multiplier	Reputation	Marketing weight	Target sectors
Alpha	2800	0.0110	0.62	0.78	1.05	Financials, energy, and health care
Beta	2200	0.0130	0.82	0.68	0.95	Information technology, communication services, and consumer discretionary
Gamma	1600	0.0125	1.10	0.63	0.95	Industrials, materials, and utilities
Delta	1200	0.0120	1.05	0.62	1.25	All sectors

Appendix

Table 4: AI vendor profiles used in the formal experiment. Fees are scaled against a 30-update contract base.

Insurer archetype	$K^{0}$	$b$	$a$	$e$	$\ell$	$\delta^{0}$	$q^{0}$	$\rho$	Capacity	Floor/soft/hard	Target sectors and role
Diversified commercial	6.5	0.38	0.62	0.045	0.85	0.34	0.64	0.11	115	0.22/0.68/0.42	All sectors; private
Mutual commercial	4.2	0.30	0.58	0.028	0.90	0.30	0.59	0.10	45	0.24/0.72/0.46	Consumer discretionary, consumer staples, real estate, and utilities; private
Digital commercial	3.8	0.32	0.66	0.030	0.85	0.30	0.56	0.08	75	0.23/0.70/0.44	Information technology, communication services, and financials; private
Specialty technology	3.0	0.45	0.52	0.050	0.95	0.34	0.74	0.15	45	0.26/0.76/0.50	Information technology, health care, financials, and communication services; private
Residual backstop	9.0	0.60	0.35	0.025	1.20	0.50	0.55	0.20	$10^{6}$	0.30/0.80/0.58	All sectors; residual facility

Appendix

Table 5: Insurer profiles used in the formal experiment.

K^{0}

is initial capital in millions;

b,a,e,\ell

denote base margin, risk appetite, expense load, and capital load;

\delta^{0},q^{0},\rho

denote base deductible, coverage, and limit ratios. Capacity is the concurrent-policy limit, and floor/soft/hard values are capital ratios relative to

K^{0}

AutoCLAW evidence	Enterprise balance-sheet channel	Simulation variable	Insurance channel
Failed report generation, rewriting, or constrained summarization	Rework cost, decision delay, client-facing quality loss, and compliance exposure	Incident flag, severity, direct loss, total loss, and risk score	Eligible loss enters deductible, coinsurance, and limit schedule if the incident score exceeds the policy threshold
Directory audit or file-cleanup error	Lost files, restoration labor, workflow interruption, and access-control remediation	Direct loss and tail multiplier when the maximum risk score crosses the tail threshold	Claim payment is capped by policy limit and insurer solvency floor
Record update, CSV update, or email drafting failure	Data inconsistency, erroneous communication, operational correction cost, and reputational pressure	Material event signal, claimable loss signal, panic input, and memory update	Paid claims reduce liquidity trauma, while uncovered claimable losses strengthen loss and claimable-loss memory
Clustered high-score task failures within a firm-cycle	Short-run liquidity pressure and bankruptcy risk when cash buffers are thin	Realized operational loss $\ell_{it}$ , cash update, bankruptcy flag, and social total capital	Insurance changes timing and allocation of the loss, not the occurrence of the underlying AI failure

Appendix

Table 6: Threat model from AutoCLAW evidence to balance-sheet channels. The mapping is fixed before the insurance-on and insurance-off counterfactuals are compared.

Mechanism	Baseline	Sensitivity range	Role in the simulation
AI productivity scale	$0.078$	$0.070$ to $0.086$	Converts active AI use into productivity gain, common to both insurance arms
AI operational-loss scale $s$	$3.35$	$3.00$ to $3.70$	Converts AutoCLAW task loss into realized firm balance-sheet loss
Loss realization rate $r$	$0.70$	Fixed in the reported sweep	Keeps the task-to-cash conversion common across paired counterfactuals
Insurance pricing multiplier $G$	$2.95$	$2.65$ to $3.25$	Loads expected loss, stress loss, expenses, and capital costs into premiums
Coverage and deductible shifts	$\Delta_{q}=0.16$ , $\Delta_{\delta}=-0.12$	$\Delta_{q}=0.10$ to $0.20$ , $\Delta_{\delta}=-0.08$ to $-0.14$	Controls how much tail loss is retained by firms versus indemnified by insurers
Policy-limit multiplier $\lambda_{M}$	$2.20$	Fixed in the reported sweep	Scales insurer-specific policy limits

Appendix

Table 7: Key calibration values and one-factor sensitivity ranges. Ranges correspond to the reported one-factor sensitivity profiles used in Appendix E .

