How many underwriting and claims decisions can we hand over to AI? It depends on how we design the process

Key takeaways:

• AI’s greatest potential often comes before the decision itself. Data analysis, context building, and recommendation preparation reduce manual work and allow experts to move more quickly to the actual risk assessment or the appropriate next step in the claims process.
• Automation depends on several layers working together. Generative AI, scoring, business rules, and workflow serve different purposes. Together, they create a process that can analyze a case, determine what happens next, and stop when human judgment is required.
• Human in the Loop does not mean manually approving every step. It means deliberately defining where people oversee the process, when they approve an exception, and when they take responsibility for the decision.
• Scale depends on the quality of the entire process. Data, integrations, decision authority, escalation rules, and auditability determine how broadly AI can be used without losing control.

AI is moving deeper into insurance processes, but the boundaries of control are still taking shape

European insurers remain cautious about AI adoption. According to an EIOPA survey conducted between May and July 2025 and published in February 2026, 65% of the 347 insurers surveyed across 25 EU and EEA countries were already actively using generative AI. Most solutions, however, were still in the pilot stage, while 64% of reported use cases focused on back-office productivity, including data extraction and underwriting support. At the same time, insurers reported strong levels of human oversight.¹

The market is no longer asking whether AI has a role in underwriting and claims. The harder challenge is moving from isolated use cases and pilots to processes in which technology operates closer to the actual assessment while the organization remains firmly in control.

What makes it difficult to move from pilot to production?

Contrary to what many organizations initially assume, choosing the AI model is rarely the hardest part. The greater challenge is embedding it in existing processes, data, and operating rules. Common barriers include:

• inconsistent, incomplete, or fragmented data,
• unclear escalation and exception-handling rules,
• poorly defined accountability and decision authority,
• an insufficient audit trail for reconstructing how a case was handled,
• inadequate control over changes to models and rules,
• weak integration with workflows, permissions, and the systems used throughout the process.

Effective AI adoption therefore requires technology, process, and governance to work together. AI can analyze information and prepare a recommendation, but the process determines what happens next, when a case should stop, and when it should be escalated. Governance, in turn, defines accountability, change controls, monitoring requirements, and how the organization can reconstruct what the solution did.

What if we are focusing too much on the decision itself?

Before an underwriter accepts a risk or a claims adjuster determines the next step in a claim, information has to be collected and verified, context has to be established, inconsistencies identified, relevant factors assessed, and the case prepared for proper analysis. According to Capgemini, administrative activities such as data entry and documentation consume 41–43% of underwriters’ time across commercial and personal lines. Only 32–33% is spent on core underwriting work.²

This is exactly where AI can already do far more than read a document or generate a summary.

AI before the decision: where does the real value come from?

The first level of AI adoption is already familiar: OCR, data extraction, document classification, and automated summaries. These capabilities are especially useful in processes that involve large volumes of documents and significant manual movement of information between systems.

The next level begins when AI does more than organize information and starts helping prepare the case for actual assessment.

In underwriting, AI can prepare the submission for risk assessment

AI can:

• consolidate data from documents and systems into a single view of the submission,
• identify missing information, inconsistencies, and contradictions,
• compare information against underwriting policy criteria,
• highlight factors that may affect the risk assessment,
• prepare data for pre-scoring or support an initial assessment,
• flag areas that require closer attention,
• prepare a recommendation with supporting rationale based on the available data.

The underwriter then receives a submission that has already been reviewed, with missing information, risk factors, and points requiring expert judgment clearly highlighted. There is no need to reconstruct the full context from scratch, and the places where underwriters usually lose time stop setting the pace of the process.

In claims, AI can prepare the case for the next step

The priorities in claims are different, with customer response time playing a critical role, but the underlying mechanism is similar. AI can combine data from multiple sources, compare the claim and supporting documentation with policy coverage, identify contradictions, assess case complexity, flag anomalies that require further review, and identify missing information. Based on that analysis, it can recommend the next course of action and explain the rationale.

This gives the claims adjuster an immediate view of which parts of the case require attention and what may determine the next step in the claims process.

What does an insurer gain by moving AI closer to the decision?

If AI only reduces the time spent reviewing documents, the main benefit is productivity. Greater value emerges when it also changes how the organization handles the overall portfolio of cases: straightforward cases move faster, exceptions reach the right experts earlier, and the insurer can handle greater volume without a proportional increase in manual work.

More mature implementations are already showing these kinds of results.

Efficiency and cost

Based on its work with commercial P&C insurers in the United States and the United Kingdom, BCG reports efficiency gains of up to 36% in complex lines of business. These gains come primarily from supporting manual underwriting activities, including data analysis, working with unstructured information, and preparing materials for risk assessment more quickly.³

In claims, BCG points to automation in FNOL data extraction, document processing, and triage. In the implementations observed, this translated into cost reductions of up to 20% and claims processing speeds up to 50% faster.⁴

Faster claims handling

Aviva provides another strong example of the scale of change. Its claims transformation involved more than 80 AI models operating at different points in the process. According to a McKinsey case study, the average time required to determine liability in more complex cases was reduced by 23 days, routing accuracy improved by 30%, and complaints fell by 65%. Cases that require more sensitive or specialized handling, such as personal injury claims, are routed to a human expert.⁵

Broader access to knowledge

In the United Kingdom, Allianz uses BRIAN, a solution designed to help underwriters work with extensive documentation and underwriting guidelines. The system answers questions using approved materials and identifies the source behind each answer, allowing the underwriter to verify it quickly.

