How Much of the Carrier Has to Be Left?

Sep 30, 2026

IIAI Article Hero Image

AI, Delegated Underwriting, and the Limits of Insurance Unbundling

Analysis
The Economics of AI in Insurance, Part 4
By James W. Moore

Key Takeaways

  • Delegated underwriting and the separation of insurance functions long predate modern artificial intelligence. AI did not create this structure.
  • AI is beginning to make outside underwriting more visible by helping convert fragmented and unstructured information into usable monitoring signals. That is different from proving that the underwriting judgment or price was correct.
  • Better verification may allow carriers and capital providers to delegate more, but it does not eliminate the need for financial alignment or the issuing carrier’s legal responsibility.
  • The boundary should move faster where performance becomes credible quickly and more slowly where losses take years to develop.

Who Is Watching the Underwriter?

Bamboo Insurance describes itself as an underwriting-first, capital-light homeowners managing general underwriter. That description becomes more interesting when you look at what sits on either side of it.

Bamboo controls much of what insurance executives would normally associate with an insurer. It manages pricing, underwriting, policy administration, distribution, technology, and claims. Program carriers issue the policies. Reinsurers and institutional investors provide much of the risk capital. Bamboo retains limited participation through a captive.

You could spend a great deal of time deciding which organization is the “real” insurance company. That is probably the wrong question.

The more useful question is why the institutions standing behind the risk are willing to support underwriting decisions made inside another organization.

For most of the industry’s history, that answer has involved contracts, underwriting guidelines, referral requirements, reporting, audits, performance history, and some form of financial alignment. Bamboo’s 2026 registration statement adds another possibility. It describes an AI-enabled portfolio platform and says its capacity providers receive high-fidelity, real-time analytics for portfolio monitoring and performance review.

That raises the question at the center of this article: if artificial intelligence gives the institutions standing behind the risk better visibility into outside underwriting, what might they be willing to delegate that they would not delegate before?

The Structure Is Older Than the Technology

Insurance began separating its functions long before anyone called the result an ecosystem.

Independent agents controlled distribution. Managing general agents exercised delegated underwriting authority. Third-party administrators handled claims. Reinsurers supplied capacity. Reciprocal exchanges separated a policyholder-owned insurance entity from the attorney-in-fact managing its operations. Lloyd’s connected underwriting judgment with outside capital centuries before the arrival of modern software.

The program-business numbers make it difficult to attribute the current structure to AI. The Target Markets Program Administrators Association estimated the market at $24.7 billion in premium in 2011. Its 2025 State of Program Business Study puts 2024 premium at $110.8 billion.

That growth crossed soft and hard insurance markets and occurred largely before the current generative-AI wave. The movement toward delegated underwriting appears secular. It is not an AI creation.

That does not mean the underwriting cycle has disappeared. The same 2025 study reports that 84 percent of carrier respondents exited at least one program, primarily because of poor performance. It also identifies dependence on capacity and reinsurance, uneven underwriting discipline, and fragmented data infrastructure among the sector’s continuing weaknesses. Most respondents described their use of AI as early-stage, with a majority saying they were “just scratching the surface.”

This looks less like AI suddenly inventing a new insurance model and more like an old model encountering a potentially important new capability.

The old problem was never whether underwriting talent could operate outside the carrier. It was how the carrier controlled the consequences after authority left its walls.

The institutional record is unusually clear on that point. The National Association of Insurance Commissioners’ Managing General Agents Model Act requires written contracts, defined underwriting guidelines, monthly transaction reporting, access to usable records, carrier cancellation rights, and timely claims reporting. It says the acts of the managing general agent are considered the acts of the insurer on whose behalf it operates.

Lloyd’s delegated-authority standards similarly require standardized risk, premium, and claims information. Its current oversight requirements call for continuous monitoring capable of identifying emerging risks, deteriorating performance, noncompliance, and data-quality problems. Lloyd’s also requires the managing agent to retain the internal expertise needed to challenge the information, exercise judgment, and approve important decisions.

Monitoring was not the only limit on delegated underwriting. Capital availability, regulation, contractual control, and the underwriting cycle all mattered. But the cost and delay involved in seeing what an outside underwriter was doing has clearly been one persistent constraint.

That caveat matters in the current market. Gallagher Re estimates that global dedicated reinsurance capital rose 5 percent to a record $688 billion in the first half of 2026, while property and casualty reinsurance premiums among its composite fell 6 percent. If delegated structures gain ground in that environment, abundant capital will be part of the explanation. The harder evidence for AI will still be whether better monitoring changes authority, collateral, retention, or compensation terms.

