Case study · Credit and lending
Turning a bank statement into a cashflow report.
A tool that takes a hard-copy bank statement, digitises it, and turns it into a finished cashflow report, giving reliable insight into credit behaviour when validating originations.
- Bank formats
- 11
- Countries
- 3
- Reports generated
- 2,300+
The problem
Credit analysts assessed lending applicants by reading client bank statements by hand. Often hundreds of pages, across 11 different bank formats and 3 countries. They reconstructed turnover, cashflow patterns, returned payments and funder relationships line by line. On a large statement, capturing and analysing it could take an analyst a couple of days, before any credit thinking started.
What we built
An AI platform that reads bank statements for credit assessment. It extracts every transaction from each statement and produces a finished cashflow analysis report, without a human transcribing a single line.
The same depth on every deal
Scale is not limited to short, simple statements. A 400-page statement gets the same treatment as a 20-page one, so the depth of analysis no longer depends on how much time an analyst can spare.
Secure by design
The whole thing runs on a workstation inside the client's network, including the language model. Nothing is sent to an external service, which is what made it acceptable to put real customer bank data through it.
Where it got to
More than 2,300 credit reports have been generated automatically. Each one covers the information needed to reach a credit decision, with an AI-drafted recommendation for the analyst to review and confirm. Volume is five times what it was, on the same team and the same running cost.
What it delivered
- 5x
- Volume has grown five times over with no added headcount and no increase in running cost.
- Under 10 min
- A finished cashflow report from a raw statement, in minutes rather than days.
- On-premise
- Deployed on a workstation inside the network, language model included. No client bank data leaves the building.
Got a problem that looks like this?
Most of what we build starts with a conversation about what is not working yet.
Start a conversation