How AI Helps Underwriters Process Insurance Submissions

AI can help underwriters by assembling a submission into a reviewable file: identify the documents, extract specified fields, compare conflicting values, flag missing information, and prepare a handoff into the underwriting workbench. The useful outcome is less reconstruction before a person can assess the risk. Whether it saves time depends on the quality of the incoming material, the effort to check the output, and the connection to existing systems.

For a commercial carrier or MGA, submission processing is a sensible first candidate when staff repeatedly reopen attachments or rekey the same information. It is a weaker candidate when most of the delay comes from broker responses, limited underwriting capacity, or unresolved appetite decisions. Those constraints need their own changes.

The time is there to recover. In Accenture's P&C underwriting survey with The Institutes, conducted in 2021, the average underwriter spent 40% of working time on administrative tasks and another 30% on negotiation and sales support, leaving about 30% for underwriting itself. Submission preparation is where much of that administrative time sits.

Explore an illustrative submission handoff with missing-document and identity checks.

Give the underwriter a file they can interrogate

A useful submission view should show the named insured, requested coverage and effective date, document inventory, material extracted values, unresolved questions, and the next owner. Every material value needs a route back to its source. A polished narrative without that evidence makes review harder: the underwriter has to locate the original fact before deciding whether to trust the summary.

Cytora's intake lesson describes combining submission, internal, external, and inferred data before routing a file. That is a vendor description of a design pattern, not evidence of performance on your submissions.

Keep those data categories distinct in your own design. An address supplied by a broker, an address from the current policy, a third-party property attribute, and a model's inferred occupancy are different kinds of evidence. Displaying all four as if they were supplied facts conceals the disagreement the underwriter needs to resolve.

Separate reading, checking, and deciding

The following is a proposed operating design for submission preparation. Adapt the fields and controls to the line of business and your underwriting authority.

A checklist, date comparison, or duplicate check may need ordinary software rules rather than AI. Use document AI where variable layouts or wording make extraction difficult. Do not ask a language model to invent a rule your underwriting team has not agreed.

A conflicting property schedule is a better test than a clean application

This fictional example illustrates the review design; it is not a client case or a measured result. A property submission includes an application listing 12 locations, a spreadsheet listing 13, and a broker message saying one site has been sold. The loss runs cover an earlier period than the intake checklist requests.

A weak output silently removes a location, reports 12 as the confirmed total, and calls the submission complete. A useful output preserves the discrepancy: application count 12, schedule count 13, disposition of the sold site unconfirmed, loss-history period incomplete. It links the broker's message and both documents without deciding which is authoritative.

The intake coordinator can confirm which files were received and request the intended schedule. The underwriter decides whether the available loss history is sufficient to proceed and how the location change affects the assessment. After clarification, the accepted schedule becomes a new version; the earlier one remains available in the record.

The handoff is ready when the reviewer can see the accepted facts, the unresolved issues, and who owns each issue. It need not imply that every uncertainty has disappeared. That distinction allows a team to prepare a useful file without falsely certifying its completeness.

What regulators expect when AI supports underwriting

Using AI in underwriting is permitted, but the insurer remains accountable for it. The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, says decisions or actions made or supported by AI must comply with all applicable insurance laws and regulations, and it sets expectations for how insurers govern their use of AI.

As of April 1, 2026, 24 states and the District of Columbia had adopted the bulletin, including Illinois on March 13, 2024, and California, Colorado, New York and Texas had issued their own insurance-specific AI regulation or guidance. A preparation system that shows its sources, keeps decisions with authorized underwriters and records what changed is far easier to explain to a regulator than one that silently fills gaps.

Connect to the workbench without creating a second book

Start with the inbox or portal, document repository, clearance process, rating inputs, and policy administration system you already use. Identify the authoritative record for each field before mapping data between them. A submission identifier should survive a broker follow-up, a revised attachment, and a reassignment to another underwriter.

For an initial evaluation, place proposed changes in a staging view. Let staff accept them into the working record. If the system cannot complete a write, it should show a pending handoff with an owner; an extracted document should not disappear into a completed queue while its fields never reach the workbench. Test retries so they do not create duplicate submissions.

Require access by role and account, and review how the provider retains documents, uses them for model training, and supports deletion and export. Ask your compliance owner to define the obligations that apply to the actual product, states, data, and decisions involved. Calling a tool an intake assistant does not settle that assessment.

Measure the review burden as well as the extraction

Compare a representative set of files using your current process and the proposed process. Include scanned loss runs, unfamiliar broker templates, corrected schedules, multiple named entities, and incomplete submissions. Keep a separate test set that was not used to tune the workflow.

• Preparation effort: total staff minutes to produce an accepted handoff, including corrections and follow-up drafting.
• Material errors: wrong entities, dates, location counts, amounts, or periods that would change the next action.
• Missed gaps: files called ready despite a missing requirement, including a sample of apparently clean outputs.
• Queue performance: time from receipt to an assigned reviewer, with time waiting on brokers reported separately.

Have the underwriting owner set acceptable errors and stop conditions before the trial. A high average field-accuracy score can hide a wrong insured or a material schedule omission. Faster preparation also does not prove better risk selection, a higher quote rate, or lower losses; those outcomes require their own evidence.

The broader insurance operations guide connects submission work with service and renewal queues. Within underwriting, the first decision is more specific: can staff accept a source-linked preparation output with less total effort and sufficient visibility into what remains uncertain?