That is where AI may change the economics.

What AI Adds to the Monitoring Stack

Conventional systems are good at checking structured information against known rules. They can flag a policy written outside an approved territory, above an authority limit, or below a filed rate. Insurers have used rules engines, dashboards, and periodic portfolio reports for years.

The harder problem is that underwriting evidence rarely arrives as one clean dataset. It sits in submissions, inspection reports, loss runs, emails, claim notes, external data feeds, and differently formatted reports from multiple organizations. A carrier may receive the information it technically asked for and still struggle to see what the portfolio is becoming.

AI can help turn more of that material into something that can be monitored.

Accelerant provides the clearest current example, although the distinction in its filing matters. The company’s 2025 Form 10-K says its Risk Exchange ingests structured and unstructured data from policy, claims, and other systems. Once ingested, the data is validated, transformed, and governed into underwriting intelligence available to both managing general agents and risk-capital partners. Accelerant says it uses several third-party technologies, including large language models, to support ingestion, analytics, and the broader exchange.

Elsewhere, the filing says the resulting dataset enables automated portfolio monitoring and provides actionable information to underwriters and capital providers.

That does not mean an AI model independently watches the portfolio. AI is one component of a larger data and monitoring platform. Much of the work remains data engineering, analytics, workflow, and human review. But it is a meaningful component because it can help recover information that conventional structured reporting leaves behind.

Bamboo offers supporting evidence. Its filing describes extracting information from structured and unstructured home-inspection reports for use in proprietary underwriting models. It describes an AI-enabled portfolio-orchestration layer and separately says its capacity providers receive real-time portfolio analytics. Bamboo attributes strong renewals and expansion of its capacity panel partly to its transparency and data quality.

That final connection remains Bamboo’s interpretation. Neither Bamboo nor Accelerant establishes that AI caused a reinsurer to accept less collateral, demand less risk retention, or grant broader authority. Their filings support the first link in the causal chain: AI is beginning to expand what can be made visible inside a broader monitoring system.

The practical significance is easier to see when monitoring is grouped into three questions.

First, did the underwriter stay within the authority it was given? That includes classes, territories, limits, pricing parameters, referral requirements, and exceptions.

Second, is the portfolio changing in ways that individual transactions conceal? Geographic concentrations, industry accumulations, attachment points, data quality, and movement toward the edge of appetite may become apparent only when the book is examined as a whole.

Third, can the carrier reconstruct an important decision? It may need to know what information was available, which model or rule influenced the recommendation, what a human changed, why an exception was approved, and who held the authority to approve it.

Those are questions about process. Better answers can reduce the amount of blind trust required across organizational boundaries. They cannot settle the most important question in underwriting: was the risk priced correctly?

Visibility Is Not the Same as Alignment

The current market contains an apparent contradiction.

Technology is making it easier to separate underwriting operations from carrier paper and risk capital. Yet several sophisticated operators retain some or all of the economics themselves.

Bamboo participates through its captive. Accelerant retains a portion of the business moving through its own underwriting entities. Bowhead uses American Family’s issuing paper, then reinsures the business back to its own insurance subsidiary. Bowhead has separated the legal paper from the underwriting operation without separating the underwriting operation from the economic result.

That does not prove American Family demanded the retention. Bowhead’s underwriters may retain the risk because they want the profit, because their partners value the alignment, or both. What the structure proves is narrower: operating specialization does not inevitably lead to capital separation.

Retention is not the only way to create alignment. Ryan Specialty’s 2025 Form 10-K describes contingent commissions tied to profitability, volume, or growth. Accelerant reports sliding-scale commissions that change with actual loss experience. Its filing describes those commissions as substantive participation in the results even when Accelerant does not retain the underlying insurance risk.

Contracts can also preserve cancellation rights, restrict authority, delay compensation, and require referrals. Ownership can align the people making underwriting decisions with the long-term value of the book.

These mechanisms are not simply substitutes for weak monitoring. Monitoring and alignment address different uncertainties. Monitoring can reveal whether an underwriter followed the agreed process. Alignment gives the underwriter an economic reason to care about the result that cannot yet be seen.

AI may change the mixture. If better visibility reduces uncertainty about conduct, a carrier or reinsurer may become comfortable with broader authority or less frequent intervention. But the uncertainty surrounding price and judgment remains until claims develop.