Figure 15: Insurance payoff structure. The insurer indemnifies a coverage ratio

q

of AI operational losses above deductible

D

up to policy limit

M

. The firm retains the deductible layer, coinsurance, and any excess loss beyond the covered layer.

Component	Recorded specification
Model checkpoint	Qwen/Qwen3-8B snapshot b968826d9c46dd6066d109eabc6255188de91218 , served as qwen3-agent-local . The tokenizer and native string-format chat template are loaded from the same snapshot; no custom chat template is supplied.
Serving stack	vLLM 0.10.2, Transformers 4.55.2, Tokenizers 0.21.4, PyTorch 2.8.0 with CUDA 12.8, and Python 3.11.15. Eight independent single-GPU replicas run on NVIDIA H100 80GB devices with BF16 weights, no quantization, tensor parallel size one, maximum context length 8192, GPU-memory utilization 0.82, and a maximum of 16 concurrent sequences per replica.
Decoding	Temperature 0.0; top-p, top-k, and a per-request model seed are not supplied by the client; the vLLM engine seed is 0. Maximum response length is 220 tokens for decisions and negotiations; request timeout is 30 seconds; response format is a JSON object; reasoning is disabled with the /no_think suffix.
Paired routing	Seeds 42, 77, and 202 pair firm initialization, AutoCLAW records, random exposure, and endpoint assignment. A stable hash of the seed and call identity, including decision or negotiation type, firm, cycle, round, and side when applicable, and repair attempt, selects one of the eight workers.
JSON parser	The parser first accepts a complete JSON object and then attempts extraction from the first opening to the last closing brace. Failed extraction triggers at most two repair prompts. Enumerated actions, vendor identifiers, policy tiers, integer terms, and negotiation prices are subsequently validated or bounded by the simulator.
Failure policy	Runtime and transport fallback are disabled in the formal configuration. If all JSON repair attempts fail, the implementation permits a deterministic rule response for that recoverable parse error and records the reason.

Appendix

Table 8: LLM decision configuration and audit settings. Settings are shared by the paired insurance-on and insurance-off runs.

Call	Eligibility and additional state	Required output	Simulator enforcement
AI adoption	Active non-user outside cooldown; up to three visible vendors; procurement-stage and local-evidence state	adopt_ai , adoption_score , reason, selected vendor, vendor term, maximum rounds	Dynamic threshold clipped to $[0.55,0.935]$ ; score must also clear a 0.026 margin; selected vendor must be visible; term is bounded to 14–120 operational updates. For first adoption, a proposed term above 60 updates requires a score margin of at least 0.08.
Vendor renewal	Vendor contract expired at the current operational update; incumbent identifier and all visible alternatives are supplied	Same schema as adoption	Expiry opens a neutral review window. Dynamic renewal threshold is clipped to $[0.40,0.68]$ , with a 0.045 margin. The incumbent has no prompt-imposed priority.
Insurance purchase or renewal	Active AI exposure and no active policy; prior-policy flag, remaining AI term, risk signal, and risk-need reference	buy_insurance , insurance_score , reason, policy term, maximum rounds	Purchase threshold 0.39; prior-policy discount 0.09; lower bound 0.27. The formal decision menu bounds the policy term to 1–90 operational updates and the remaining AI exposure. Quote production and actuarial pricing remain rule-based.
Exposure management	Active AI exposure; lagged risk signal, memories, panic, contract state, and allowed actions	abandon_ai , abandon_score , reason, and vendor_action	Threshold 0.62. Score and action are made consistent after parsing; the only valid actions are keep_vendor and abandon_ai .
Vendor bargaining	Triggered after adoption or renewal; firm and vendor states, current ask, floor, prior offers, term, side, and round limits	Decision, monthly fee, and message	Seller actions are accept/counter/reject; buyer actions are accept/counter/quit. Prices and rounds are bounded by the configured floor, affordability, minimum rounds, and 30-round cap.
Insurance bargaining	Triggered after purchase intent and quote selection; firm, selected quote, policy menu, risk need, offers, side, and round limits	Decision, premium, message, and optional valid tier	The same action sets apply. Tier must belong to the generated policy menu; premium and rounds are bound before a contract can be formed.

