The AI-Powered Financial Document Reader

The situation

This was part of a bigger move: the client was migrating their entire platform off a proprietary system onto a modern database (Supabase/PostgreSQL). Buried inside the old platform was a built-in document scanner, a "black box" nobody could inspect or improve, that manually extracted numbers from uploaded bank statements, alongside a hidden email connector doing something similar. Someone still had to check the scanner's work by hand.

The bigger migration also meant moving years of tax, invoice, and ledger history from a loose, flexible old-style database into a strict one with real relationships between records. Every ID, every link between an invoice and a customer, had to survive the move intact.

What we built

A custom AI reader built to replace the old black-box scanner entirely. Every uploaded bank statement gets opened, read, and the exact transaction data (amounts, dates, descriptions) gets pulled out and written straight into the new database.

How it works

1. 01A bank statement is uploaded into the platform.
2. 02A background function sends the document to the AI, along with strict instructions for exactly what shape of data to hand back, not just "read this," but "give me these exact fields, every time."
3. 03The AI extracts the transaction details and returns them in that exact structure, so the numbers always land in the right place instead of needing a human to double-check them.
4. 04The structured data is written into a secure database record that only that one client's account can ever see.

The challenges we solved

• Legacy documents came in inconsistent formats, so the reader had to be told exactly what shape of data to hand back every time, or the results would come back messy.
• The company's financial records were moving into a completely rebuilt, relational database at the same time, so the reader had to feed the new structure, not the old freeform one, without losing any historical data in the process.
• Every business using the platform needed its own private, walled-off data: one account could never see another account's numbers, including inside the new AI pipeline.

The result

The manual data-entry step disappeared completely. Every bank statement now gets read, structured, and filed automatically with the same accuracy every time, freeing the team from a task that used to eat hours every week, and putting them in full control of a process they used to depend on a black-box vendor for.