Better visibility answers one question while leaving the other open.

How Far, and How Fast?

The economic boundary can move as far as the verifiable process extends. It stops where only emerging loss experience can validate price and judgment.

That boundary will not sit in the same place for every line of business.

Neptune offers a useful endpoint. Its 2025 Form 10-K says the flood-insurance platform takes no balance-sheet insurance risk and has no claims-handling responsibility for the policies it sells. It provides capacity providers with real-time access to performance metrics while using automated, data-intensive systems to select and price property-level flood risk.

Neptune should not be treated as proof that AI has eliminated the carrier. Flood is comparatively bounded, heavily modeled, and capable of producing faster feedback than long-tail casualty. Its structure shows that extensive separation is possible under the right conditions. It does not show that the same structure will travel easily into every market.

The NAIC model act offers a surprisingly clear expression of the timing problem. If a managing general agent shares in interim profits and can influence those profits through reserves or claim payments, the act says the profits cannot be paid until one year after they are earned for property business and five years after they are earned for casualty business. Even then, they must be verified.

Those periods are regulatory rules for a particular contractual situation, not a universal measure of loss development. But the distinction they embody is fundamental. Property results generally become credible sooner. Casualty results can appear healthy for years before deficient pricing, legal changes, medical inflation, or adverse claim development becomes visible.

AI can compress a monthly reporting cycle toward continuous oversight. It can identify an accumulation or authority exception much earlier. It cannot make a liability claim mature faster simply because the monitoring platform operates in real time.

That suggests unbundling can proceed faster where process monitoring is meaningful and performance becomes observable relatively quickly. It should proceed more slowly where judgment is difficult to standardize and losses remain uncertain for years.

The market will tell us whether this is occurring through more than company growth. The stronger evidence will be changes in contract terms: broader delegated authority, different collateral requirements, reduced retention, revised sliding-scale commissions, or less restrictive cancellation and referral provisions. Those changes would show that capital providers believe better monitoring can replace at least some of the controls and alignment they previously required.

We do not have that evidence yet. Even if monitoring is improving now, capital providers may wait for enough loss experience to decide whether greater visibility justifies different terms. The lag in the evidence may be part of the mechanism rather than proof against it.

The Second Boundary

There is also a legal boundary, and it is different from the economic one.

The economic question is how much responsibility a carrier or capital provider is willing to delegate. The legal question is who remains obligated when the arrangement fails.

State National has built a business around that distinction. Through its program-services operations, it supplies licenses, ratings, regulatory filings, and issuing paper to programs supported by outside underwriting and capital. The State National discussion in Markel’s 2025 Form 10-K explains that the company reinsures substantially all of the risks written through these arrangements.

It also states that reinsurance does not discharge the issuing insurer from its primary liability to policyholders.

That is not another version of the economic limit. A capital provider may become comfortable delegating more because it receives better information. The issuing carrier cannot monitor its way out of the insurance promise. If an underwriter, model, reinsurer, collateral arrangement, or service provider fails, the carrier remains accountable to the policyholder and regulator.

This is the same principle that applies to AI decision architecture inside the carrier. Authority can be delegated. Ownership of the delegation cannot. The work can move. The promise cannot.

What Has AI Actually Changed?

Bamboo demonstrates how much of an insurance operation can be assembled outside the balance sheet carrying most of the risk. Accelerant shows how AI can contribute to a broader platform that makes fragmented underwriting information more usable to capital providers. Bowhead shows why operating specialization does not necessarily produce capital separation. Neptune shows that near-complete separation is possible in at least one bounded market. State National shows why the carrier remains legally present even when most of the economics moves elsewhere.

The underlying architecture is not new. Delegated underwriting was expanding long before modern AI arrived, and the willingness of carriers and reinsurers to supply capacity will continue to move with performance and the market cycle.

What is newer is the ability to make more of the underwriting process visible across organizational boundaries, with less delay and at greater scale. AI can contribute by extracting and connecting information that previously remained trapped in documents, incompatible systems, or periodic reports.

That may eventually change how much authority and capital move outside the traditional carrier. For now, the evidence supports a more measured conclusion.

AI can tell the carrier more about what an outside underwriter is doing, and tell it sooner. AI can verify the process long before it can verify the price. It cannot make losses emerge faster. The final boundary will still be drawn by experience, alignment, and who remains responsible when the answer arrives.

Additional Reading

Sources

Subscribe to AI Insights Newsletter

AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.

Related Posts

468