Underwriters spend 30% to 40% of their time on administrative tasks, according to McKinsey research on commercial P&C technology. The rest goes to opening broker emails, rekeying ACORD forms, chasing loss runs, and sorting submissions that will never bind. AI for commercial insurance underwriting can recover that time through document automation: extracting structured data, validating it, and routing each risk to the right underwriter or workflow.

The practical decisions sit in the confidence thresholds and routing rules carriers set for exceptions and referrals, and in the governance documentation that has to sit alongside them.

What document automation for underwriting covers

The scope spans data ingestion and process automation. AI extracts data from documents in multiple formats for new business and renewal submissions, including loss runs, then APIs or robotic process automation pass that data into underwriting systems. This reduces rekeying and shortens submission-to-quote time.

The level of automation varies by risk. Commercial-lines automation falls into two categories: complete straight-through processing, or partial automation of specific sub-processes. For complex commercial and specialty business, automation generally prepares the work. An underwriter still decides whether to bind or decline and what to quote. Straight-through processing suits risks whose underwriting rules and decisions can be explicitly codified; the more judgment a risk requires, the less it belongs on that path.

Extraction: turning submission packages into structured data

A commercial submission can contain many document types, each holding part of the risk picture in a different format.

How does extraction turn documents into structured data?

Intelligent document processing handles this in two steps. Classification identifies broker emails, ACORD forms, statements of values, loss runs, spreadsheets, and supplemental applications, then selects the right extraction workflow so underwriters don't manually sort submission packages.

Field extraction then pulls names, addresses, class codes, COPE attributes, TIVs, and claims histories from those varied files, so carriers can review the extracted fields, flag gaps, and map accepted data into underwriting systems.

Which extraction method fits each document type?

For forms and complex tables, layout-aware models combine text with its position on the page, which helps them interpret forms and preserve relationships within complex tables. For documents with varying structures, content-aware extraction adapts better than workflows tied to a single fixed template.

Carriers can combine LLMs with rules and model configuration, including prompt engineering and fine-tuning, to extract and validate critical data points from variable documents while keeping human review for uncertain outputs.

What accuracy should carriers require?

Vendor-reported accuracy figures provide indicative benchmarks, but they lack independent audits. Requirements should reflect the document type, field importance, document quality, and the consequences of an incorrect value:

• Standard ACORD forms: set field-level targets for consistent, structured information, and separately monitor fields that require interpretation

• Schedules of values: evaluate individual-field accuracy and whether relationships among locations and their values across coverage parts stay intact

• Loss runs: test performance across carrier formats and track the claims fields whose errors would materially affect risk assessment

Before committing to straight-through processing, carriers should define acceptance criteria for each document class. Lower accuracy creates manual work downstream; a single portfolio-wide target can conceal weak performance on decision-critical fields.

How widely has intelligent ingestion been adopted?

Adoption is early but accelerating. A Sollers survey of ten markets found that one-fifth of commercial insurers already use AI to triage submissions and extract data from unstructured documents, the most visible area of AI progress in underwriting so far.

Validation: confidence scores and exception queues

A practical validation design uses field-level confidence scores. Reliable data passes downstream; uncertain fields enter a human exception queue. An extraction error that nobody catches can become a mispriced policy, so carriers should set distinct thresholds based on field importance.

Thresholds require trade-offs. A system that sends too many documents to review undermines the economics of automation; one that accepts uncertain data creates silent errors. Use human corrections as structured feedback for extraction models, and monitor whether recurring exceptions decline over time. Before production, test agents against historical submissions with known values on the carrier's actual document mix.

Carriers applying AI or LLMs need enough explainability to show how automated outputs and resulting referrals or decisions were produced, and the insurer retains that responsibility even when it uses a third-party vendor.

Routing and triage: matching risks to underwriters and appetite

Routing determines whether clean data reaches an authorized underwriter with the right expertise and available capacity. Its value depends on identifying in-appetite risks early without unnecessary referrals.

How does automated routing work?

Manual triage consumes processing capacity before an underwriter can assess the risk, and submissions that never bind may still require substantial handling. The workflow has two stages. Clearing the submission completes applicable operational and compliance checks before underwriter review, and determines whether it requires an operational exception.

Assessing and routing the risk combines appetite criteria with indicators of submission complexity, such as scale, sanction flags, risk type, completeness, and required authority; a rules-based process can then route submissions automatically to the right underwriter or underwriting operations professional.

What causes referral overload?

Routing creates referral overload when near-appetite cases escalate because the system can't represent conditional approvals. Carriers should test whether their routing logic can handle missing data or authority limits without sending every borderline case to the same queue.

What outcomes can better routing improve?

The payoff appears at the top of the funnel. In-appetite business may never get quoted if it reaches an underwriter too late. AIG reported that by its Q3 2025 earnings call, it was processing 100% of applicable private and nonprofit submissions through its AIG Assist tool, with faster handling and higher submit-to-bind ratios as a result.

Governance: what regulators expect from automated underwriting

The NAIC Model Bulletin on the use of AI systems by insurers, adopted in December 2023, requires a written AI Systems Program covering governance and risk controls proportional to potential harm, testing for bias and unfair discrimination, third-party vendor oversight, and documentation sufficient to reconstruct any consumer-facing decision under examination. As of mid-2026, several states and the District of Columbia had formally adopted it, with several more in progress.

Supporting operational controls for automated triage and related pricing or referral decisions can include:

• Decision logs that support examination readiness and faster incident investigation

• Escalation triggers and field-level review rules

• Model inventories that make automated systems traceable during reviews

• Human review points

Governance requirements vary by market

In the UK, a practical implementation approach maps AI controls to existing applicable requirements, which continue to apply alongside any AI-specific regime. The LMA published an AI Adoption Toolkit in April 2026, developed with Barnett Waddingham, to help Lloyd's managing agents build governance frameworks as AI use moves from pilots into core underwriting workflows. Lloyd's syndicates and MGAs should align intake rules with delegated authority and line-specific appetite under capacity-provider governance.

Carriers building these controls for competitive reasons are simultaneously building examination readiness.

From document handling to decision advantage

Agentic AI lets commercial and specialty insurers move from manual, case-by-case underwriting toward what McKinsey calls a machine-first, human-governed model. Execution still lags ambition: Deloitte's survey of 200 US insurance executives found that lack of business-line support was the top reason generative AI implementations failed, ahead of data quality or technology gaps. Intake automation and pricing succeed or fail together, as one connected governance system.

Start where friction is highest. Measure how long submissions sit before an underwriter sees them, then choose a bounded first workflow with a measurable downstream outcome. Assign operational owners for exceptions and for model governance, including change logs, before scaling further.

How the hx platform connects intake to the pricing decision

For carriers that want intake data to flow into governed pricing, the hx platform is the agentic underwriting workbench for commercial P&C insurance, connecting ingestion and triage to pricing while retaining underwriter control.

hx's submission triage maps extracted submission data into governed appetite and workflow rules, then attaches model-driven indicative pricing at the point of intake. A risk that meets appetite can progress into full pricing with one click. Customers control schemas and workflow rules, including dashboards, per line of business, and the schema underwriters clear at intake is the same data actuarial pricing models consume, without rekeying.

Built-in audit trails and in-model safeguards, supported by API ingestion, cut rekeying, while stronger triage focuses underwriter attention on in-appetite risks.

Book a demo to explore how hyperexponential helps underwriting and actuarial teams connect submission intake and validation to triage and pricing decisions.