Introduction

Two conversations about AI in insurance are running side by side, and they rarely acknowledge each other. In one, vendors promise autonomous systems that underwrite risk and make coverage decisions with little human involvement. In the other, regulators are turning oversight into practice: more than two dozen US states have adopted AI governance guidance for insurers, examiners are piloting a standardized AI evaluation tool across twelve states, and the Lloyd’s Market Association has issued a toolkit to help managing agents govern these same technologies. Underwriting leaders have to work in the gap between the two, and the question they keep running into is a practical one: which parts of the workflow should an agent actually own?

AllDigital Specialty operates from an AI-native position. Our underwriting workflows were designed around AI from the start instead of being retrofitted onto processes built before AI existed, and we run live specialty submissions through that architecture today. Working this way makes us optimistic about what agents can do, and it also means we run into their limits early. Risk is complicated and context-dependent, and the consequences of getting it wrong often land years later. That is why underwriting has always been a judgment profession, and why giving an agent the wrong task is a liability decision however it looks on a roadmap.

What follows is the framework we use to decide how much autonomy belongs at each point in a specialty underwriting workflow. It also covers the regulatory boundaries that framework has to respect and what a mature version of it looks like in practice.

A Working Framework: Three Tiers of Autonomy

Most commentary on AI in underwriting stops at “ensuring human governance.” That’s true in spirit, but it doesn’t say where that governance actually sits. We assign every task to one of three tiers, using a four-part test. How reversible is the error if the system gets it wrong? How much money is on the table? How much genuine ambiguity is involved? And who has to answer for the outcome?

Tier 1: Autonomous execution. The agent runs the task end to end. This is reserved for work where errors are recoverable, outputs can be verified, and little judgment is required. Review happens by exception (low confidence scores, anomalies, periodic sampling) rather than on every file.

Tier 2: Suggest and confirm. The agent produces an output that materially shapes the outcome, but a human has to ratify it before anything binds. That confirmation has to be a real review. Whether it functions that way in practice depends on underwriter workload as much as on policy.

Tier 3: Human decision, AI support. The decision stays with a named underwriter. AI can surface precedent, summarize exposure, or draft language, but it does not decide. This tier covers outcomes that are adverse, hard to reverse, financially significant, or genuinely ambiguous.

The tiers are not fixed. A task can move from Tier 2 toward Tier 1 as confidence data accumulates and controls mature, which is what responsible expansion of autonomy looks like. The failure mode is Tier 3 work drifting downward because a vendor labeled it “agent-ready.”

Signal and Noise: Getting the Vocabulary Right

“AI agent” has become one of the most overloaded phrases in insurance technology, and that sloppiness is how organizations end up automating decisions they never meant to hand over. Three terms routinely get collapsed into one, and they shouldn’t be.

An AI agent is a component: a system that perceives context, reasons about a goal, uses tools, and takes a sequence of actions, often without step-by-step instruction. A submission-intake agent that reads a broker email, extracts risk parameters, and populates a workbench is an agent, which is to say an architectural unit and nothing more.

Agentic AI describes a system property: a workflow in which agents plan, delegate, act, and self-correct across multiple steps with reduced human intervention. A pipeline that classifies a submission, enriches it, scores appetite fit, and routes it by confidence level is agentic AI. The extraction step within it is just one agent doing its job.

AI-native describes an organizational stance. The process was designed assuming AI from day one, staffing and unit economics and controls included, instead of being bolted onto a legacy workflow. Incumbents can buy agents easily enough. Redesigning how the work flows is the harder part, and it isn’t something a vendor can sell you.

In vendor materials, a deterministic document classifier and a system marketed as “making underwriting decisions” both get called agents. The tiers above give us a way to tell them apart before anything reaches production.

Where AI Agents Deliver Real Value Today

The best candidates for automation share a profile. They are data-intensive rather than judgment-intensive, structured or semi-structured, high in volume, and forgiving of a wrong answer. In every example below, the agent informs the decision without making it.

Submission intake and triage is the clearest early win. An agent receives a submission, extracts risk parameters, checks them against appetite guidelines, flags gaps, and routes the file to the right underwriter within seconds. Work that used to take a junior underwriter twenty to thirty minutes per submission now happens in the background before they’ve opened the file.

Data gathering and enrichment follows the same logic. An agent that pulls public records, carrier history and regulatory filings into a structured summary means the underwriter starts from a fuller picture instead of spending the morning assembling one.

Document classification and extraction handles the unstructured content that dominates the operation: applications, loss runs, financials, engineering reports. It moves faster than manual review, and it keeps underwriters on analysis rather than data entry.

Loss run analysis works well because reviewing multi-year histories and flagging frequency and severity outliers is exactly the kind of pattern-matching agents are good at. The underwriter still decides what the pattern means.

