ACORD and loss run document AI: extract, review, and route without replacing judgment

Document AI should read a packet, pull candidate fields onto a review screen, flag missing pages, and route a task to the right person. That is logistics. It should not decide that a loss history is acceptable, that a driver is eligible, or that a limit is enough. Those are licensed judgments.

The useful mental model is a prep desk. A person used to open the PDF, write names and dates on a sticky, and pass the file. Extraction can do the sticky. The person still looks at the file before the sticky becomes the record.

Do not use a general chatbot as the document system. Pasting an ACORD packet or a loss run into a consumer window creates a data-control problem and a weak audit trail. An approved extractor with retention you can explain is a different tool from a writing assistant.

Do not auto-issue certificates, auto-submit applications, or auto-send a client a summary of their losses. Routing to a human is the product. Skipping the human is how a confident wrong VIN becomes a real AMS problem.

On ACORD-style applications, extraction often helps with named insured, mailing and garage addresses, FEIN or other business identifiers the agency already collects, vehicle year/make/model/VIN candidates, driver names and license states, and coverage-request checkboxes as they appear on the form. Each of those is a candidate until a person confirms.

On loss runs, useful extraction is structural: carrier name, named insured as printed, policy period, number of pages, whether three years appear to be present, and line items the reviewer can scan. Characterizing frequency or severity in narrative form is not a first workflow. Completeness flags are.

Handwriting, photos of screens, and scanned packets with stamps will fail more often. The workflow should expect low-confidence output on those files and send them to a slower human path instead of guessing. Guessing is how wrong garage ZIPs enter the system.

Keep original files in the AMS or document store the agency already uses. Extraction is a layer on top of the packet, not a replacement for the packet. If the extracted VIN and the PDF disagree, the PDF wins until a person says otherwise.

Set a review rule before you set a confidence number. A practical default is that every extracted field is unconfirmed until a person accepts it. Higher-confidence fields can be visually grouped so reviewers move faster. They should still be accepted, not silently written.

Low-confidence fields should block writeback. Missing VIN characters, mismatched named insureds, and loss-run years that do not add up are stop conditions. The task becomes “CSR to read the PDF,” not “system to pick the most likely string.”

Measure how often reviewers change fields. If they always correct garage address, the extractor or the packet quality needs work. If they never look at high-confidence VINs, slow the workflow down until looking is the habit. Rubber-stamps are not review.

Coverage and premium fields deserve extra friction even when the extractor is confident. A requested limit on an ACORD is still a request. It is not a bound coverage. Do not let writeback language imply that the agency has agreed.

Route by product and owner, not by whichever inbox is open. An ACORD for commercial auto should create a commercial auto intake or renewal task with the assigned producer. A loss-run packet should attach to the same account’s outstanding-document checklist. Random document libraries without owners are still piles.

Match to an existing client when you can, and require a human merge when you cannot. Creating a second named insured because the extractor read an extra DBA line will split the file. Matching on phone and email plus a person confirming the account is safer than aggressive auto-match.

Keep certificate requests on a different task type from new-business ACORs when that is how the office already works. Speed work and submission work compete on the same desk if you route them as one “document” blob.

Notify the owner in the channel they will see. A quiet AMS attachment with no task is how packets wait until someone asks where the loss runs went.

Write back only confirmed fields, and write them to the fields the agency already trusts. Do not create a shadow set of “AI columns” that nobody reports on. Do not overwrite a producer-corrected VIN with a later re-extraction unless a person triggers it.

Stamp the record with source and reviewer: extracted from file X, confirmed by Y, on this date. That stamp is how you reconstruct a bad value later. Silent overwrite is how dirty data becomes mysterious.

Leave narrative claim comments in the PDF unless a person copies a needed fact into a note. Bulk writeback of loss descriptions is how sensitive detail spreads into reports and prompts that did not need it.

If the AMS rejects a field format, fail visibly. Do not truncate a VIN to make the save succeed. Visible failure creates a task. Quiet truncation creates an E&O conversation.

ACORD packets and loss runs include personal and business details you would not paste into a random tool. Use an approved vendor, know where files live, how long logs remain, and how deletion works. If those answers are vague, the extractor is not production-ready.

Limit access to the review queue. Not every producer needs every packet. Role-based access should match how the agency already handles applications and claims documents.

Do not keep extra copies on laptops, group chats, or export folders after writeback. The AMS copy plus the vendor’s documented retention is enough. Extra copies are how retention policies fail.

Redact before any secondary summarization. A routing note can say “three-year loss runs attached, CSR confirmed page count.” It does not need to restate medical or claimant detail in a prompt.

Pilot one document type. Loss-run completeness checks or ACORD vehicle schedules are better first pilots than “every PDF that hits the inbox.” One type teaches confidence, routing, and writeback without boiling the ocean.

Use real packets from the last month, redacted if you are in a vendor demo, and time the human review. If review takes longer than opening the PDF, the screen is wrong. If review is instant because nobody reads, the control is wrong.

Define done: packet stored, fields confirmed, task routed, no silent overwrite. Expand to a second document type only when CSRs use the queue without a workaround.

If you want help choosing the first document type against intake and renewal leaks, book the free 15-minute workflow audit. Document AI is usually a second or third build, after unowned leads and missing-document follow-up are visible.