The Loss Run Arrives as a Scanned Fax With a Handwritten Note in the Margin - and Someone Keys It Into PolicyCenter

Forty-one minutes. That is a fair figure for what it takes an underwriting assistant to turn one small-commercial submission from a broker into a cleared, keyed record in PolicyCenter — and almost none of those minutes are underwriting. They are retyping.

The submission came in as an email at 8:02 with four attachments: a scanned ACORD 125 with the applicant section filled in by hand, an ACORD 126 for the general liability, a five-year loss run from the expiring carrier that was clearly faxed to the broker and then photographed, and a statement of values spreadsheet with 41 locations and a merged header row. Somebody has to get all of that into a system that will only accept clean fields.

The submission does not arrive the way the software expects it

Commercial insurance still runs on documents designed for a fax machine. The ACORD forms are standardised, which helps, but the copies are not: printed, signed, scanned at 200 dpi, often annotated. The loss run is worse. It comes from the incumbent carrier in whatever shape that carrier's claims system spits out — sometimes a proper PDF, frequently a scan of a printout, occasionally with a handwritten line in the margin reading "per phone call w/ Marcy 4/12 — claim reopened, reserve now 118k."

That margin note is the whole account. A five-year loss run showing $34,000 incurred is a different risk from one showing $152,000, and the difference lives in pen ink no field on the form was going to capture. The assistant either catches it or does not.

Here is the clean version worth quoting: multimodal document reading means the software looks at the page as it actually is — the fax streaks, the crooked scan, the pencil in the margin — and pulls the fields out of the image, instead of demanding a tidy electronic file that commercial insurance almost never receives.

What the retyping actually costs, beyond the retyping

Ask any commercial lines manager for their hit ratio, then ask for their declination-for-no-response rate. The second number is the embarrassing one. Submissions age. A broker with a 1 September effective date sends the same account to four carriers on 12 August; the carrier quoting on 14 August is in the conversation and the one quoting on 22 August is not. In the 1 January and 1 July renewal crushes, when a small-commercial team sees three times its normal volume in ten working days, the aging gets worse exactly when the premium is thickest.

So teams triage. They take the submissions from the agencies that bind, they let the rest sit, and everyone agrees this is fine. It is not fine — it is a hidden underwriting rule that nobody wrote down and nobody can audit, and it means your appetite is being set by the speed of your data entry rather than by your rate.

What changed: readers that cope with the bad copy, not just the good one

Optical character recognition has been in carriers since the 1990s, and every underwriting manager has a story about the zone-based reader that worked beautifully on the vendor's sample pack and fell apart on the first real fax. The difference in 2026 is that the page no longer has to sit in a fixed position. The reader takes the whole page the way a person does, treats a column headed "Incurred" and one headed "Total Inc." as the same thing, notices the handwritten note belongs to the claim on line six, and says so.

It also handles the spreadsheet nobody wants to open. A broker's statement of values is not a table, it is a 41-row negotiation with merged cells, a subtotal in the middle, construction described as "JM/frame w mansard" and a protection class typed into the wrong column. You can hand it the entire workbook at once, ask for each location's total insured value, construction class and sprinkler status, and it will show which cell each answer came from.

flowchart TD
  A["Broker email lands 8:02 a.m."] --> E["Reader takes each attachment as it sits"]
  B["Scanned ACORD 125 and 126"] --> E
  C["Five-year loss run, faxed then photographed"] --> E
  D["Statement of values, 41 locations, merged cells"] --> E
  E --> F{"Figures disagree or handwriting present?"}
  F -->|Yes| G["Held for the underwriting assistant with the page marked"]
  F -->|No| H["Cleared and drafted into PolicyCenter"]
  H --> I["Underwriter opens a submission already priced-ready"]

Tuesday, 8:20 a.m., in the small-commercial queue

By 8:20 the four attachments have been read. Applicant name, FEIN, mailing and location addresses, years in business, prior carrier and expiring premium are drafted into PolicyCenter. Classification is proposed with the ISO general liability class code and the NCCI code for the workers' compensation piece, each beside the sentence from the application it came from. Clearance has already run against the book to catch the same account submitted by two agencies.

