Casefile checks documents in four stages, and only the last one uses a model. Nothing is checked by hand, and nothing leaves the United Kingdom. This page explains exactly how, for brokers, compliance teams and applicants who want to know.
Each stage only passes on what it could not settle. A statement downloaded from a banking app is usually finished by stage three.
File type, size, page count, password protection, corruption, and whether the same document has been uploaded twice. Instant, and it catches problems before anything else runs.
A PDF downloaded from a banking app or from HMRC carries real text inside it. We read that directly. No image recognition, no model, and far more accurate than anything that interprets a picture.
Sort codes, account numbers, dates and statement periods follow predictable patterns, and the major UK banks, HMRC and the common payroll providers each have a recognisable layout. We also read the file's own metadata, which is how editing and tampering tend to show up.
Photographs, unusual layouts and poor scans go to an AI model. It runs on Amazon Bedrock in the London region, so the document stays in the United Kingdom. The provider does not keep it and does not train on it.
Reading a document is only half of it. The checks that matter compare documents against each other, and that part is arithmetic, not AI.
Net pay on the payslip, matched to the credit on the bank statement. Once both numbers are read, comparing them is a subtraction.
Direct debits on the statement and agreements on the credit report, checked against the fact find. A list comparison.
Name, date of birth and National Insurance number checked for consistency between documents. String matching.
The period on the document against the period the broker requested. A date comparison.
Every result records which stage produced it, so any check can be traced back and reproduced. The same document always gives the same answer.
Sub-processors and locations. We tell brokerages before any of this changes.
| What | Who | Where |
|---|---|---|
| Uploaded documents | Amazon S3 | London, eu-west-2 |
| Database | Amazon RDS, PostgreSQL | London, eu-west-2 |
| Stage 4 document reading | Amazon Bedrock | London, eu-west-2 |
| Sign-in, names and email addresses | Clerk | United States |
| Invitation and notification emails | SendGrid, a Twilio company | United States |
Documents are held only in the United Kingdom. Account identity data, meaning names and email addresses, is processed in the United States by Clerk and SendGrid under the UK International Data Transfer Addendum. Neither of them ever receives a document.
No. Nothing is checked by hand. The stages above run automatically, then your mortgage adviser reviews the result and decides. Casefile staff do not routinely have access to documents at all.
No. The model used at stage four runs inside Amazon's London region on infrastructure we control. Your document is not sent to a public AI service, is not retained by the provider, and is not used to train anything.
Because people photograph documents. A photo has no readable text inside it, so something has to interpret the image. Uploading the PDF from your banking app avoids that stage entirely and gets a faster, more accurate answer.
Your adviser, every time. Casefile shows what it found and where it found it. It never approves or rejects a document, and a flagged document can still be accepted if your adviser is satisfied.
Every check shows its evidence and the source it came from, so a wrong result is visible rather than hidden. Your adviser can open the document at the point the check refers to and overrule it.
Yes. Compliance teams and networks can ask for our Data Security Policy and we will complete a supplier security questionnaire. Email marvel@gocasefile.com.
Updated 12 October 2026