A New Framework for AI Autonomy in Insurance

The insurance industry stands at a critical inflection point. As artificial intelligence capabilities accelerate from simple automation to sophisticated autonomous agents, insurers face a pressing question: How much autonomy should we grant to AI systems in high-stakes processes like claims processing?

Recent research from the University of Washington introduces a framework that reframes this question entirely, original paper here. Rather than treating autonomy as an inevitable consequence of AI capability, the framework proposes that autonomy should be a deliberate design decision—one that balances efficiency gains against accountability, transparency, and regulatory compliance.

For insurance companies navigating the complex landscape of AI adoption, understanding and implementing appropriate levels of autonomy isn't just a technical consideration. It's a strategic imperative that will determine competitive advantage, customer trust, and regulatory standing in the years ahead.

Understanding the Five Levels of AI Autonomy

The framework introduces five distinct levels of autonomy, each defined by the role a user plays when interacting with an AI agent. Critically, these levels are independent of an AI system's underlying capabilities—a highly capable AI can operate at low autonomy if designed with appropriate controls, while a less sophisticated system might function autonomously within narrow, well-defined parameters.

Level 1: User as Operator

At this foundation level, the AI functions as an on-demand assistant. Users maintain complete control over planning and decision-making, with the AI providing contextual suggestions only when explicitly invoked. Think of claims processing systems that offer smart recommendations but require explicit approval before any action. The human operator drives the workflow while AI accelerates specific subtasks. Think an AI copilot for reviewing claims and searching for related legal documents & references.

Level 2: User as Collaborator

Here, both human and AI can plan, delegate, and execute tasks in parallel. The AI works independently on assigned responsibilities while maintaining transparent communication about progress and blockers. In claims scenarios, this might involve AI handling document extraction and preliminary fraud checks while human adjusters focus on complex liability assessments. Crucially, users can take control at any point and directly edit AI outputs.

Level 3: User as Consultant

The AI takes primary responsibility for planning and execution, but proactively seeks human guidance on critical decisions. Rather than hands-on collaboration, humans provide directional feedback, domain expertise, and approval of major plan changes. For claims processing, the AI might independently triage incoming First Notice of Loss (FNOL) reports but consult adjusters when encountering ambiguous policy language or unusual circumstances.

Level 4: User as Approver

Human involvement becomes more passive, limited primarily to resolving blockers the AI cannot handle independently and approving consequential actions. The AI makes most decisions autonomously, only requesting input when it encounters failure states or actions flagged as requiring approval. In claims workflows, this could mean AI processing straightforward claims end-to-end but requesting human sign-off on settlements above certain thresholds.

Level 5: User as Observer

Full autonomy. The AI plans and executes all tasks without requiring human involvement, with users relegated to monitoring via activity logs. The only control mechanism is an emergency off-switch. While theoretically possible for highly constrained scenarios, Level 5 autonomy raises significant concerns in insurance contexts where regulatory compliance, fairness, and explainability are paramount.

The Critical Challenge: Claims Processing Automation

To understand why this autonomy framework matters, consider the claims processing landscape. Traditional manual claims handling faces cascading challenges that directly impact both operational efficiency and customer satisfaction.

The Manual Processing Crisis

Insurance claims processing remains stubbornly labor-intensive despite decades of digitization efforts. Claims agents navigate fragmented systems to gather information, manually extract data from unstructured documents including bills, invoices, pharmacy receipts, and medical diagnoses, then categorize everything for adjudication. This process creates multiple pain points:

Heavy operational costs from labor-intensive data entry and document review
Inconsistent delivery, with different agents reaching varying conclusions on identical claims
Poor data integration across multiple sources leading to inaccuracies and outdated information
Lengthy processing times that frustrate policyholders and damage insurer reputations
High error rates from manual data entry creating downstream processing failures

The adjudication workflow itself compounds these challenges. Claims move through initial review, automated checks, manual review for flagged items, payment determination, and finally payment delivery. At each stage, manual processes create bottlenecks. The initial processing review alone—where basic errors like missing patient names, incorrect service codes, or invalid diagnosis codes are caught—represents a significant source of denials that could be prevented with better automation.

