AI email filing for tax advisors at Crowe Peak (now OneLane)

The opportunity

Tax advisors at Crowe Peak spend a quiet but real part of their day on admin: reading each incoming email, working out which client project it belongs to, and filing it and its attachments in the right place. It takes around five minutes per email. Multiplied across a firm, that is hours a week of senior, billable people doing work nobody hired them for.

My role

I was one of five engineers who built the product from scratch over ten weeks, as the software project of my degree at TU Delft, with Crowe Peak as the client. We all worked across the whole system.

The part I built on my own was the model-agnostic classification layer. It routes each email to one of about 2,800 SharePoint destinations, and the model behind it can be swapped without touching the rest of the system.

Strategy and key decisions

The hard constraints weren’t technical, they were human. We agreed on four rules early, and every decision followed from them:

1. Fast enough that people actually use it. If filing with the tool felt slower than doing it by hand, adoption would be zero, however clever the AI.
2. Fit the existing workflow. Advisors live in Outlook, so we built an Outlook add-in instead of a separate app. Nobody had to change how they work.
3. Maintainable without us in the room. The team would leave after ten weeks, so the system had to be simple for the firm’s own people to run.
4. The human stays in control. The AI suggests where an email should go; the advisor confirms. With client data on the line, trust beats full automation.

The project also sat between three stakeholders with different priorities. Crowe Peak needed something that fits real consulting workflows, TU Delft held us to academic rigour, and end users just wanted something that didn’t slow them down. Balancing those taught me as much as any architecture decision.

Execution

We built the add-in inside Outlook. It reads each email, classifies it through the model-agnostic layer, matches it to the right client project, and suggests the destination for the advisor to accept.

• Accuracy: 72–89% top-3 accuracy per field.
• Speed: a median latency of 2.0 s, against a 10 s target.
• Documents: our native PDF approach was 5× faster than Azure Document Intelligence, and more accurate.

Results

Filing an email went from about five minutes of manual work to under thirty seconds. We handed over a working product at the end of the ten weeks.

What I learned

Good engineering starts from the people who will use the product. The most useful decisions on this project were the four rules we wrote before writing any code.