The example shows that AI does not have to replace the expert to create value. It can instead give underwriters faster access to the context they need to make a decision. Following a pilot involving nearly 3,000 questions from 190 users, the solution was moved into production.⁶

Potential before the decision

We are pursuing similar opportunities in the Underwriter Portal. Our assumption is that organizing the work before the actual risk assessment can reduce handling time and improve the quality of the data that reaches the expert.

In the process we have modeled, the potential benefits include reducing Time to Quote by 20–30% and bringing Time to Respond down to 24 hours. These are project estimates, and actual results will depend on the organization’s starting point, data quality, scope of automation, and chosen operating model.

When support starts to become automation

In our conversations with insurers, we have found that once AI begins supporting risk assessment and preparing recommendations, the discussion quickly shifts toward automating parts of the process.

And that is where one of the biggest concerns emerges: as more steps become automated, are we effectively handing the decision over to AI and removing the human from the process?

Does an automated decision mean handing the decision over to AI?

Process automation does not mean that a single model independently analyzes a case and makes the final determination. In practice, the outcome depends on several components, each with a different role.

Does a human need to be involved in every case?

There is no single answer that applies to every product or type of risk.

A standard case with complete data that falls within the organization’s risk appetite and established authority thresholds requires a different level of control than a complex corporate risk, an unusual claim, or a case in which the consequences of an incorrect decision could be significant.

Factors that matter when determining the appropriate level of human involvement include:

• the consequences of the decision,
• the quality and completeness of the data,
• whether the recommendation can be independently verified,
• the model’s confidence level,
• underwriting policy or claims-handling rules,
• authority and decision limits,
• requirements for additional approval, for example under a four-eyes or dual-approval principle,
• the ability to stop the process and escalate an exception.

Human in the Loop does not mean manually approving every case

The level of expert involvement can vary depending on the risk and complexity of the case.

This approach is consistent with the direction of European regulation. In its August 2025 opinion on AI governance and risk management, EIOPA advocates a risk-based and proportionate approach. Areas it identifies as particularly important include data governance, documentation of how the system operates, security, explainability, and human oversight.⁷

Human oversight should therefore be distinguished from manual approval of every standard case. The goal is to ensure that the organization knows where the process may operate automatically, when it must stop, and when an expert must take over.

Not every use of AI in insurance is considered high risk

Using AI in underwriting or claims does not automatically mean the solution qualifies as a high-risk system under the EU AI Act.

In insurance, the high-risk category includes, among other things, systems used to assess risk and determine pricing for individuals in life and health insurance. For other applications, the specific purpose of the system, its role in the process, and how its output is used all matter.

That does not mean other AI applications fall outside governance requirements. The level of governance should reflect the risks associated with the specific solution and its impact on the customer, the process, and the decision.

Where does the human fit into this process?

Greater automation changes when expert judgment is needed. The expert remains part of the process, but enters later and in a different role.

Underwriters and claims adjusters create the most value in ambiguous cases, exceptions, situations involving conflicting signals, and decisions with material consequences for the customer or portfolio. If technology has already prepared the data, context, and recommendation, the expert can focus on that point in the process instead of starting by assembling the case.

This trend is likely to continue as more autonomous solutions develop. McKinsey describes future underwriting models in which specialized AI agents handle tasks such as submission intake, risk profiling, pricing, and compliance checks, while an orchestration layer determines which cases can proceed and which should be referred to a senior underwriter.⁸

The choice, then, is not simply “AI or human.” It is about creating a more precise division of responsibility among technology, rules, and experts.

The closer AI gets to the decision, the more important auditability becomes

If AI, scoring models, business rules, workflow, and an expert all contribute to the outcome of a case, the organization must be able to reconstruct the role each of them played. Knowing the final decision is not enough. It must also be possible to determine:

• what data the system used,
• which version of the model and rules was active,
• what factors influenced the recommendation,
• why the case was routed to a particular path,
• whether the expert changed the recommendation and why,
• when the process was expected to stop and hand the case over to a human.

This requires version-controlled rules, clearly defined thresholds and escalation logic, a trace of AI activity, and a record of the human decision. Many of these elements rest on the same foundation as broader data and AI work across the insurance value chain: data quality, a clear data model, and clearly assigned accountability.

That is why AI is difficult to treat as a standalone tool added to an existing process. Its output must feed into workflow, use data from the organization’s systems, and respect established permissions and decision authority. This does not necessarily require replacing the core system, but it does require connecting these layers into a single controlled process.

This stage often proves more difficult than launching the pilot itself. In the previously cited 2026 EIOPA survey, insurers identified data privacy and security, regulatory compliance, and skills gaps among the main barriers to GenAI adoption. At the same time, 49% of surveyed insurers already had their own AI policy, compared with roughly one-quarter in 2023.⁹

So how many decisions can we hand over to AI?

There is no single answer for the entire market. The appropriate scope of AI will vary depending on the product, type of risk, data quality, consequences of the decision, and regulatory requirements.

Even today, however, technology can take over a significant share of the work leading up to a decision: analyzing information, identifying gaps and inconsistencies, building case context, highlighting risk factors, and preparing recommendations. Instead of starting by assembling information, the underwriter or claims adjuster can begin with a case that is already prepared for expert review.