[ Bank · mortgages and personal loans ]


From 40 days to 19 minutes to pre-approve a mortgage

19 min

In short

The problem
Mortgage and loan applications arrive all the time, and advisors spent half-days on files that did not meet the requirements to go ahead.
What we built
An agent reads the application and its documents as they arrive, checks the prerequisites and passes to the advisor only the files that can move forward.
The impact
From application to pre-approval goes from 40 days to 19 minutes.

The context

The bank receives mortgage and loan applications all the time, through different channels. Each one arrives with its own bundle of documents, filled in however it happens: from income to bank statements, sometimes scanned crooked, sometimes incomplete.

Until then, every application landed on an advisor’s desk. The advisor opened it, read the attachments, checked the data one by one and only at the end worked out whether the file met the requirements to go ahead. Very often it did not. A document was missing, or a figure did not add up between one form and another.

The result was a long, slow queue. Advisors’ time went mostly on files that would never have gone ahead, read with the same care as the good ones. People with a valid application waited weeks for an answer, and in the meantime had all the time they needed to ask somewhere else.

We wanted the advisor to see only the files worth working on, and the customer with a valid application to know within minutes.

  • 99.7% accuracy in recognising the files that hold up
  • The advisor receives the file with the checks already done
  • +340% conversion from lead to customer

How it is built

Reading documents the way an advisor reads them

The first step is extraction. The agent opens every attachment, recognises what kind of document it is and pulls out the data the assessment needs: amounts, dates, account holders, deadlines. Then it compares them with each other, because an inconsistency between two documents in the same file is exactly the kind of detail that, by hand, on the thirtieth file of the day, slips through.

The bank's criteria, applied the same way every time

With the data in hand, the agent checks the prerequisites the bank uses to decide whether a file can go ahead. They are the bank's criteria written down, and they are applied to every application in the same way, the first of the morning like the last of the evening. If something is missing it flags it straight away, instead of leaving an advisor to discover it after half an hour of reading.

The decision stays with a person

The agent approves and rejects no mortgage. It prepares the file, clears off the table what cannot go ahead and passes to the advisor only what holds up, with the checks already done. In a bank that is the condition for a project like this to exist, and it is also what makes it defensible in front of whoever answers for compliance.

How it is built in Microsoft 365

Brought into Microsoft 365, the pattern rests on a few pieces. An agent in Copilot Studio acts as the orchestrator, AI Builder document processing models read the attachments, Power Automate moves the verified file into the advisor's queue, and the documents stay in SharePoint with the permissions and sensitivity labels of Microsoft Purview. No data leaves the perimeter the bank already governs.

The pilot started in Italy. Accuracy in recognising valid files reached 99.7%, above manual review.

19 minfrom application to pre-approval, down from 40 days
99.7%accuracy, better than manual review
+340%conversion from lead to customer

What changed

From application to pre-approval goes from 40 days to 19 minutes. For applicants who meet the requirements, that means knowing almost immediately whether they can go ahead, with no time to turn to another bank. For the bank it means conversion: +340% from lead to customer.

Advisors, meanwhile, spend their days on the files that make sense and on the customers who can close. And compliance risk goes down, because the basic checks are always done, and always done the same way.


Other case studies