Appetite matching rounds out the core set. Against documented guidelines, an agent sorts a submission into clearly in, clearly out, or grey-zone, so underwriter time goes where it’s actually needed. Renewal preparation, pre-issuance compliance checks, and routine broker communication follow the same pattern.

The economics back this up. BCG’s research with US and UK commercial P&C insurers found AI can improve underwriting efficiency in complex lines by up to 36%, largely by augmenting manual processes, plus up to three additional points of loss-ratio improvement from better use of unstructured data. WTW’s 2026 Advanced Analytics and AI Survey, drawing on 59 P&C insurers across the US and Canada, found that analytics-sophisticated insurers ran combined ratios six points lower than slower adopters between 2022 and 2024. None of that is really in dispute. The harder question is where a given organization can safely collect the gain, which is what the tiering is meant to settle.

That is how our own workflow is arranged. Intake, document classification and enrichment run at Tier 1, so a specialty submission is parsed, checked against appetite and routed before an underwriter opens the file, with review triggered by low confidence scores and anomaly checks rather than applied to every case. Appetite fit scoring and preliminary pricing sit at Tier 2, where an underwriter signs off before anything moves toward a broker. Declinations, appetite exceptions and manuscript wording sit at Tier 3 with a named underwriter. Because the workflow was designed around AI rather than adapted to it, those boundaries were part of the original design rather than controls bolted on after agents were already running.

Where Human Underwriters Remain Essential

Tier 3 isn’t a temporary concession to immature technology. It reflects how specialty underwriting actually works.

Complex risk assessment often involves novel or combined exposures that require understanding how an industry actually operates, not just how it reads on an application. A large construction project in a flood zone, paired with unusual risk transfer terms and a contractor with a mixed loss history, fails the ambiguity test on its face.

Ambiguous submissions belong with humans for a structural reason: specialty exists to cover what standard markets won’t touch, so the submissions landing on these desks are the ones least suited to pattern-matching against historical data. Agent performance tends to degrade exactly where expertise matters most.

Coverage interpretation fails the materiality test. Policy language, endorsements and exclusions interact in ways that only surface at claim time, sometimes years later and rarely reversibly. Structuring coverage or drafting manuscript language needs to stay under close human control.

Appetite exceptions fail the accountability test. Writing outside standard appetite carries financial and reputational weight, and that requires someone with both the authority to make the call and the obligation to answer for it.

Relationship management and negotiation stay human because the underlying skill is human. Trust with a producing broker takes years to build and one bad renewal conversation to damage. Knowing when to hold the line on terms and when to bend to keep an account is not something current AI systems can do.
Wherever a wrong decision is hard to reverse, financially significant, genuinely ambiguous, or carries individual accountability, human governance applies — not a single approval step, but ongoing accountability for the decision. None of that is a gap better technology will eventually close. It is a design choice, and it is increasingly the one regulators expect to find when they look.

The Regulatory Perimeter

Regulators have been sharpening their answer to this question since 2023, with growing specificity.
In the US, the NAIC’s December 2023 Model Bulletin on the Use of AI Systems by Insurers, now adopted by 24 states plus DC, makes one point every underwriting leader should internalize: there is no AI exemption. AI-supported decisions still have to comply with existing unfair trade practice and governance law, and accountability can’t be delegated to a vendor or assigned to the agent itself. For an MGA writing on carrier paper, that creates a two-layer chain: the MGA owns its outcomes, and so does the carrier whose paper it binds. Oversight is also moving from principle to examination. The NAIC is piloting a standardized AI Systems Evaluation Tool across twelve states between March and September 2026.

State-level detail keeps sharpening. New York’s DFS Circular Letter No. 7, issued in July 2024, requires documented fairness testing, bias assessment, and board-level accountability for AI outcomes in underwriting and pricing. Colorado’s rules for predictive models and external consumer data, originally written for life insurers and expanded in 2025 to auto and health, are widely seen as a template other lines and states will adopt. With a national framework still contested at the federal level, multistate writers should design to the strictest applicable state standard rather than wait for consensus.

London is moving in the same direction. The Lloyd’s Market Association’s April 2026 AI Adoption Toolkit, developed with Barnett Waddingham, helps managing agents build governance frameworks on top of existing PRA model-risk expectations. One nuance worth noting: the EU AI Act’s high-risk classification targets AI used for risk assessment and pricing of natural persons in life and health insurance specifically, so commercial specialty lines generally fall outside it. Transparency obligations still apply.

