Three separate apps, no connection between them, and hours of copying the same thing by hand.
A mortgage broker's job is to help clients get the best loan. But a huge chunk of their day was just moving data between software tools that didn't talk to each other. By the time they'd filled everything in everywhere, they'd already spent hours on admin work before making a single real decision.
Reconstructed from broker interviews. Some were doing this on every single application.
The AI could read documents and pull out numbers. But brokers couldn't tell which numbers to trust.
When a broker uploaded a client's income slip or appraisal, the AI would scan it and fill in the numbers. Most of the time it was right. But sometimes it wasn't, and there was no way to tell the difference.
So I added a confidence score to every field. A small, color-coded signal right next to the value. Green means the AI is confident. Amber means take a look. Red means you need to fill this in yourself.
The broker stays in control. The AI earns trust by being honest about what it doesn't know.
Compliance checks used to happen at the very end. By then, it was expensive to fix anything.
In mortgage lending, there are rules about what you can and can't submit to a lender. Before Autowrite, brokers only found out they'd broken one of those rules after all the work was done. Then they had to start the whole application over.
I made compliance its own tab, always visible while the broker was working. Problems showed up early, when they were still small and easy to fix.
Brokers were tracking their deals in spreadsheets. We built a better home screen.
The pipeline view became the first thing brokers saw when they logged in. Every active deal in one place, with its current stage, outstanding flags, and last activity. Scannable in seconds.
The goal was to replace the spreadsheet brokers were maintaining on the side, not by asking them to change behaviour, but by making the app more useful than the spreadsheet.
The original form asked for everything upfront. Most brokers gave up halfway through.
The intake form had every single field on one screen. You couldn't see any AI output until the whole thing was done. It was a lot to take in, and brokers regularly dropped off before finishing.
I split it into three stages: who the borrower is, what they earn, and what property they're buying. Complete stage one, the AI starts working. You don't have to finish everything before you see any value.
When brokers stopped using their spreadsheet, we knew it was working.
The deals list shows every client in one place: where their application is, what still needs doing, and any flags the AI raised. Each row is readable in a couple of seconds.
A few brokers told us they stopped maintaining the spreadsheet they used to keep on the side. That was the real sign it had clicked.
The hardest part of this project wasn't the UI. It was figuring out how much a broker should have to do manually, and how much it was safe to hand off to the AI, without them feeling like they'd lost control.
Four weeks isn't a lot of time to figure that out in a regulated industry neither of us had worked in before. But that constraint also forced good decisions. Every design choice had to be justifiable in one sentence.
Want to hear the
real story?
Four weeks, a lot of fast decisions, and a domain I had never worked in before. I'd rather walk you through it properly.