Use caseInsurance mailroom
Monday brings 4,000 items by paper, email and portal. Each needs a type, a case and a team, and the morning goes into reading envelopes.
Every channel lands in one sorting flowEach item reaches its case on its contentA person sorts what falls below the threshold
A plumber's invoice for a burst pipe in a kitchen. Type Repair invoice, confidence 97 percent, above the threshold for that type.
source · page 1 of the scanOne point needs a person · the operator confirms or corrects
A regional insurer receives about four thousand items on a Monday. The paper is scanned overnight, the emails sit in the claims mailbox, the portal uploads in a third queue. The hard part is the first reading: finding the policy number on page two, spotting a medical report inside an envelope, guessing which team owns a letter that names three contracts.
One flow now takes the first reading. Every channel feeds the same capture chain, the AI gives each document a type and a case on its content, and a threshold per type decides what goes straight through. The operators spend the morning on the items below the threshold, and decide each one.
Seen one at a time, the screens look familiar. Put them in order and the point shows: every channel is read the same way, each document reaches its case before the teams log in, and a person sorts what the AI cannot place. The walkthrough opens on the overnight batch and closes on the noon review of the corrections.
Before anyone reads a word, paper, email and portal uploads are gathered and split into documents.
The scanning provider dropped batch B-1145 in a shared folder overnight: 1,262 envelopes, each one a single file that may hold a letter, an invoice and a photo. The capture chain picks the batch up and splits each file document by document.
The claims mailbox holds 418 new emails and the customer portal has 301 uploads. Each email becomes an inbound item: the message is kept as a document, each attachment becomes its own. Portal uploads arrive through the API with the customer's account, so they already carry a customer number.
The AI classifies each of the 3,107 documents on the insurer's own scheme of 46 types: claim declaration, cancellation request, change of banking information, medical report, repair invoice, complaint. Each type comes with a confidence level.
The threshold is set per type. A complaint needs 90 percent, because a missed complaint starts a response deadline. Below its threshold, a document waits for a person.
A typed document still needs an owner. The AI finds the case on the content, the process finds the team, and the lead sees what is left.
On the letter from James Porter, the AI reads policy HAB-55120-02 in the reference line and claim SI-26-04471 in the second paragraph. The repository confirms the claim is open on that policy. The letter is attached to it, whatever the file was called.
The claims process routes by document category, product and region. A home claim document about a house in Columbus lands with the home claims team there. A cancellation goes to customer service, a change of banking information to the back office.
Medical reports go to a team whose rights are restricted to that document type. Nobody else sees them in a queue.
Nadia, who leads the mailroom in Columbus, opens her home page at 8:30 a.m. One report counts the morning's documents by result. 296 wait for a person, and the possible complaints are at the top because of their deadline.
296 documents below the threshold. Nadia's operators clear them before eleven, in three moves.
D-30414 is a letter about a five-week repair delay. It calls the situation unacceptable and asks for nothing specific. The AI proposed Complaint at 71 percent, under the 90 percent threshold, so it stopped.
Nadia reads it in the viewer and confirms the type. The letter goes to the complaints desk with its arrival time intact.
D-30502 is a plumber's invoice for Rachel Lawson, who has two open claims: a water leak in March and a broken window in October. The AI proposes the water leak, because the invoice describes a pipe and dates the work March 18. It leaves the choice to Nadia.
At noon, Marcus, the capture administrator, reviews the morning's corrections. Eleven letters typed as Reminder were changed to Complaint by the operators. He adds three of them as examples to the complaint type, and adds a route for the pet insurance product launched in November.
By 10:52 each of the 3,107 documents sits in one of three states, and each state sends it somewhere else.
The type is above its threshold and one case matches. The document is filed in its case with its fields.
Then: the team works it from its own queue.A type below threshold, or two cases that fit. The queue says which, with the line it rests on.
Then: the operator confirms or corrects, and the item moves on.No policy or claim found for this customer. The AI lists what it read and where it looked.
Then: a person opens a new case, or sends the item back to the customer.It opens no file on a guess. It picks no case between two candidates. It changes no setting. Across the nine steps on this page, every uncertain item is settled by an operator, with her name on the decision.
The AI reads, types and proposes. The thresholds and the routes are written by people, so the mailroom keeps control when Monday volumes double.
Taken from Monday's batch in Columbus. Your mail mix and your list of types will give other numbers.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
The mailroom, the capture administrator and IT each have a reason to want this, and a reason to doubt it.
Mornings go into opening and reading, and the hard items come last.
Every new product or form needs a sorting rule that nobody wrote down.
Three channels, three tools, three ways of failing.
With one channel and ten document types. Take the scan batch, type it, and let the operators sort what stops. A few weeks of real mail will tell you whether it holds.
The other channels join the same chain afterwards, with the same types and the same routes.
See it live
Bring an anonymized sample of a day's inbound mail and your list of document types. We will read it the way this page does, then sort what stops with your operators.