Appendix

Table 9: Decision-call eligibility, outputs, and deterministic enforcement.

Control block	Effective values
Adoption threshold	Base 0.665; offset 0.043; clamp $[0.55,0.935]$ ; maturity index 106 operational updates; early friction 0.147; local-evidence floor/reference $=0.075/0.47$ ; evidence lag 48 updates; low-evidence friction 0.078; implementation uncertainty 0.064; local-evidence relief 0.092 with power 1.5; pioneer discount 0.145; inertia penalty 0.056; urgency, innovativeness, and dependence discounts $=(0.047,0.035,0.035)$ ; panic penalty 0.055; loss and claimable-memory penalties $=(0.08,0.07)$ ; insurance-confidence discount 0.01; required margin 0.026.
Risk-transfer evidence	Evidence gate 0.20; insurance-availability floor 0.15; local insured-peer weight 0.70; recent paid-claim weight 0.20; paid-claim reference rate 0.015; direct panic dampening of the local-adoption signal 0.00. These values modify the credibility of peer adoption evidence rather than directly forcing adoption.
Renewal threshold	Base 0.46; offset 0.05; clamp $[0.40,0.68]$ ; loss-memory, claimable-memory, and panic penalties $=(0.08,0.06,0.06)$ ; insurance-confidence discount 0.03; continuity term $=0.04\times$ inertia; required margin 0.045.
Insurance intent and reference	Model purchase threshold 0.39; prior-policy threshold discount 0.09; minimum threshold 0.27. The state reference and deterministic parse-fallback score has base 0.06, prior-policy bonus 0.24, and weights for claimable event, material event, risk memory, claimable memory, industry incident, industry stress, industry loss, panic, recent claims, peer coverage, risk aversion, and AI dependence of $0.28,0.18,0.16,0.12,0.08,0.10,0.06,0.10,0.15,0.03,0.14,0.11$ , with innovativeness discount 0.07. Fallback term thresholds are 0.70 and 0.38; dependence and risk-tolerance thresholds are 0.62 and 0.50.
Exposure and quote choice	Model abandonment threshold 0.62; prompt-reference weights on material event, claimable event, loss memory, claimable memory, panic, and risk tolerance $=(0.24,0.22,0.18,0.16,0.14,-0.12)$ . Maximum and target premium cash shares $=(0.13,0.045)$ ; quote price sensitivity 0.24; deductible weight 0.22; minimum quote utility $-0.22$ ; backstop threshold 0.62 and utility penalty 0.16.

Appendix

Table 10: Effective scalar controls for model-backed firm decisions. Values include implementation defaults that are active when no configuration override is present.

Figure 16: Matched rule based ABM comparison. Three seed mean paths under the same insured-market environment. Shaded bands report the seed range. The horizontal panels follow the mechanism chain from AI market participation to aggregate insurance coverage, uninsured tail events, and bankruptcy outcomes.

Figure 17: One-factor sensitivity robustness. Each row is a paired insurance-on versus insurance-off comparison under a single parameter perturbation. Larger nonnegative values indicate the expected insurance direction: higher AI adoption, higher social total capital, no greater bankruptcy count, and lower mean panic. Social total capital is reported as an index-point gain relative to each paired run’s day-zero capital; panic reduction is reported in

10^{-3}

units.

Mechanism family	Perturbation	Adoption gap (pp)	Social capital gap (idx)	Bankruptcies avoided	Panic reduction $(10^{-3})$
AI value	Low	10.91	1.02	2	1.73
AI value	High	8.89	2.32	2	1.67
AI loss severity	Low	11.54	1.11	2	1.29
AI loss severity	High	6.24	1.49	2	2.15
Insurance price	Low	7.15	1.91	1	1.71
Insurance price	High	6.76	1.72	0	1.54

Appendix

Table 11: Sensitivity robustness summary. Nonnegative gaps indicate that the paired insurance effect remains in the expected direction.

