Use caseDocument processing services
Northgate captures mail for an insurer, a utility and a lender, each on its own setup. Every new client costs weeks before its first batch.
One capture line, one configuration per clientOperators open only the fields the AI is unsure ofNothing leaves before the quality lead signs
The batch is split document by document and classified on Cedar Valley's own scheme: meter cards, ACH authorizations, complaint letters.
source · batch B-26-3108, 3,615 pagesA field is unsure or fails a rule · a person corrects it
Three contracts at Northgate Document Services mean three capture setups, three sets of templates and three sets of screens to learn. Clients now ask harder questions too: what share of fields a person checked, and how they know their documents never sat in another client's queue. Each new client is a new tool to learn, and each audit is a spreadsheet built by hand.
One capture line replaces the three, with a configuration per client: its own classification scheme, its own fields and rules, its own threshold and its own export. The AI classifies and extracts, operators check what it is unsure of, and the quality lead signs before anything leaves.
Every batch on this line passes the same gates: a batch belongs to one client from intake, the AI reads it with that client's rules, people spend their time on doubt, and the export waits for a signature. The steps follow one Cedar Valley batch through them, then watch a fourth client join.
Morning at Northgate: scanned mail, email attachments and files sent by clients, eleven batches by 8:30.
By 8:30, Kwame Asante, operations manager, has eleven batches waiting. Each client has its own intake: a scanner output folder, a mailbox, a partner API. Each intake feeds that client's capture chain, so a batch belongs to one client before any page is read.
The AI splits Cedar Valley's 3,615 pages document by document, then classifies each one on Cedar Valley's own classification scheme. A complaint letter is a complaint letter for Cedar Valley, and a claim form is a claim form for Larkfield.
On each document the AI extracts the values the client asked for, up to fifty per document, assigns each a confidence level, and checks them against the client's rules. On a handwritten meter card, the reading is clear, but the account number has a smudged digit and fails its checksum.
The AI handles the clear pages. People set the rules, check the doubt and sign.
Mei Chen, solution architect, keeps one capture configuration per client: its classification scheme, its fields, its rules, its confidence threshold and its model. Cedar Valley's contract keeps documents inside the network, so its configuration runs a small model on premises. She improves it with instructions and a few examples.
Sara Lindgren is on team blue, assigned to Cedar Valley Water and Brightway Lending. Her queue shows only those two clients; access is set by role and by client. She opens the flagged card, reads the digit on the image, types 1, and the checksum passes.
Before a batch leaves, Dana Hart, quality lead, runs the checks each contract asks for. Brightway's contract sends every loan agreement to a person whatever the confidence: reviews can be triggered by document type. Dana reviews the open points, then signs.
Signed batches leave for each client's system, each client reads its own figures, and a fourth one joins.
Once Dana signs, the operations engine loads the documents and their approved values into each client's repository. Cedar Valley Water runs SharePoint, Larkfield Insurance runs FileNet. Brightway Lending loads its own system from the Excel file of approved values.
Every Monday Kwame sends each client its report, built from the line's own records: documents processed, queues, delays against the agreed deadlines, fields corrected by a person. AI consumption is tracked by use, and each client is its own use.
Northgate signs Ashcombe Housing, an affordable housing provider sending lease forms and repair requests. Mei creates its configuration with two types and a few examples, sets its intake and its export, and runs 200 anonymized documents. Kwame reviews the result with Ashcombe's manager before go-live.
No document leaves on the AI's word. Each one lands on one of three paths, set by that client's threshold and rules.
Each value is above the client's threshold and every rule holds.
After that: the document waits for the batch sign-off.A value under the threshold, or a rule that fails. The flag says which.
After that: a validation task for the team assigned to that client.The document matches nothing on the client's scheme.
After that: an operator classifies it by hand, and it becomes an example.It signs no batch. It sends nothing to a client. It changes no client's rules. Across the nine steps on this page, the export starts only after the quality lead signs, with her name on the batch.
The AI splits, classifies, extracts and checks. The thresholds and the cases that need a person are set upfront, per client, so the control holds when volumes rise.
Batch B-26-3108 and the three contracts behind it supply these counts. Other clients and other forms will give you other ones.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
A processing center moves its clients onto one line once operations, the head of service and the people at the validation screen agree, each for a reason of their own.
One tool per client, and a staffing plan built from guesses.
Each new contract costs weeks before its first batch.
You reread clear pages to find the one unclear field.
With one client and one document type. Take the form your operators key most often and run a test batch of real, anonymized pages through its own configuration.
Then compare the flagged fields with what your team corrects today, before you move the next client.
Your move
Send us a sample batch per client and the systems they load into. We will run a test line with you: intake, validation, sign-off and export.