The problem
A mid-market mortgage and finance broking firm was assembling every loan-application pack by hand. For each application, a broker or their support staff collected the documents a lender needs — payslips, identity documents, bank statements, rates notices — checked each one was present, current and legible, keyed the key figures into the lender's portal, and lodged the pack.
The work was slow and easy to get wrong at the margins. Preparing and checking a single pack took an experienced broker around forty minutes. Worse, an incomplete or out-of-date document that slipped through came back days later as a lender "more information" request — which stalled the application, frustrated the client, and pulled the broker back into a file they thought was closed.
At their volume the firm was facing a choice between adding administrative headcount and capping how many applications each broker could carry. They wanted neither.
What we did
Four weeks of build, three weeks of pilot on a subset of brokers, three weeks of staged rollout. The deployed system:
- Read each uploaded document and classified it by type (payslip, driver licence, passport, bank statement, rates notice, and so on) using a document-intelligence model
- Extracted the key fields per document with a confidence score, and flagged anything expired, illegible, or below the confidence threshold
- Checked the assembled pack against the relevant lender's document checklist and listed exactly what was missing or stale before anything was lodged
- Produced a lodgement-ready summary the broker reviewed and signed off — the system never lodged anything on its own
The pipeline ran inside the firm's existing environment. No client document left their control, and every classification and check was recorded so a completed pack could be reconstructed later.
The hard part was not the extraction. It was the checklist logic: lenders differ, and the same document means different things depending on the product. We spent two weeks with the firm's most experienced broker turning "what a good pack looks like" into rules the system could apply — and, just as importantly, into the cases it should hand straight back to a human.
The outcome
| Before | At 6 months in production | |
|---|---|---|
| Time to prepare and check a pack | ~40 minutes | ~8 minutes (including the broker's sign-off) |
| Lender "more-information" requests | Frequent | Materially fewer |
| Applications per broker | Capped by admin time | Higher, with no added admin headcount |
| Administrative FTE added despite growth | Would have needed more | 0 |
What's interesting
The win was not "AI read the documents faster". It was three quieter things:
- The completeness check. Catching a missing or expired document before lodgement — not after the lender bounced it — removed most of the rework that had been eating broker time invisibly.
- Confidence-scored routing. The files that went to a human were the files that genuinely needed one. The routine, clean packs went through quickly, and brokers stopped treating every file as if it might be the problem one.
- Broker sign-off stayed. Nothing was lodged without a person confirming it. That kept the broker accountable for their own client relationship — which is the part of the job that was never the bottleneck.
What we'd do differently
We would build the lender-checklist rules with more than one broker from the start. The rules we wrote with a single senior broker were excellent for how they worked, and needed a second pass to fit the rest of the team. Capability transfer across the whole desk — not just the expert — is the part worth planning for early.
We did not replace the broker's judgement, and we did not lodge applications automatically. We took the assembling and checking off the desk so the team could spend their time with clients, not with paperwork.