The implication for agentic systems is direct. A multi-step agent that plans and self-corrects produces a decision trail that’s harder to explain to an examiner than a static model. “Why did the system decline this risk?” has to survive a market conduct exam. That is a practical argument for keeping agents at the workflow layer, keeping interpretable logic at the decision point, and keeping human accountability wherever the outcome is adverse. Read alongside the bulletins and circulars, a tiered model is close to what regulators are already asking firms to demonstrate.

AI Assistance Before AI Autonomy

The AI deployments that hold up over the next several years will be the ones that make underwriters substantially more effective at the parts of the job that genuinely require them. Very few will look like replacement.

Take a rough illustration. An underwriter handling 200 submissions a month, with a large share of that time lost to intake and administration, can plausibly handle 350 to 400 once agents absorb the preparatory work, and the time that comes back is spent on judgment-intensive decisions. The operation gains capacity without a proportional headcount increase. There is also a reasonable case that the decisions themselves improve, simply because attention is no longer fragmented across low-value tasks.

The approach compounds. It builds capability and trust gradually, keeps a human on every high-stakes call, and generates the observational data needed to move tasks between tiers safely, all while preserving the accountability that regulators and carrier partners require. Firms that race to maximize autonomy usually discover its limits through a failure nobody saw coming. Deploying AI as an assistant first, validating what works and expanding from there is slower, and it tends to hold up.

From Augmentation to Orchestration

Augmentation is the foundation for what comes next, and the next stage is worth naming precisely. It is agentic AI in the sense defined earlier: multiple specialized agents and human underwriters operating as one governed system.

In an orchestrated environment, a submission triggers an intake agent that classifies and extracts data. An enrichment agent retrieves third-party information. A summarization agent prepares a structured view for the underwriter. Confidence-based routing then decides whether the file proceeds automatically, waits for confirmation, or escalates to a human who makes the coverage and pricing call. Compliance checks run at each step along the way, with issuance following once everything clears.

There is a strategic point underneath this architecture. The foundation models powering these agents are available to any organization willing to rent them, so the model itself confers no lasting advantage. Composition is what matters: how an organization assembles agents, data, human judgment and controls into a workflow that holds up under scrutiny. That has to be built into how the organization actually works, which is why operators designed around tiered autonomy from the outset tend to have a real edge over incumbents retrofitting agents onto legacy processes.

For most underwriting leaders the live question is sequencing rather than whether to build toward orchestration at all. Start with Tier 1 work. Instrument it thoroughly. Graduate tasks as the evidence supports it. The rule worth holding to is that governance stays ahead of autonomy, never the other way round.

Conclusion

AI agents have a real and growing role in specialty underwriting. Organizations that deploy them well for intake, enrichment, classification and analysis will write more business, respond to brokers faster, and free underwriters for the work that actually needs human expertise. Those advantages are available inside the regulatory perimeter, which is where they have to be earned.

That value depends on where AI is put to work. Applied to tasks that need human judgment and clear accountability, agentic AI makes underwriting riskier, and those risks now tend to surface in exam findings rather than only in hindsight.

The choice in front of underwriting leaders was never AI versus human underwriters. It is whether the orchestration gets built deliberately, with agents handling what agents handle well and underwriters holding the decisions that carry accountability, or whether it accumulates one vendor pitch at a time until nobody can say clearly who decided what.

AllDigital Specialty is an AI-native specialty insurance organization that designs, builds, and operates intelligent underwriting infrastructure in live production, improving underwriting precision, operational efficiency and broker responsiveness across specialty lines.

References

[1] National Association of Insurance Commissioners (NAIC). Model Bulletin: Use of Artificial Intelligence Systems by Insurers. Adopted December 4, 2023.

[2] Quarles Law Firm. “Nearly Half of States Have Now Adopted NAIC Model Bulletin on Insurers’ Use of AI.” April 2, 2025.

[3] Fenwick & West. “NAIC Expands AI Systems Evaluation Tool Pilot Program to 12 States.” March 9, 2026.

[4] Lloyd’s Market Association (LMA) and Barnett Waddingham. “LMA Launches AI Adoption Toolkit to Support Governance-Led Implementation Across the Lloyd’s Market.” April 23, 2026.

[5] Boston Consulting Group (BCG). “How Insurers Can Supercharge Their Strategy with AI.” April 11, 2025.

[6] WTW. “Insurers Using Advanced Analytics and AI Report Strong Returns on Investment and Premium Growth.” March 19, 2026.

[7] New York Department of Financial Services (NYDFS). Insurance Circular Letter No. 7 (2024). Issued July 11, 2024.

[8] Colorado Division of Insurance. Regulation 10-1-1. Adopted September 21, 2023; expanded to auto and health, effective October 15, 2025.

[9] European Parliament and Council of the European Union. Regulation (EU) 2024/1689 (EU AI Act), Annex III, Point 5(c).