First Notice of Loss: The Critical Entry Point

The FNOL process represents perhaps the most critical juncture for automation. When policyholders report an incident, insurers must capture accurate information quickly to initiate claims processing, arrange emergency assistance, and set expectations. Yet traditional FNOL processes struggle with reporting delays, incomplete information capture, and manual handling that creates friction at precisely the moment customers are most stressed.

AI-powered FNOL automation promises dramatic improvements: up to 70% faster claim initiation, 24/7 availability, instant validation of policy coverage, and real-time document capture with automated data extraction. The business case seems compelling—yet implementation without careful attention to autonomy levels creates new risks.

Where AI Autonomy Goes Wrong in Insurance

The insurance industry's rush toward AI automation has revealed critical failure modes that the autonomy framework helps explain and prevent.

The Black Box Problem

AI systems, particularly those using deep learning, often function as opaque "black boxes" where decision logic remains inscrutable. For a heavily regulated industry where transparency, explainability, and auditability are regulatory requirements, this opacity poses existential risk. When an AI system denies a claim or adjusts a settlement, insurers must be able to explain why to both regulators and policyholders.

Explainable AI (XAI) frameworks have emerged as a partial solution, using techniques like SHAP values, accumulated local effects, and counterfactual explanations to illuminate model reasoning. However, explainability alone doesn't solve the autonomy problem—it only makes autonomous decisions more auditable after the fact.

Over-Reliance and Skill Degradation

When claims adjusters grow accustomed to accepting AI recommendations without critical evaluation, two problems emerge: errors of commission where incorrect AI suggestions are rubber-stamped, and gradual erosion of the adjuster's ability to handle novel situations or consider multiple perspectives. This skill degradation represents a long-term organizational risk that undermines human expertise over time.

Human-in-the-loop (HITL) approaches attempt to mitigate this by keeping insurance professionals involved at strategic decision points. Rather than replacing underwriters and adjusters, HITL workflows equip them with AI tools while maintaining human accountability for final decisions. The most effective implementations create feedback loops where human corrections actively improve AI performance over time.

The Expectation Mismatch

Perhaps most tellingly, early automation deployments have revealed disconnects between what AI optimizes for and what customers actually want. One insurer implemented automated claims payment that functioned flawlessly from a technical standpoint—claims were paid immediately in line with policy coverage. Yet customers were deeply dissatisfied because they wanted expense reimbursement, not benefit payments. The automation worked exactly as designed but solved the wrong problem, with efficiency savings immediately consumed by complaint handling.

Regulatory and Compliance Headwinds

As AI capabilities advance, regulatory scrutiny intensifies. The European AI Act classifies certain AI applications in insurance as high-risk, particularly around pricing and risk assessment. Insurers must demonstrate that automated systems provide transparency at least equivalent to human underwriter decisions. Systems that operate at higher autonomy levels face proportionally greater regulatory burden.

Designing Appropriate Autonomy for Claims Processing

The autonomy framework provides a structured approach to deploying AI in claims workflows that balances efficiency, accuracy, and accountability.

Level 1-2 for High-Stakes Decisions

For complex claims involving significant settlements, ambiguous liability, or potential fraud, Level 1 or Level 2 autonomy proves most appropriate. AI can accelerate document processing, flag potential issues, and suggest preliminary assessments, but human operators maintain control over planning and final decisions. This approach preserves the domain expertise and nuanced judgment that separates excellent claims handling from merely adequate processing.

A Level 2 collaborative approach might involve AI extracting structured data from medical bills while the adjuster focuses on evaluating medical necessity and policy coverage. Both work in parallel with transparent visibility into each other's progress. If the adjuster notices the AI misclassifying a procedure code, they can intervene immediately and correct it—creating training data that improves future performance.

