Use caseRetail banking
Every home loan gets the same completeness review, from a PDF checklist that changes twice a year. Each processor reads it a little differently.
The administrator writes the rules as plain sentencesEvery verdict names the rule and the document behind itThe processor keeps the final word
One rule needs a person · the AI names it, with its document
Eleven rules pass. One does not: three recent pay stubs are needed for each borrower, and Anthony's August pay stub is not in the file.
source · 9 documents on the applicationColumbus passes a file that Cincinnati sends back. Same bank, same checklist, the same PDF in the procedure folder, rewritten twice a year with an email from operations each time. Each processor applies the list from memory. The checklist exists. Nobody can say which version a given file was checked against.
Sophie Langford, who owns the procedure, writes the checklist once, as sentences. The assistant applies it to every application on arrival and writes a verdict with its reasons. She changes a sentence without a release, and the processor decides with the reasons in front of him.
Rules in a console, verdicts in a queue: neither screen is new. The order is what earns the trust: the rules are written in sentences, tried on closed files, approved by compliance, then run on every new application with a reason for each result. The steps below start with Sophie's edit on November 3 and end with her month-end report.
The procedure changed on November 1, and the administrator updates the checklist herself.
Sophie Langford runs credit operations for the bank's Ohio region. She opens the task Check completeness, home loan in the console. Its instructions hold eleven rules, each one a sentence. The new procedure wants pay stubs dated within 90 days, so she rewrites rule 4.
Some rules end the review. At this bank, the home loan covers a primary residence or second home; a condo bought to rent out goes through the investment property loan instead. Sophie writes it as rule 12, with the verdict it returns and the sentence the processor will read.
Loan files hold income, tax and identity data. Sophie and the IT team pick a model hosted on the bank's own servers for this task, and set it in the task. The routing rule on data residency keeps every call for this case type on premises.
The rules touch every application in the region, so they are tried on closed files and read by compliance first.
Sophie opens three applications the bank decided in October and runs Check compliance on each. Two verdicts match what the processors found at the time. The third flags a missing pay stub on a file the bank accepted.
The borrower works for a hospital that issues two pay stubs a month, one for salary and one for allowances. Sophie adds one sentence to rule 4 and runs the check again on the same file. It now returns Complete.
Sophie files the twelve rules as a document on the procedure case and launches its approval. Anne Doyle from compliance reads each rule next to the procedure section it cites, and approves. Only then does Sophie make the new instructions live.
On Monday a new application reaches a processor's queue, already checked.
Julie and Anthony Carver apply for $285,000 over 25 years for a condo in Columbus. The check runs as a step in the process and writes two fields on the case: a status and the reasoning. Here the status is Attention required: Anthony's August pay stub is not in the file.
Matt calls Anthony, who was on unpaid leave in August. He asks for the employer's letter, which arrives the same afternoon on the case. Matt records his decision in the case fields, with his reason, and sends the file to underwriting.
At the end of the month, Sophie opens her report of applications by AI verdict, then searches the files marked Attention required on rule 4. The processors' reasons repeat: unpaid leave, parental leave. She drafts a sentence on documented leave for the next approval.
The Carvers' application came back with one rule open. Every application gets one of three results, and the result decides what the processor does next.
Identity, income, tax return, agreement, purpose. The reasoning shows each of the twelve rules as passed.
Next: the processor checks and sends the file to underwriting.A pay stub missing, a signature absent. The reasoning says which rule, on which document.
Next: the processor requests the document or records his own reason; the check runs again when a document arrives.The AI lists what passes first, then names the rule that fails and the product that fits.
Next: the processor reviews it with the client and points to the right product.It refuses no loan. It sends nothing to underwriting. Across the nine steps on this page, every binding choice is made by a person: the administrator writes the rules, compliance approves them, the processor decides on each file.
The AI applies the sentences and writes its reasons. The control is in how the rules are written and approved, so it holds when volumes rise.
Counts from Sophie's November: one edited rule, one new one, three closed files. A longer checklist would call for more trial files.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
This one needs an administrator, a processor and a compliance officer on board, and each of them sees a different part of it.
When the procedure changes, you cannot tell who is applying the new version.
The missing page turns up while you prepare the underwriting file, days after the application came in.
Approving rules you cannot read is approval in name only.
Write the completeness review your processors do by hand as sentences, and run it on a handful of closed files before it touches a live one.
Your loan applications, documents and procedure stay put. The checklist runs inside the case you already use.
Let's talk
Bring your completeness procedure and a few anonymized files. We will write the rules with your administrator and run them on your closed files together.