The loss run has been turned into a table: five policy years, claim count, paid, incurred, open and closed, with the valuation date pulled off the header. The margin note is flagged in red with the scanned line reproduced underneath, and the incurred figure for that claim is marked "disputed — handwritten amendment, not reflected in typed total." The assistant's job on this account is now four minutes: confirm the handwriting, decide whether to ask the broker for a re-run valued this month, and release it.

The underwriter opens it at 9:00 with the loss history in a shape she can price against, not a PDF she squints at. On a hail-exposed property account she wants the statement of values, the roof ages and the construction, and those are drafted too, cell references included. What she does not do is retype.

The arithmetic of one loss year that got keyed wrong

Time savings are the boring half. The expensive half is error. Assume a small-commercial team handles 900 submissions a quarter, and three per cent carry a keying error in the loss data — a year omitted, an incurred figure read as paid, a reopened claim missed. That is 27 accounts. Assume half are material enough to change the price, and that a mispriced account comes out $2,400 light against what your own rate would have produced.

Halve the error rate and you have recovered $67,000 of premium that was already yours. Add the 420 assistant hours a quarter and the argument stops being about software cost. And note which way the leakage runs: mis-keyed loss data almost always makes an account look cleaner than it is, because omissions are more common than inventions. You are not losing accounts to this. You are winning the wrong ones.

The three things a person still initials

Classification is the first. A reader can propose a class code and quote the sentence it came from, but "concrete contractor" on the application and what the insured actually does on a job site are two different facts, and the second one is a phone call. Getting the code wrong on a workers' compensation account is not a data problem, it is a premium audit dispute eighteen months later.

The second is the valuation date on the loss run. A run valued eight months ago with three open claims is a different document from the same run valued last week, and the reader will tell you the date but it cannot tell you whether the reserves have moved since. Somebody has to decide whether to ask for a re-run, and on a large account that decision belongs to the underwriter.

The third is anything the reader marked low confidence. Resist auto-accepting those to clear the queue during the 1 January rush. The whole value of the reader is that it flags its own uncertainty instead of quietly guessing, and that value evaporates the week somebody turns the flags off to hit a volume number.

Frequently asked questions

Will it read a loss run from a carrier whose format we have never seen?

Yes, and that is the real change from older readers: there is no template to build per carrier. What still trips it is a run photocopied so many times the numbers have closed up, and there it should say so rather than return a figure. Test it on your worst twenty PDFs first.

Do we have to change anything in PolicyCenter or Duck Creek?

No structural change. The practical work is deciding which fields get drafted automatically and which stay blank until a person fills them, and building the review screen where the assistant sees the drafted value next to the image of the page it came from. That side-by-side view is the single most important design decision in the whole project.

What about the broker's own submission system?

Agencies on Applied Epic or EZLynx can send you structured data, and where a broker will do that, take it. But the loss run they forward almost never originates with them — it comes from the incumbent carrier and arrives as an image no matter how modern the agency is. That is why the reader matters even for your most digital producers.

How do we know it did not just make a number up?

Require the citation. Every drafted field should carry the page and the location it came from, so the assistant is confirming rather than trusting. If a vendor cannot show you the drafted figure and the source image on the same screen, that is the demonstration to ask for before anything else.

Where to start, and with how many

Take fifty submissions you have already processed — from a busy week in the last renewal season, so the ugly documents are represented. Run them and compare field by field against what your assistant keyed. You will find two things: places the reader is wrong, and places your keyed record was wrong and nobody knew. The second list is usually longer, and it is the one that funds the project.

One related pressure worth naming: brokers who get answers faster call more often — about status, about a missing schedule, about whether you will look at a risk. CallSphere builds AI voice and chat agents that pick up the agency line and the web chat, take the account name and submission number, answer the where-is-it questions from what you have on file, and book the underwriter call when a person is genuinely needed. Faster reading is worth less if the phone rings out at 4:40 on a Friday in December.