Building manager’s certificates, read by machine

The problem

A building manager’s certificate is written to be read by a person. It arrives as a scan, and its layout depends on whoever issued it. The parts the bank needs, the housing company and its identification number, the loans, and the renovations already carried out, sit wherever that particular template puts them.

So the bank did what everyone does. Someone opened the scan on one screen and typed into the other. Transcription adds nothing to a customer’s case, and here it happened with the customer sitting on the other side of the desk, watching it.

What we built

A smart scanning system for this one document type, installed inside the bank’s own private cloud. Nothing leaves the bank’s infrastructure, which was a condition rather than a preference.

It runs on our own extraction engine, so setting it up means describing the document rather than labelling thousands of copies of it. For each field we write a few lines saying what that field looks like. The company identification number probably has the words “company ID” near it, matches a known pattern, and has to parse as an ID. The housing company name probably has “name” or “company” nearby, in Finnish or in Swedish, so the Swedish keywords go in as well at a slightly lower weight.

Then the fields check each other. Housing companies and their identification numbers are public record, so the system compares the name it read against the number it read, and both against the official registry. A pair that disagrees is worth flagging before it reaches a customer’s file. The machine learning stays inside the engine; what we write down is the structure.

What changed

• 80% less manual typing to serve a mortgage or real estate customer.

• Around 12 minutes saved per customer and certificate.

• Certificate data arrives in structured, database-ready form instead of a retyped approximation of it.

• The awkward fields came out too. A renovation history is a list inside a document, and it is extracted as a list.

Twelve minutes is not a dramatic figure on its own. It is twelve minutes on every certificate, for every customer, handed back to the conversation instead of the keyboard.

Why small data

The usual recipe for document automation is millions of documents and a team of data scientists per document type. It does not scale down to one certificate used by one bank, and it responds slowly when a template changes.

We trained on 848 documents. Structural hints do the work that volume would otherwise have to do, which is what makes eight weeks realistic and what keeps the accuracy where a bank needs it.