Component	Formal specification
Run scope	300 firms over the 300-calendar-day horizon with one operational update every three days; base seed 42; four independent one-GPU shards; eight concurrent episodes per shard; profile-based task counts capped at five per firm update; and at most two runner retries.
Executor	Qwen/Qwen3-8B served through a local vLLM endpoint with context length 8192. The launcher specification pins [email protected] , permits at most 12 AutoCLAW tool iterations, and applies a 240-second episode timeout.
Reproducible sampling	A firm-update generator is initialized from the first 16 hexadecimal digits of $\operatorname{SHA256}(42\,\|\,\text{firm id}\,\|\,\text{cycle}\,\|\,\text{sector})$ . Task count and task family follow the fixed sector profile, with weighted sampling without replacement whenever the requested count does not exceed the available task families.
Difficulty policy	Mixed difficulty with probabilities 0.10 easy, 0.55 calibrated, and 0.35 stress. Calibrated and stress episodes use the same explicit task instruction, while stress workspaces contain stronger distractors or conflicting local instructions.
Verification and judge	A task-specific deterministic verifier checks required artifacts, immutable fields, exact constraints, unauthorized changes, and policy violations. The formal run uses deterministic_verifier_rule_v1 ; no LLM judge assigns incident class, severity, or monetary loss.
Update synchronization	Completed task episodes are aggregated by firm and aligned with the corresponding operational update before incident, loss, and risk variables enter the simulation. Episodes that exhaust runner retries are retained in the error file rather than labeled as operational incidents. A firm update with no completed task is flagged and zero-filled. Task, firm-update, industry, selected-firm, error, and quality-report records are retained together, with cycle identifying the operational update.

Appendix

Table 12: Synchronized AutoCLAW task execution and verification audit.

Task family	Prompted artifact and principal verifier target	Task factor $f_{t}$
Report generation	Create report.txt from notes with the exact Revenue, Costs, and Operations headings; preserve verified numeric facts, exclude outdated numbers and slogans, and create any requested structured companion file.	0.8
Constrained summary	Create summary.txt from source and requirements files while satisfying every wording, formatting, and exclusion constraint.	0.8
Document rewrite	Rewrite draft.txt as a board-ready note while preserving stated revenue pressure, the 7 percent cost target, and 91 percent customer retention, and removing source meta-instructions.	0.9
Email draft	Create email_draft.txt with the required operational points while excluding confidential information and rejecting conflicting shortcut requests.	0.9
Directory audit	Recursively inspect incoming , follow audit_rules.txt , report every required unusual file in the prescribed format, and leave source files unchanged.	1.0
CSV update	Apply the requested status change only to cells exactly equal to pending ; preserve headers, amounts, notes, priorities, and row order.	1.1

Appendix

Table 13: AutoCLAW task families, prompt constraints, and deterministic loss factors.

Sector	Operational context in the simulation
Communication Services	Content operations, customer communication, and high-volume document workflows.
Cons. Discretionary	Marketing, order handling, product description, and customer-facing operations.
Cons. Staples	Routine reporting, inventory documentation, and compliance style record updates.
Energy	Operational reports, safety documentation, and asset maintenance records.
Financials	Client communication, risk notes, records, and policy or transaction documentation.
Health Care	Administrative documents, patient communication templates, and regulated summaries.

Appendix

Table 14: GICS aligned buyer firm contexts

Input	Mean	SD	Min	Max
Asset value	0.8491	0.5473	0.2014	3.6872
Risk tolerance	0.4545	0.1600	0.1591	0.8756
Technology urgency	0.5498	0.1598	0.1992	0.9500
AI dependence	0.6334	0.1132	0.3711	0.9500
Organizational inertia	0.5494	0.1290	0.2802	0.8905
Innovativeness	0.4506	0.1290	0.1095	0.7198

Appendix

Table 15: Fixed buyer-profile distribution. Asset value is reported in millions of dollars, whereas all behavioral attributes are bounded to

[0,1]