Level 3 for Standard Claims with Complexity

Many claims fall into a middle category: not simple enough for full automation, but common enough that AI can handle most processing steps. Level 3 autonomy positions the AI as the primary executor while keeping human experts available as consultants on critical judgment calls.

For auto claims processing, the AI might independently assess damage from submitted photos, consult repair cost databases, and draft a preliminary settlement. However, when it encounters unusual damage patterns inconsistent with the reported incident, or when policy language around coverage remains ambiguous, the AI consults the adjuster for guidance rather than proceeding autonomously. This preserves adjuster expertise for high-value decisions while automating routine workflows.

Level 4 for High-Volume, Low-Complexity Processing

Certain claims categories lend themselves to higher autonomy. Simple liability determinations, straightforward coverage assessments, and routine claims below materiality thresholds can operate at Level 4, with AI handling end-to-end processing and requesting approval only for exceptional cases.

Property damage claims below a certain value threshold might be entirely AI-processed: the system verifies coverage, assesses damage via computer vision analysis of submitted photos, calculates settlement based on repair estimates, and issues payment—all without human intervention unless the claim exhibits fraud indicators or exceeds approval thresholds. This frees adjusters to focus on complex cases requiring human judgment.

Level 5: Proceed with Extreme Caution

Full autonomy (Level 5) remains inappropriate for most insurance contexts, for now, given regulatory requirements, fairness concerns, and the high stakes of claim decisions. The limited scenarios where Level 5 might be justified—perhaps for micro-claims in sandboxed test environments—require extraordinary safeguards and should be approached cautiously.

How Denklinie Transforms Insurance AI Deployment

At Denklinie, we've built our entire go-to-market strategy around Forward Deployed Engineering precisely because we recognize that AI autonomy in insurance cannot be solved with off-the-shelf products.

Our Approach

When Denklinie begins working with an insurance carrier, our engagement follows a structured FDE lifecycle designed to implement appropriate AI autonomy for their specific claims operations:

Phase 1: Discovery and Autonomy Mapping (2-4 weeks)
Phase 2: Rapid Prototyping with Real Data (3-6 weeks)
Phase 3: Production Hardening and Scaling (8-12 weeks)
Phase 4: Continuous Optimization and Knowledge Transfer (Ongoing)

The Competitive Advantage of Getting Autonomy Right

Insurers who thoughtfully implement appropriate AI autonomy levels gain multiple strategic advantages:

Regulatory Confidence: Clear autonomy frameworks and audit trails demonstrate to regulators that AI deployment is deliberate, controlled, and aligned with consumer protection principles.
Customer Trust: Policyholders appreciate fast claims processing but also want assurance that complex decisions receive human consideration. Graduated autonomy delivers both speed and judgment.
Talent Attraction and Retention: Claims professionals don't want to be replaced by AI, but they also don't want to spend their expertise on data entry. Appropriate autonomy elevates their role to focus on work that requires human judgment.
Risk Management: By restricting high autonomy to well-bounded scenarios while maintaining human oversight for complex cases, insurers reduce the risk of AI making consequential errors.
Continuous Improvement: Systems designed with appropriate autonomy levels and HITL feedback loops get progressively better as they learn from human corrections, while fully autonomous systems can compound errors without oversight.

The Path Forward

The insurance industry's AI transformation is inevitable, but how autonomy is implemented will separate winners from losers. Carriers that treat autonomy as a deliberate design decision—carefully calibrated to each use case's complexity, risk profile, and regulatory context—will capture efficiency gains while maintaining the human judgment that complex insurance decisions require.

The question isn't whether AI will transform insurance operations—it's whether you'll implement it thoughtfully or reactively. The autonomy framework provides the strategic foundation. Forward Deployed Engineering provides the execution model. Together, they represent the future of insurance AI done right.

Ready to explore how appropriate AI autonomy could transform your claims operation? Denklinie is ready to embed with your organization and build solutions tailored to your specific needs. Book a discovery call to begin the conversation.