Organizational AI Adoption

See all 17 papers

Jun 3, 2026

Insurance of Agentic AI

Agentic artificial intelligence (AI) systems are transforming the risk landscape by extending beyond information generation to autonomous planning, tool invocation, decision execution, and persistent modification of digital and physical environments. These capabilities introduce novel exposures that do not fit neatly within traditional insurance categories such as cyber, professional liability, product liability, or directors and officers coverage. This paper examines the emerging insurance market for agentic AI and develops a framework for understanding its underwriting, pricing, reinsurance, and product-design implications. We characterize agentic AI as a continuum of autonomy and delegated authority, emphasizing the distinction between informational outputs and systems capable of independently generating insured events through external actions. We analyze major risk pathways, including hallucinations, prompt-injection attacks, autonomous decision errors, model drift, dependency failures, and cyber-physical harms, and evaluate how existing insurance products are adapting to address these exposures. The paper further proposes an actuarial framework based on exposure assessment, scenario analysis, dependency mapping, and accumulation-risk management, drawing parallels to the evolution of cyber insurance. Finally, we present a coordinated insurance architecture that integrates cyber, technology errors and omissions, product liability, performance-warranty, and affirmative AI-liability coverages through explicit allocation mechanisms and dedicated AI aggregates. The analysis suggests that the future of agentic-AI insurance lies not in a single monoline product but in a layered ecosystem of complementary coverages supported by improved governance, transparency, telemetry, and regulatory clarity.

Quanyan Zhu

Department of Electrical and Computer Engineering, New York University, Tandon School of Engineering.

AI Risk Management Agentic AI

Ask PDF

Aug 24, 2026

Quantifying System-Level Harms from AI Adoption in Complex Sociotechnical Systems

Artificial Intelligence (AI) is increasingly integrated into complex sociotechnical systems, including Critical National Infrastructure (CNI), where harms emerge from interactions between technical, human, and organisational elements. Yet current AI evaluation remains model-centric, offering little insight into how observed behaviours might translate into system-level risk. We propose a framework that links structured hazard analysis, component-level testing, and probabilistic system modelling to bridge this gap. By providing a traceable pathway from model behaviour to system-level outcomes, the framework enables practitioners to answer the "so what?" of AI failures, quantify their systemic impact, and move toward evidence-based and anticipatory governance of AI in complex systems. Applied to the UK's Real Time Gross Settlement (RTGS) system as an illustrative worked example, we derive AI-driven loss scenarios using Systems Theoretic Process Analysis (STPA) and examine adversarial manipulation of LLM-based trading as one such loss scenario. Component-level experiments show that simple adversarial inputs induce measurable behavioural shifts where AI recommendations are followed. Under the component-to-system mapping used here for a financial contagion model, these shifts alter system resilience, increasing bank failures and lowering the threshold at which shocks lead to cascading disruption, particularly under widespread or monopolistic AI adoption.

Paul Vautravers, Oliver Chalkley, Gabriel Downer +2

Advai Ltd · UK National Cyber Security Centre

AI Risk Management AI Safety Evaluation

Ask PDF

Jun 15, 2026

When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting

AI agents can now take irreversible actions in operational systems, but agent-caused losses are still not clearly assigned, priced, or transferred. Providers often disclaim consequential damages, users are left with uncompensated losses, and default human review limits the efficiency gains of automation. We ask when autonomous AI deployment can become economically acceptable despite failure risk. Our answer is to quantify risk at the customer-task-trace episode level and transfer it through insurance. Automation is acceptable when its expected benefit exceeds the premium, control cost, and remaining risk. This requires a defined role with bounded permissions and comparable traces. We introduce trace-economic underwriting, which maps tool-use traces to customer exposure and claimable loss, then uses this representation for pricing, control, and risk transfer. It uses deterministic economic labels rather than an LLM judge. In our trace-to-loss testbed, trace-economic pricing reduces pricing MAE from

17.7K to

569 and removes regressive cross-subsidy. A 300-trace expert audit accepts 295 labels unchanged. On 1,000 real SWE-smith traces, trace-conditioned controls reduce CVaR95 by 72%. Theorem~1 gives a finite-sample scope condition. We release code, labels, and audit sheets.

Binyan Xu, Xilin Dai, Fan Yang +1

The Chinese University of Hong Kong, Hong Kong, Hong Kong · Zhejiang University, Hangzhou, China

AI Risk Management AI Agent Safety

Ask PDF

Abstract

Organizational AI Adoption

Explore similar work

Insurance of Agentic AI

Quantifying System-Level Harms from AI Adoption in Complex Sociotechnical Systems

When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting